Serving IoT Communications over Cellular Networks

Serving IoT Communications over Cellular Networks

Challenges and Solutions in Radio Resource Management for Massive and Critical IoT Communications

AMIN AZARI

PhD Thesis in Information and Communication Technology School of Electrical Engineering and Computer Science KTH Royal Institute of Technology Stockholm, Sweden 2018

TRITA-EECS-AVL-2018:73 ISBN 978-91-7729-973-8

KTH School of Electrical Engineering and Computer Science SE-164 40 Kista SWEDEN

Akademisk avhandling som med tillstånd av Kungl Tekniska högskolan framlägges till offentlig granskning för avläggande av doktorsexamen i Informations- och kommunikationsteknik fredagen den 23 november 2018 klockan 10:00 i Sal C, Electrum, Kungl Tekniska högskolan, Kistagången 16, Kista. © Amin Azari, November 2018 Tryck: Universitetsservice US AB

iii Abstract Internet of Things (IoT) communications refer to the interconnections of smart devices, with reduced human intervention, which enable them to participate more actively in everyday life. It is expected that introduction of a scalable, energy efficient, and reliable IoT connectivity solution can bring enormous benefits to the society, especially in healthcare, wellbeing, and smart homes and industries. In the last two decades, there have been efforts in academia and industry to enable IoT connectivity over the legacy communications infrastructure. In recent years, it is becoming more and more clear that the characteristics and requirements of the IoT traffic are way different from the legacy traffic originating from existing communications services like voice and web surfing, and hence, IoT-specific communications systems and protocols have received profound attention. Until now, several revolutionary solutions, including cellular narrowband-IoT, SigFox, and LoRaWAN, have been proposed/implemented. As each of these solutions focuses on a subset of performance indicators at the cost of sacrificing the others, there is still lack of a dominant player in the market capable of delivering scalable, energy efficient, and reliable IoT connectivity. The present work is devoted to characterizing state-of-the-art technologies for enabling large-scale IoT connectivity, their limitations, and our contributions in performance assessment and enhancement for them. Especially, we focus on grant-free radio access and investigate its applications in supporting massive and critical IoT communications. The main contributions presented in this work include (a) developing an analytical framework for energy/latency/reliability assessment of IoT communications over grant-based and grant-free systems; (b) developing advanced RRM techniques for energy and spectrum efficient serving of massive and critical IoT communications, respectively; and (c) developing advanced data transmission/reception protocols for grant-free IoT networks. The performance evaluation results indicate that supporting IoT devices with stringent energy/delay constraints over limited radio resources calls for aggressive technologies breaking the barrier of the legacy interference-free orthogonal communications. Keywords: 5G, Battery lifetime, Grant-based and grant-free access, Massive and critical IoT communications, Radio resource management.

iv Sammanfattning Sakernas Internet (IoT) handlar om sammankoppling av smarta enheter med liten mänsklig inblandning. Införandet av en skalbar, energieffektiv och tillförlitlig anslutningslösning för IoT kan ge enorma fördelar för samhället, framförallt inom hälso- och sjukvård, välbefinnande, smarta hem och industrier. Under de senaste två decennierna har det därför gjorts stora ansträngningar inom akademi och industri för att möjliggöra anslutning av storskalig IoT över nuvarande kommunikationsinfrastruktur. Bara under de senaste åren har det dock blivit allt tydligare att egenskaper, krav och tekno-ekonomiska modeller av IoT skiljer sig markant från nuvarande trafik såsom röstsamtal och internetsurfning. Av denna anledning har flera revolutionerande lösningar, inklusive cellulär smalbands-IoT, SigFox och LoRaWAN, föreslagits och implementerats. Var och en av dessa lösningar fokuserar dock endast på en delmängd av möjliga prestationsmått. Samtidigt finns det fortfarande en brist på dominerande aktörer på marknaden för att leverera en skalbar, energieffektiv och tillförlitlig IoT-lösning. Detta arbete går ut på att karaktärisera de främsta lösningarna för möjliggörandet av IoT-kommunikation, deras begränsningar, samt våra egna bidrag på området. I synnerhet fokuserar vi på icke-ortogonal multipelåtkomst (NOMA) och undersöker dess tillämplighet för att stödja massiv och kritisk IoT-kommunikation. Nyckelord: 5G, Batterilivslängd, Massiv och kritisk IoT-kommunikation, Icke-ortogonal multipelåtkomst (NOMA).

In memory of my grandfather, Karbala-e Mohammad Javad Haghghi (1925; 22-10-2018), who taught me the knowledge that one cannot find in any college, I devote my thesis to the One, who created an interconnected world of live things when Internet of Things was not a fashion, the one who created the learning brains when machine learning was not a fundhunting term. To the one, who gives us the opportunity to live, love, and evolve. Every day is an extra for us to identify where did we come from, where are we going, and why have we been created?(1)

‫آرخ ننمائی وطنم‬

‫هب کجا می روم؟‬

‫آدمنم بهر هچ بود؟‬

،‫زکجا آدمهام‬

‫یا هچ بودهست رماد وی ازین ساختنم‬

‫ زک هچ سبب ساخت رما‬،‫مانده ام سخت عجب‬

‫قفس‬ ‫چند روزی ی ساخته اند از بدنم‬

‫نیم از عالم خاک‬

،‫رمغ باغ ملکوتم‬

‫هف‬ (1) Jalal Al-Din Rumi, 13th-century Persian poet. ‫)موالان جالل الدین رومی شارع افرسی رقن تم‬1(

vi

Acknowledgments I finished my bachelor and master studies in University of Tehran, Iran. Right after, I started my PhD studies in January 2014 in the radio systems lab (RS-Lab) of KTH Royal Institute of Technology on enabling large-scale IoT connectivity over cellular networks. This thesis summarizes my contributions in this field. During my PhD studies in KTH, I found the RS-Lab a very exciting place for research, empowered by scholars with a variety of experiences in academia and industry. To name the people, first I would like to express my sincere gratitude to my advisors, Professor Jens Zander, Associate Professor Guowang Miao, and Assistant Professor Cicek Cavdar for offering me this PhD position and their kind support during these five years. Special thanks to Cicek for offering supervision after licentiate, Guowang for offering supervision until licentiate, and Jens for his long-term forethought supervision. Outside KTH, I would like to extend my thanks to Professor Petar Popovski and Associate Professor Cedomir Stefanovic for their support during my visit to Aalborg University in 2017, and afterwards. Also, Dr. Aurelian Bria from Ericsson AB, thanks for your help in filing my first patent, and your advises afterwards. Special thanks to Professor Merouane Debbah for accepting the role of opponent for my thesis, and to Professor Marina Petrova for reviewing my thesis. I am also grateful to members of the grading committee, Professor Matti Latva-Aho, Professor Viktoria Fodor, and Dr. Stefan Parkvall. Also, Dr. Ali Behravan, thanks for reviewing my PhD proposal. I am also grateful to all other members of the RS-Lab because of their valuable comments during the discussion seminars, specially Ki Won. Slimane, Mats, Anders, Göran, Claes, Ki Won, Chip, Lena, Håkan, Robert, Jan, and Bengt: thanks for having your office doors open for questions. Serveh, Amirhosein, Ali, Andres, Tatijana, Luis, and previous members of the RS-Lab: thanks for transferring your valuable knowledge to us. Haris, I am lucky to start my PhD studies on the same day as you, and thank you for offering your kind help when I required, and even when you realized (like the Swedish abstract). Yanpeng, I am really proud of having a kind friend like you. You have helped me a lot in my career, and made me happy by your good progress in learning Persian. Istiak, thanks for sharing your knowledge, experience, ideas, and offering cooperation in research. Peiliang, you are my hardworking colleague, and have contributed a lot in my career. Meysam, thanks for your real cooperation in our joint works. Seyed Amirhosein, Rehan, Ashraf, Sara, Milad, Forough, Michael, Abdulrahman, Mustafa, Peng, and Mohammad Galal: thanks for all the good memories and our interactive discussions. Sarah, Jenny, Susy, Emanuel, Madeline, and other PhD and HR officers: thanks for offering help in my administrative affairs. Sarah, you are really a helpful colleague who eased my stay in Sweden, especially in the first year. My further thanks to the migrationsverket and the embassy of the I. R. of Iran for their kind support when needed. For writing this thesis and its relevant publications, I have frequently shorten

vii my vacations and holidays, which in essence belong to the family. My sincere thanks to Mohaddeseh, my dear wife, and MohammadBagher, my dear son, for understanding me, and I hope to compensate for these years in the near future. Last but not least, special thanks to my parents, Alimadad and Fatemeh; siblings, MohammadKazem, Iman, and Zahra; and family in law for their endless support and encouragement in my life. Amin Azari

Contents Contents

viii

List of Figures

x

List of Frequently Used Acronyms

xiii

1 Introduction 1.1 Background . . . . . . . . . . . . 1.2 Previous Work . . . . . . . . . . 1.3 Challenges and the Thesis Focus 1.4 Research Questions . . . . . . . . 1.5 Methodology . . . . . . . . . . . 1.6 Contributions . . . . . . . . . . . 1.7 Thesis Outline . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

2 IoT Connectivity over Cellular Networks: and key results 2.1 Serving IoT Traffic over Shared Systems . 2.2 Serving IoT over Dedicated Systems . . . 2.3 Concluding Remarks . . . . . . . . . . . .

. . . . . . .

1 2 4 13 15 17 19 26

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27 27 42 46

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

Challenges, solutions,

3 Enabling Ultra-Energy-Efficient IoT Connectivity: Challenges, solutions, and key results 3.1 Grant-free Access: A performance analysis . . . . . . . . . . . . . . . 3.2 Deployment and Operation Strategies for Grant-free Access Networks 3.3 Enhanced Transmission and Reception Schemes for Grant-free Access 3.4 Distributed Coordination of Grant-free Access Networks . . . . . . . 3.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . .

49 49 54 58 62 67

4 Enabling Ultra-Reliable IoT Connectivity: Challenges, solutions, and key results 69 4.1 RRM for Non-Scheduled Critical IoT Communications . . . . . . . . 69 4.2 Proposed Solutions and Key Results . . . . . . . . . . . . . . . . . . 71 viii

CONTENTS 4.3

Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix 74

5 Conclusions and Future Research Directions 77 5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5.2 Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . 79 Bibliography

81

A Relevant Publications

89

Reprint of Publications

93

Patent: System and Method for Providing Communication Rules Based on a Status Associated with a Battery of a Device

95

Paper A: Network Lifetime Maximization for Cellular-Based M2M Networks 101 Paper B: Latency-Energy Tradeoff based on Channel Scheduling and Repetitions in NB-IoT Systems 117 Paper C: Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks 127 Paper D: Grant-Free Radio Access for Cellular IoT

141

Paper E: Self-organized Low-power IoT Networks: A Distributed Learning Approach 155 Paper F: Serving Non-Scheduled URLLC Traffic: Challenges and Learning-Powered Strategies 165

List of Figures 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9

Illustrative examples of the system model. . . . . . . . . . . . Our focus in studying the IoT connectivity service . . . . . . . The requirements of IoT communications. . . . . . . . . . . . . SigFox and LoRaWAN network architecture . . . . . . . . . . . Simplified access protocol exchanges in SigFox . . . . . . . . . Simplified access protocol exchanges in LoRaWAN . . . . . . . Interference measurement in the ISM band . . . . . . . . . . . Research approach: steps toward answering research questions. Research methods . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

2 3 5 8 10 10 15 19 19

2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 2.10 2.11 2.12 2.13 2.14

Access protocol exchanges in LTE/LTE-A . . . . . . . . . RA resource provisioning and traffic arrival modeling . . . Access rate versus number of preambles in each RA . . . Leveraging CFO for collision detection and resolution . . Empirical PDF of individual lifetimes . . . . . . . . . . . SIL/LIL network lifetime comparison . . . . . . . . . . . . Resource provisioning for IoT . . . . . . . . . . . . . . . . Performance tradeoffs . . . . . . . . . . . . . . . . . . . . b Tradeoff between Econs , D1 , and D2 . . . . . . . . . . . . . b Tradeoff between Econs and energy efficiency of P2 devices NB-IoT technology . . . . . . . . . . . . . . . . . . . . . . Data exchanges and their respective power consumptions Queuing model of the NB-IoT . . . . . . . . . . . . . . . . Performance evaluation . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

28 31 31 32 38 38 39 40 41 42 43 44 44 47

3.1 3.2 3.3 3.4 3.5 3.6 3.7

Graphical description of motivation for RQ2.1 . . . . . . . . . . Validation of analytical and simulation results . . . . . . . . . . Contributions related to RQ2.2 . . . . . . . . . . . . . . . . . . Tradeoff between network cost, reliability, and battery lifetime The optimized deployment strategy . . . . . . . . . . . . . . . . The optimized operation control . . . . . . . . . . . . . . . . . Scalability analysis . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

51 54 56 56 57 58 59

x

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

List of Figures

xi

3.8 3.9

61

3.10 3.11 3.12 3.13 3.14

The proposed receiver design . . . . . . . . . . . . . . . . . . . . . . . . Received packets at the receiver. VFs have different carrier frequencies, start times, and slot sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . Battery lifetime and latency analysis . . . . . . . . . . . . . . . . . . . . Tuning design parameters for ADP(N1 , N2 , dth ) . . . . . . . . . . . . . . Learning for power and SF selection control . . . . . . . . . . . . . . . . The constellation of selected SFs . . . . . . . . . . . . . . . . . . . . . . Learning for transmit power and SF selection control . . . . . . . . . . .

61 61 62 65 66 67

4.1 4.2 4.3

An illustrative example of the system model . . . . . . . . . . . . . . . . Hierarchical ML-powered RRM . . . . . . . . . . . . . . . . . . . . . . . Performance evaluation of the proposed RRM . . . . . . . . . . . . . . .

70 73 74

List of Frequently Used Acronyms

3GPP LTE 5G AP BS cIoT EE eMBB IoT ISM band LoRaWAN MAC mIoT ML MNO NB-IoT PDCCH PDSCH PPP PRACH PUCCH PUSCH QoS RA RRM UE URLLC

3rd generation partnership project long-term evolution The fifth generation of wireless networks Access point Base station Critical Internet of thing Energy efficiency Enhanced mobile broadband Internet of thing The industrial, scientific and medical radio band LoRa wide area network Media access control Massive Internet of thing Machine learning Mobile network operator Narrowband IoT Physical downlink control channel Physical downlink shared channel Poisson point process Physical random access channel Physical uplink control channel Physical uplink shared channel Quality of service Random access Radio resource management User equipment Ultra-reliable low-latency communications

xiii

Chapter 1

Introduction Internet of Things1 (IoT) refers to the ever-growing network of smart physical objects that are capable of sensing or acting on their environment, and the communications between these objects and other Internet-enabled systems. IoT enables smart devices to participate more actively in everyday life, business, industry, healthcare, and hence, in a more intelligent connected society. Providing such a large-scale IoT connectivity first became a target in the communications society during research on the fourth generation of wireless networks (4G) [2, 3]. However, regarding the urgent need to serve high volumes of human-generated communications’ traffic, e.g. web-surfing, IoT connectivity was excluded from the major design objectives of 4G wireless networks [4]. During the transition period from 3G to 4G, the first attempts for providing IoT connectivity were made by companies outside the 3GPP2 , and early IoT solutions were introduced over the unlicensed spectrum, e.g. SigFox, before the public introduction of 4G networks [4]. After successful and widespread deployment of 4G networks, enabling large-scale IoT connectivity became a major research subject in academia and industry, and was embedded in most big projects aimed at shaping the 5G networks, e.g. METIS I and II [5]. Based on the early-stage expectations from the 5G networks, telecommunications (telecoms) equipment vendors predicted more than 50 billion of connected devices by 2020 [6], mobile network operators (MNOs) counted on big contribution of the IoT traffic on filling their revenue gaps [7, 8], IoT was scaled in literature to the Internet of everything (IoE) [9], and more than 10 years of device battery lifetime became a favorite goal across the 5G projects [10]. However, in recent years it has become more and more clear that realizing IoE, connecting billions of devices to work for years cannot be achieved easily. Indeed, we foresee that providing scalable, energy efficient, and reliable IoT connectivity will remain as a goal even in the post 5G era [11]. Especially, enabling large-scale IoT connectivity with strict delay1 Parts of this chapter have been previously published in chapter 1 of the author’s licentiate thesis in 2016 [1]. 2 3GPP: 3rd generation partnership project.

1

2

CHAPTER 1. INTRODUCTION

Active IoT device Inactive IoT device

non-IoT service IoT service

Active IoT device Inactive IoT device

non-IoT service IoT service

Figure 1.1: Illustrative examples of the system model.

reliability constraints, which is the baseline for many promising 5G applications, is a less-explored direction, and calls for intense further research. This thesis presents the key challenges in enabling large-scale IoT connectivity, including scalability, energy efficiency (EE), and reliability of communications, and aims at addressing them. The long-term objective is to design new network architectures, protocols, and algorithms such that billions of loT devices could be supported in cellular systems on a global scale. Before going to the details of challenges in enabling large-scale IoT connectivity, we need to make sure the reader has the same background and terminology as we use throughout the thesis. Towards this end, the necessary background on the IoT traffic, device, quality of service (QoS), and technologies of interest, is given in the next section. Later, we come back to the challenges, and our proposed solutions for tackling them.

1.1

Background

This section provides the necessary background for the thesis, starting from a description of system model that leads to the problems covered by this research.

1.1.1

System Model

In this thesis, we aim at enabling large-scale IoT connectivity service, mainly in the uplink direction. By the IoT connectivity service, we mean receiving devices’ transmitted packets using base stations (BSs), and delivering them to the network layer, to be transmitted to the destination(s) of interest. The communications system considered in this work is content-blind, i.e. it doesn’t care about the packets’ contents, and just works as a medium for connecting IoT devices to the cloud, as depicted in Fig. 1.2. The networking protocols and higher layers of the

1.1. BACKGROUND

3

Cloud

MAC

MAC

PHY

PHY

The main focus Figure 1.2: Our focus in studying the IoT connectivity service

protocol stack are out of the scope of this thesis. The IoT connectivity service could be provided by systems and resources which are shared with non-IoT services, or are dedicated to the IoT service, as depicted in Fig. 1.1. Direct device to device communications are out of the scope of this thesis, and hence, all communications are assumed to pass through the access network. The set of IoT devices served in the network is vastly heterogeneous, and may include anything which benefits from being connected, e.g. sensors, actuators, and vehicles. The considered IoT devices might be powered by the grid or batteries, however, throughout the thesis, our main focus is on battery-powered devices. But, which wireless connectivity solution can handle such a heterogeneous large-scale system of IoT devices?

1.1.2

The requirements

To answer the above question, we first need to further explore the general requirements of IoT communications. Scalability: One key difference between IoT communications and the legacy human-oriented communications consists in the potential number of concurrently active users in the service area [12]. In an IoT-enabled network, thousands of devices may require access at the same time, e.g. after a power outage or fire event [13]. This calls for communications protocols able to handle a massive number of concurrent IoT connections. Energy efficiency: In contrast with the legacy user equipment (UE) in 1G4G wireless networks, i.e. mobile phone, which has a daily recharging pattern, operation of most IoT devices is dependent on the lifetime of their embed batteries. Then, human intervention is required when they become out of energy [12, 14]. As a result, shortcoming in providing long battery lifetime incurs huge maintenance

4


costs, and is also in contrast with the main goal of IoT, i.e. reducing human intervention. This calls for design of communications protocols able to serve IoT communications in an energy efficient way. Cost efficiency: The feasibility of many IoT applications from the technoeconomic perspective mandates that the offered IoT service must be achievable by low-cost devices [15]. This is mainly due to the fact that having low-cost devices enables achieving reliability by over-deployment of devices. Furthermore, the number of required sensors in many reporting applications, e.g. smart agriculture, increases significantly with the increase in the operation area, and hence, cost per module significantly contributes in economic viability of any IoT solution. This calls for design of communications protocols able to communicate with low-cost IoT devices. Delay-Reliability: The tolerated delay by different IoT applications may vary from milliseconds, e.g. in alarm applications, to hours, e.g. in reporting applications [16, 17]. Furthermore, depending on the application, the level of reliability in achieving such delay bound may vary significantly, e.g. from 90% to 99.999% [17]. In massive IoT (mIoT) communications3 , successful data transmission rates of order 0.9-0.999 in relatively long time intervals, ranging from seconds to hours, are usually acceptable [18]. In contrast to the mIoT, in critical IoT (cIoT) applications4 , e.g. in intelligent transportation systems (ITS), ultra-reliability within a tightly bounded time interval is a requirement [19]. The stringent and unseen requirements of cIoT communications call for design of communications protocols able to serve ultrareliable low-latency communications, when and where needed. Given the system model and general requirements of IoT communications, in the next section, we present the related works, in terms of existing solutions and state-of-the-art research projects aimed at offering a large-scale IoT connectivity solution.

1.2 1.2.1

Previous Work Existing solutions

The continuing growth in demand for IoT connectivity has encouraged different industries to investigate evolutionary and revolutionary radio access technologies for serving IoT communications [20]. Broadly speaking, one can categorize the dominant IoT connectivity solutions based on their radio resources into two categories: i) solutions over the licensed spectrum like LTE category M and NB-IoT [21]; and ii) solutions over the unlicensed spectrum like SigFox and LoRa wide area network (LoRaWAN) [18]. In the following, we investigate these categories in details. 3 In the context of 5G, the mIoT communications are also called massive machine-type communications (mMTC). 4 The cIoT communications are classified under the ultra-reliable low-latency communications (URLLC) in the context of 5G.

1.2. PREVIOUS WORK

5

Delayreliability

Energy efficiency

IoT IoT connecti connectivity vity

Scalability

Cost efficiency

Figure 1.3: The requirements of IoT communications.

1.2.1.1

Cellular networks over the licensed spectrum

Cellular networks are expected to play a critical role in realization of a networked society because of their ubiquitous coverage and roaming capability [20]. Following the early-stage predictions of a huge increase in the IoT traffic by Cisco [22], MNOs have started serving IoT traffic over cellular networks in hope of decreasing their revenue gaps [8, 20]. The IoT traffic over cellular networks is usually served in three different ways: (i) it is served similar to the legacy traffic, i.e. it passes the same procedures as non-IoT traffic passes; (ii) it is served along with the nonIoT traffic but leveraging special procedures, which have been implemented for enabling efficient IoT connectivity; and (iii) it is served over a dedicated pool of radio resources using IoT-specific algorithms and protocols. The first item, i.e. IoT/non-IoT serving using the same radio resources and protocols, is usually the case in 2G networks, and it out of the scope of this work. The other two items are investigated in the following subsections. IoT/non-IoT traffic serving over the same resources using dedicated communications protocols The legacy connectivity procedure in LTE-Advanced (LTE-A) networks includes synchronization, connection establishment, authentication, scheduled data transmission, confirming successfulness of data transmission, and connection termination [23]. Random access channel (RACH) of the LTE and LTE-A systems is the typical way for IoT devices to establish connection with the BS, and become connected to the network. In RACH, each device randomly chooses a preamble from a set of available preambles for contention-based access to the BS. Since the total number

6


of available preambles is limited, the random access (RA) procedure can introduce collisions and energy wastages, especially when a massive number of devices try to get access to the network [23]. Once a device successfully passes the RA procedure, it sends scheduling requests to the BS through the physical uplink control channel (PUCCH), the BS performs the scheduling and sends the scheduling grants back through the physical downlink control channel (PDCCH), and the granted devices send data over the granted physical uplink shared channel (PUSCH) resources to the BS. The connection establishment, scheduling, and scheduled data transmission procedures in the existing cellular networks have been designed and optimized for a totally different traffic pattern than the IoT communications. While they work well for a limited number of long-length communications sessions, for battery-limited IoT devices with a massive number of short-length communications sessions these procedures are the bottlenecks. Then, network congestion, including radio network congestion and signaling network congestion, is likely to happen [24]. Towards isolating the impact of serving IoT communications from non-IoT communications, in LTE release 13, use of dedicated resources for IoT communications has been proposed, to be discussed in the sequel. IoT traffic serving over dedicated resources and communications protocols The development of LTE for low-cost massive IoT over dedicated resources has been initiated in release 12, and has continued in release 13 [25]. In the LTE release 13, LTE category M (LTE-M) and narrow-band IoT (NB-IoT) systems have been introduced which are expected to provide IoT connectivity over dedicated 1.4 MHz and 200 KHz bandwidths respectively [25]. In the following, we focus on the NB-IoT technology. The investigation of LTE-M is out of the scope of this work. NB-IoT technology represents a big step towards accommodation of IoT traffic over cellular networks [26]. Data transmission in NB-IoT takes place in a narrow bandwidth, i.e. 200 KHz, which results in more than 20 dB extra link budget compared to the standard LTE-A systems. This extra link budget enables smart devices deployed in remote areas, e.g. basements, to directly communicate with the base stations. Since the legacy signaling and communication protocols have been designed for large bandwidths, e.g. 20 MHz in LTE-A, NB-IoT introduces five novel narrowband physical channels. The new channels include: i) narrowband physical random access channel (NPRACH); narrowband physical uplink shared channel (NPUSCH); narrowband physical downlink shared channel (NPDSCH); narrowband physical downlink control channel (NPDCCH); and narrowband physical broadcast channel (NPBCH) [27, 28]. NB-IoT also introduces four new physical signals: i) narrowband reference signal (NRS); ii) narrowband primary synchronization signal (NPSS), iii) i) demodulation reference signal (DMRS), which is sent along with user data on NPUSCH; and iv) narrowband secondary synchronization signal (NSSS). When a device has data to transmit, it first listens for receiving cell information, e.g. NPSS and NSSS, through which the device makes itself synchro-

1.2. PREVIOUS WORK

7

nized with the BS. Then, it waits for the NPRACH to establish connection, for NPDCCH to receive the scheduling grants, and finally for NPUSCH/NPDSCH to send/receive data to/from the BS. One major difference between LTE-A and NB-IoT systems consists in the fact that in the latter, IoT devices can go from the idle mode to the deep sleep mode during which, device is still registered to the BS. Thus, upon waking up, it can become connected by sending/receiving a fewer number of messages that the case of LTE-A [21]. This new functionality saves a considerable amount of energy in the connection establishment procedure [29]. Furthermore, NB-IoT systems benefit from coverage extension, which is achieved by repetition of transmitted signals in the time domain. In an NB-IoT cell, BS assigns each device to a coverage class, from a limited number of defined classes, based on the estimated device-BS path-loss. A coverage class is characterized by the number of replicas per original data packet that must be transmitted. For example, based on the specifications in [28], each device belonging to the group i shall repeat 2i−1 times the preamble transmitted over the NPRACH, where i ∈ {1, · · · , 8}. Using this repetition, it is expected that NB-IoT will be able to offer deep coverage to most indoor areas [21]. 1.2.1.2

IoT networks over the unlicensed spectrum

From 1G to 4G, the telecommunications industry has spent a great deal of resources investigating how to realize high-throughput infrastructure but forgotten about scalable low-power low-rate systems in which, energy efficiency and battery lifetime are crucially important. Along with telecoms’ transition from 3G to 4G networks, some 3GPP-independent companies have tried to enter the market and provide large-scale IoT connectivity over the unlicensed spectrum, e.g. SigFox and LoRaWAN. While reliability over the unlicensed spectrum is a big concern for these solutions, they have focused on minimizing the cost per device and maximizing the battery lifetime in order to deliver a competitive service at low cost. Regarding the fact that energy consumption in the synchronization, connection establishment, and signaling is comparable with, or even higher than, the energy consumption in the actual data transmission [29], IoT solutions over the unlicensed spectrum leverage grant-free radio access for energy saving [30]. In the grant-free access, once a packet is triggered at the device, it is transmitted without any handshaking with the BS or authentication process. Furthermore, the IoT solutions over the unlicensed spectrum mainly leverage narrowband, or even ultra-narrowband, communications in order to cover a large area with the minimum possible energy consumption at the device-side [31]. These low-power wide-area (LPWA) IoT networks over the unlicensed spectrum, along with cellular NB-IoT and LTE-M networks, are expected to share 60 percent of the IoT market among themselves, a number that is expected to grow over time, and hence the competition between LPWA technologies is becoming intense [31]. In the sequel, we introduce SigFox and LoRaWAN as two dominant LPWA IoT networks in the market. Fig. 1.4 represents the network architecture

8


IoT Server

Internet

Access points

Figure 1.4: SigFox and LoRaWAN network architecture

for SigFox and LoRaWAN. In this architecture, data is collected by the local access points (APs). These access points are connected to the Internet through wired or wireless backhauling. The IoT user can access the gathered data through the IoT server, process and visualize it, and potentially send commands back to the devices through the access network. Both LoRaWAN and SigFox use the industrial safety medical (ISM) spectrum band for communications. The European Telecommunications Standards Institute (ETSI) has defined multiple European ISM bands. In this work, we focus on the 868 MHz ISM band. This band is part of the licensefree spectrum, and hence, any device can use it for communications as long as it complies with the fair spectrum use regulations [18]. The regulations vary from one sub-band to another, and are mainly restricting the effective radiated power (ERP) as well as the spectrum access method [32]. The spectrum access method can be either listen-before-talk (LBT), in which devices starts communications if the channel is free, or based on the duty cycle. In the latter, a device starts transmission once it has data to transmit by considering the constraint that it cannot transmit in more than x% of the time, where x is determined from the regulations. The communications protocols used in the physical layer of SigFox and LoRaWAN are described in details in the following sections. SigFox SigFox, founded in 2009, is the first widely announced proprietary IoT technology aiming at providing low-cost low-power IoT connectivity over the unlicensed spectrum. Thanks to its ultra-narrowband signals, 100 Hz, SigFox is able to cover large outdoor areas, and even is able to provide high link budget for indoor coverage. Once a SigFox device has a packet to transmit, it transmits one copy of the packet immediately, and two other copies consequently with a semi-random frequency hopping over 600 KHz system bandwidth [33, 32]. While in the first versions, SigFox was only supporting uplink communications; from 2015 it also

1.2. PREVIOUS WORK

9

supports uplink-initiated downlink communications. This means that once a SigFox device transmits data and requests for a response from the application, the access point has an opportunity to send the acknowledgment (ACK) as well as the queued data waiting for the device [30], as depicted in Fig. 1.5. As per other grantfree solutions, SigFox doesn’t require pairing, which means that there is no need for synchronization, authentication, and connection establishment message exchanges between devices and the APs [32, 33]. The duty cycle for the SigFox AP is 10%, i.e. a SigFox AP can transmit data at most in 10% of the time, while the duty cycle for devices is 1%. The difference comes from the fact that AP and devices are working on different sub-bands of the ISM band, and hence, they should comply with different regulations. The duty cycle for SigFox devices forces an upper-limit of 140 messages of 12-bytes per device per day, which makes it unfavorable for some applications with more data exchange requirements. As depicted in Fig. 1.4, the core SigFox servers monitor the status of the network and manage operations of the APs. SigFox modules are available at low cost, e.g. 2$ per module [34], to be used in various electronic devices. A cost comparison in [34] represents that the cost of a SigFox module is almost one-third of a LoRa module, and one-fifth of a cellular IoT module.

LoRaWAN Similar to SigFox, LoRaWAN provides low-power IoT connectivity over the unlicensed spectrum. The access protocol exchanges of LoRaWAN has been depicted in Fig. 1.6. In contradictory with the SigFox, one may build her own LoRaWAN gateway, connect it to the Internet, and use it for long-range data gathering from LoRa-compatible IoT devices. The LoRaWAN servers manage allocation of spreading factors (SFs) to devices in order to balance the data-rate and reliability of the communications. SFs, ranging from 7 to 12, denote the number of chirps used to encode a bit. On the one hand, the higher chirp rate is, the better reconstruction of the received signal is achieved. On the other hand, increase in the SF stretches the transmission time [35], and hence, increases the probability of collision with another data transmission. LoRa is utilizing chirp spread spectrum (CSS) as a modulation to mitigate the potentially severe interference on the unlicensed bandwidth and has been claimed to be robust against multi-path fading and Doppler shifts to some extent [36]. Regarding the ever-increasing interest in deployment of IoT solutions over the ISM band, and hence the coexistence issues, a high level of resilience to sporadic interference will be of paramount importance for any ISM-band solution. The main feature of CSS, which is used in the LoRa modulation, is that signals with different SFs can be distinguished and received simultaneously, even if they are transmitted at the same time and on the same frequency channel [37]. In LoRaWAN (class A), downlink transmission is possible after data transmission of a device during two non-overlapping and consecutive dedicated time intervals, called receive windows [38].

10

CHAPTER 1. INTRODUCTION Backend SigFox

IoT device data

Backend application

Frontend application IoT app./User

Call back Ack

Ack

DL confirm

Figure 1.5: Simplified access protocol exchanges in SigFox [39]

IoT device

LoRa gateway

Network server

IoT app./User

data

Ack

Figure 1.6: Simplified access protocol exchanges in LoRaWAN [18]

1.2.2

Related research

This section presents a survey of state-of-the-art publications focused on addressing challenges in enabling large-scale IoT connectivity. More detailed state-of-the-art could be found in the appended publications A-F. Broadly speaking, one may categorize the research works focused on enabling large-scale IoT connectivity based on their deviation from the legacy LTE networks as evolutionary and revolutionary solutions. Evolutionary solutions aim at serving IoT traffic along with the non-IoT traffic, i.e. over the same resources, but using dedicated communications procedures [18]. Revolutionary solutions aim at serving IoT traffic over dedicated systems and communications protocols. In the following, we first start with the evolutionary solutions and then proceed towards the revolutionary ones. 1.2.2.1

Evolutionary solutions

Evolutionary solutions mainly focus on enhancing connectivity procedure of existing LTE networks, e.g. they improve the legacy connection establishment and scheduling procedures [40, 41, 42, 43]. In the following, we study state-of-the-art proposed solutions for enhancing connection establishment and scheduling procedures.

1.2. PREVIOUS WORK

11

Connection establishment: Random access channel of the LTE systems is the typical way for IoT devices to become connected to the BS. The capacity limits of the RA procedure for serving mIoT and a survey of improved alternatives have been studied in [13]. In order to prevent network congestion, in case a massive number of IoT devices try to connect to the BS, several solutions have been proposed in [44]. Among these solutions, access class barring (ACB) has attracted lots of attention in literature [45]. In this scheme, when a device needs cellular connectivity, it decides to access the network with probability qa or defer its access with probability 1 − qa . The corresponding access probability qa for each class of nodes/traffic is broadcasted by the BS. A novel contention resolution scheme, empowered by distributed queuing (DQ), has been proposed in [46]. This DQ-based scheme improves the successful utilization of RA resources by splitting contending devices into groups for the subsequent RA transmissions. In order to isolate QoS of different traffic types in coexistence scenarios, dynamic/static strategies for separation of RA resources have been proposed in [47]. For delay tolerating IoT applications, a time-controlled network access has been proposed in [48]. In this scheme, the IoT devices are divided into classes based on their QoS requirements, and fixed RA/data transmission resources are reserved to occur for each class in (semi) regular time intervals. Finally, in [49] short packet transmission over the RA resources of LTE-A networks has been proposed in order to decrease the required signaling in connection establishment and scheduled data transmission. Resource scheduling: Scheduling is the process performed by the BS to assign radio resources to user devices. In cellular networks, scheduling is not part of the standardization work, and is usually left for vendor implementation. However, due to the fact that signaling is standardized, each implemented scheduler should comply with the control requirements specified in the standards. Literature study on IoT scheduling over LTE/LTE-A networks reveals that delay, throughput, transmit power, and impact on QoS of non-IoT traffic have been the frequently used IoT scheduling metrics in previous works [24]. In [23], major challenges in IoT scheduling over LTE networks have been studied. Energy efficiency of data communications over LTE networks has been studied in [50], and it has been shown that LTE physical layer is not energy efficient for small data communications. Design of power-efficient uplink scheduler for delay-sensitive traffic over LTE systems has been investigated in [51], where the considered delay models and traffic characteristics are not consistent with the ones of the IoT traffic [52]. Uplink scheduling with IoT/non-IoT coexistence over cellular networks has been investigated in [53]. In [53], a simple energy consumption model has been used, which takes only the transmit power for reliable data transmission into account. Thus, the other sources of energy consumptions, including energy consumptions in resource reservation and operation of electronic circuits, which are comparable with the energy consumption for reliable data transmission [54, 55], have been neglected. Providing an accurate model of energy consumption in different stages of connectivity over cellular networks, and developing optimized media access control (MAC) protocols based on such models are the research gaps in this field, which are addressed in this thesis.

12 1.2.2.2

CHAPTER 1. INTRODUCTION Revolutionary solutions

Revolutionary solutions focus on designing new communications systems for IoT services, and devising new communications protocols for efficient use of them. In the following, we study state-of-the-art proposed revolutionary solutions for serving IoT communications. NB-IoT systems: In LTE Rel-13, NB-IoT has been proposed for serving IoT traffic over a 200 KHz dedicated bandwidth. NB-IoT also introduces a set of coverage classes, each associated with a predefined number of signal repetitions in time. In [56, 57], preamble design for connection establishment over NB-IoT has been investigated. Radio resource scheduling for energy efficient uplink transmission of connected devices has been investigated in [58]. For NB-IoT deployment in LTE guard-band, which introduces external interference to the IoT communications, coverage analysis has been done in [59]. In [29], energy consumption in data transmission in three different coverage regimes, i.e. normal, robust, and extreme, has been studied, where in robust and extreme regimes extra coverage is achieved by signal repetition in the time domain. In comparison with the legacy LTE, the energy consumption analysis in [29] illustrates that NB-IoT can significantly reduce the consumed energy in connection establishment because it lets devices to go to the deep sleep mode while they are registered to the BS. Regarding the fact that NB-IoT is a newly proposed solution under development, there is lack of a comprehensive study on its capacity, energy efficiency, and scalability in literature. Especially, the coexistence of coverage classes with different repetition orders over the same system is expected to enhance energy efficiency and latency for some devices, and severely degrade for others [60]. Finally, there is no work investigating how multiplexing uplink/downlink physical channels, i.e. NPRACH, NPUSCH, NBDCCH, NPDSCH, which are time multiplexed over the same uplink/downlink radio resources, could affect performance indicators of interest. The above research gaps are addressed in this thesis. Grant-free radio access: Thanks to the simplified connectivity procedure and removing the need for pairing and synchronization, grant-free radio access has attracted lots of interests in recent years for enabling energy efficient IoT connectivity, especially when long lifetime is of interest. SigFox and LoRaWAN are the two major grant-free radio access technologies over the ISM-band. While the energy consumptions of LoRaWAN and SigFox technologies are extremely low, and their provided link budgets are enough to penetrate to most indoor areas, e.g. a LoRaWAN message can be decoded when its power is 20 dB less than the noise power, reliability of their communications is questionable, especially in coexistence scenarios [32, 61]. In [62, 63], interference and outage probability analysis for grant-free access networks, including SigFox and LoRaWAN, has been carried out by assuming a constant received power from all contending IoT devices at the receiver. One must note that this assumption is hardly met in practice regarding different path-loss values that different devices experience, and the lack of channel state information as well as sophisticated power control at the IoT-device side. The reliability of grant-

1.3. CHALLENGES AND THE THESIS FOCUS

13

free transmissions has been analyzed in [64] by assuming a homogeneous Poisson point process (PPP) distribution for interfering devices. In [32], experimental measurement results reflect a significant impact of interference from already installed ISM-band devices on the performance of SigFox and LoRaWAN. Based on the analysis in [32], LoRaWAN communications which benefit from the CSS modulation are more immune to the interference, and could survive in coexistence scenarios. One sees that in prior arts, the performance of LoRaWAN and SigFox have been investigated in isolated conditions, i.e. when each technology is the sole user of the spectrum. However, as discussed in [32], coverage and performance of these technologies are more limited by the interference than the noise. Furthermore, the correlation in locations of devices have not been investigated in prior arts, while it has a significant impact on reliability of communications [65].Hence, there is lack of a research work investigating coexistence analysis of grant-free networks, their scalability, reliability, and energy efficiency. In this thesis, we cover this research gap.

1.3 1.3.1

Challenges and the Thesis Focus Challenges

As discussed in the previous section, the existing cellular infrastructure is unable to serve a massive number of concurrent IoT service requests because it has been designed and optimized for a small number of long-lived communication sessions, mainly in the downlink direction [23, 50]. Then, network congestion, including radio network congestion and signaling network congestion, is likely to happen in massive IoT device deployment scenarios [16]. Furthermore, the energy inefficiency of existing systems in serving small data transmissions limits battery lifetime of connected devices. Consider the case that IoT traffic is to be sent through the LTE infrastructure. Then, around 34 bytes overhead data for user datagram protocol (UDP), Internet protocol (IP), packet data convergence protocol, radio link control, and the MAC overhead must be added to the original data [50]. When LTE is used for data-hungry applications, this level of overhead is reasonable. However, in IoT applications, where IoT devices may have few bits of data to transmit and need to last for years on the battery power, this level of overhead makes the communications inefficient. With the introduction of NB-IoT technology, we expect that battery lifetime of IoT devices will be significantly improved in comparison with the LTE networks. However, the preliminary performance analyses show that there is still a considerable gap between the expected battery lifetime in NB-IoT networks and the 5G expectations [29, 60]. This is mainly due to the fact that the major energy consumption sources like synchronization, connection establishment, resource scheduling, and waiting for the control signals from the BS are still there in the NB-IoT standardizations [29, 60]. Furthermore, the cost per IoT module in the NB-IoT systems is almost 5 times more than its competitors in the low-power

14


wide-area (LPWA) technologies over the unlicensed spectrum, e.g. SigFox modules [34]. When it comes to IoT solutions over the unlicensed band, the first shortcoming consists in the number of packets that could be transmitted per day by a device, which is limited by the fair-use regulations [32]. For example, by considering packets of length 2 seconds in SigFox technology5 , a device will be able to transmit at most 3 packets per hour regardless of success or failure of the transmitted packets. Furthermore, the coexistence problem over the unlicensed band results in reliability degradation. Indeed, there is no guarantee that an LPWA IoT solution which is working at a given time, will work properly in the future in case of deployment of a neighboring network working on the same band. To get insights into the interference problem in the ISM band, one may refer to Fig. 1.7, in which interference measurements in the European 868 MHz ISM band have been depicted. In this figure, one sees that in use-cases like a business park, the ISM band has been occupied almost all the time. This means that there is a high probability of collision in data transmission over the ISM band in these use-cases. Furthermore, regarding the fact in those solutions devices are not easily reachable and controllable through the network, change in communications characteristics of devices, e.g. adapting them to the network traffic, is not possible without human intervention. Moreover, from 1G-4G, cellular networks have penetrated into almost all locations, and ubiquitous and seamless coverage has been provided in almost all countries. Cellular towers have been mounted in near-optimal locations to provide as much as possible coverage with as low as possible number of BSs. However, when it comes to the LPWA networks, they are in the initial steps, and their solutions suffer from lack of coverage in many places, especially in indoor areas. Then, providing coverage to the level of existing cellular networks requires a huge amount of investment in the LPWA technologies. The need for huge investment makes the widespread usage of such technologies in future unclear. Finally, enabling ultra-reliable communications in a bounded time interval (cIoT communications) is a major challenge of next-generation wireless networks. Enabling cIoT communications is considered as an essential prerequisite of a new wave of services including drone-based delivery, smart factory, remote control, and intelligent transportation systems (ITS) [19, 66]. The ITS itself can make a significant impact on society and the daily lives of human beings by enabling safer and greener transportations. Whilst enabling cIoT communications is essential for realizing these promising applications, it is never an easy task to guarantee ultrareliability withing a bounded time interval using radio resources. More specifically, unlike the mIoT service, in which significant improvements have been made, the cIoT service is in its infancy [66]. The main reasons behind such gap consist in cIoT’s unseen delay-reliability requirements [19].

5

Recall that SigFox requires transmission of 3 replicas per data packet.

1.4. RESEARCH QUESTIONS

15

Figure 1.7: Interference measurement in the ISM band in coexistence scenarios [32]: a business part (left), a hospital complex (right). (Reprinted from [32], ©2017 IEEE, reused with permission.)

1.3.2

Thesis focus

In this thesis, our focus is on figuring out challenges of existing wireless access solutions in providing IoT connectivity, identifying the root causes, and developing evolutionary and revolutionary solutions, especially in the MAC and physical layers. Based on this focus, the high-level research problem studied in this thesis is as follows: • Is it possible to improve scalability and energy efficiency of IoT communications, compared to what is offered in the existing systems? In the following section, we break down this high-level research question, and present our detailed research questions, as well as our methodologies for solving them.

1.4

Research Questions

In order to start our research on enabling solutions for large-scale IoT communications, it is meaningful to first investigate the IoT connectivity procedures in the state-of-the-art 3GPP LTE systems, analyze their scalability, and evaluate their energy efficiency in serving massive IoT communications. Then, the first research question studied in this thesis is defined as follows. • RQ1: What are the challenges of existing cellular networks in energy-efficient serving of mIoT communications? How can one enhance their scalability and energy efficiency, and up to which extent? In response to this question, we investigate IoT connectivity procedures of the LTE-A and the newly proposed NB-IoT systems. Then, we investigate different

16


sources of energy consumptions, delays, and failures in each step of the connectivity procedure, including synchronizations, connection establishment, scheduling, data transmission, and acknowledgment. These results are subsequently employed in optimizing the MAC design of cellular networks in order to serve mIoT communications in a more energy efficient way. Our performance evaluation results indicate that the major limitations of cellular networks in enabling ultra-energyefficient IoT communications consist in following the legacy grant-based radio access scheme. This means that an IoT devices which has a few bits to transmit is required to transmit/receive several packets of data for connection establishment and scheduled data transmission. In order to break the barriers of the legacy grant-based MAC protocols, and provide solutions able to offer IoT connectivity with ultra-high energy efficiency, the second research question is defined as follows. • RQ2: Given the limitations of cellular networks in supporting mIoT, derived from RQ1, how can one design an energy-efficient MAC for IoT communications, to be integrated into future cellular networks? This question aims at investigating the ways in which, reliability, latency, spectrum efficiency, and receiver complexity could be traded to achieve extra energy efficiency in communications. In response to this question, we focus on grant-free MAC design, a class of techniques suited to support energy-efficient IoT communications in an energy efficient way. We further investigate challenges associated with the grant-free radio access, investigate its reliability/latency/energy efficiency performance, and augment it with advanced schemes, including path-loss adaptive replica transmission control and successive interference cancellation (SIC)-based receiver triggered by carrier frequency offset (CFO) of devices. We further leverage device intelligence to make IoT communications reliable against potential intended and unintended interferences from other IoT devices and jammers, respectively. Finally, we study IoT connectivity with ultra-high reliability requirements. While the literature on ultra-reliable communications is mature [67], the proposed schemes in literature mainly achieve ultra-reliability by leveraging time diversity, i.e. by retransmission of packets. In cIoT applications of our interest, one cannot leverage time diversity because the required reliability should be achieved in a short time interval, which is usually shorter than the channel coherence time. On the other hand, retransmissions of packets in hope of increased reliability is not a scalable solution for uncoordinated multi-user scenarios, as the communications’ reliability decreases by an increase in the traffic load [65]. While enabling cIoT communications is essential for providing many fantastic applications like autonomous driving, its technology is immature even in point to point communications [66]. In this thesis, we mainly focus on radio resource management (RRM) for cIoT communications. In the uplink direction, enabling cIoT requires guaranteeing delay and reliability requirements for both scheduled transmissions as well as non-scheduled transmissions. The non-scheduled transmissions could be from connected/disconnected devices which have urgent data, usually of short payload size, or devices which request

1.5. METHODOLOGY

17

resource reservation for subsequent scheduled transmissions of urgent data, usually of larger sizes. Then, the set of available resources is needed to be managed to be used for both scheduled and non-scheduled communications, and the required reliability for both communications must be assured. As discussed in the above, time-diversity, and hence retransmission, could not be leveraged here. Thus, reliable transmission of data must be carried out using frequency diversity6 . Regarding the fact that frequency resources are limited, the main challenge consists in jointly serving scheduled and non-scheduled cIoT traffic over a limited set of frequency resources. Then, the third research question is defined as follows. • RQ3: What kinds of RRM schemes are suitable for serving scheduled and non-scheduled uplink cIoT traffic over a limited set of radio resources? Isolation among these two traffic types, via allocating orthogonal slices of resources to them, is a common practice when dealing with ergodic objectives like throughput. However, in cIoT applications, due to the limitations of time-diversity and the crucial need to large spectrum bands for providing frequency diversity, isolation through orthogonal slicing is challenging [19, 68]. This is more challenging when we consider the fact that non-scheduled communications are infrequent. Hence, a considerable amount of radio resources must be idle for reliable serving of potential non-scheduled cIoT traffic. Recently, serving coexisting services over limited radio resources has attracted attention, and non-orthogonal RRM for serving cIoT has been proposed [69, 70]. In response to RQ3, we investigate the performance of overlay transmission of non-scheduled cIoT traffic over scheduled transmissions, and study the respective RRM solution, which is needed to manage the radio resources.

1.5

Methodology

The research approach and methods followed in this work has been presented in Fig. 1.8 and Fig. 1.9 respectively. The approach starts with the system and problem identification and formulation respectively. This requires study of the given system model, KPIs7 , and research goals. Furthermore, at this step, the research problem is broken into subproblems, each targeting a specific part of the system. The achieved knowledge in the first step is employed in the second step for developing analytical models and simulators, which aim at simulating parts of behavior of the system that correspond to the problem of interest. The analytical models, simulators, and initial results in the prior arts are the three strong tools which can cross-validate each other. In other words, by simulating the system models used in prior arts, we can both investigate accuracy of our simulator and check feasibility of results presented in the prior arts. The same is true for numerical results from analytical models, and simulation results. The simulators have been developed by the 6 Other sources of diversity, e.g. code domain, could also be used for reliability, however, they are out of scope of this work. 7 KPI: Key performance indicator.

18


thesis author in Matlab, and are mainly available online8 . In response to RQ1, we have further validated our Matlab simulation results by comparing them against the NS-3 LENA simulator [71] results. The analytical modelings of experienced delay and success probability in IoT communications, which also determine the battery lifetime model, are mainly based on the queuing theory and stochastic geometry. The shortcoming of using stochastic geometry for reliability modeling comes from the fact that it just gives the average probability of success in communications, and hence, only the average values of other KPIs could be derived. Then, it cannot provide the distributions of KPIs, and hence, the variance and tail of reliability, delay, and battery lifetime distributions are missing for further analysis. After cross-validating our derived models and simulators, the approach proceeds towards identification of challenges and root causes in the system. Here, we mainly leverage inductive reasoning and hypothesis testing, as effective tools for figuring out the challenges and their root causes. The research approach subsequently employs this knowledge in designing solutions for the discovered challenges. In the design step, we mainly benefit from centralized and distributed optimization toolboxes. In the former, the optimization problem is explicitly programmed based on a priori information about system and traffic model. For example, in a part of answering to RQ2, we derive the optimized number of replicas per packet in terms of experienced path-loss in communications. In the latter, the optimization problem is programmed to work in an online and distributed manner, i.e. the decision is made by each device, based on the received feedback/input from the environment. The proposed solutions in the design step are evaluated using the derived analytical and simulation models. Most of our publications related to this thesis, e.g. [60, 65, 72], contain a section in which, the numerical and simulation results are compared in order to validate the developed analytical models, simulators, and solutions. The challenge identification and solution proposal procedures could be repeated several times until convergence to a stable solution which outperforms the alternatives. The research approach terminates by validating the proposed solutions and research results. The validation of proposed results is conducted by presenting early results in group seminars, and submitting the results for peer review to IEEE conferences and journals. Moreover, in order to get early-stage feedback, we have conducted inter-university collaboration, and three of the related publications to this thesis have been written in close collaboration with scholars from Aalborg University. Interest from the industry to our results and joint patenting of parts of our proposed solutions [73], could be seen as a fact that parts of our results have been also validated in the industry. Finally, we have made some simulators used is our research available online in order to let other researchers reproduce our results, and validate them.

8 The papers and respective simulators could be found in the thesis author’s profile on www.researchgate.net portal.

1.6. CONTRIBUTIONS

19

Input: System, KPIs, Goals

Review how the system works

Develop model, simulator

Infer challenges and their cause

Infer the challenges and their cause

Design solution for the challenges

Evaluate KPIs, compare with goals

Validate: Peer review

Output: protocol, architecture

Figure 1.8: Research approach: steps toward answering research questions.

Steps towards answering research questions Modeling

Inference

Statistics

Inductive reasoning

Queuing theory

Hypothesis testing

Design Optimization

Evaluation

Validation

Monte-Carlo simulation

Group discussion

Numerical analysis

Blind review

Stochastic geometry

Inter-university collaboration

Universityindustry collaboration

Make simulator available online

Figure 1.9: Research methods: the main used methods in different steps of the research approach.

1.6

Contributions

The main contributions of this thesis are described in this section.

20


1.6.1

Overview of contributions

An overview of our main contributions, which to the best of our knowledge are presented for the first time, includes: • A rigorous analytical model of reliability of communications in grant-free access IoT communications by considering potential overlapping of signals in time, frequency, and code domains (powered by stochastic geometry). • A multi-packet reception solution by leveraging potential CFOs of IoT devices, suitable for grant-free IoT communications, as well as RACH of LTE networks. • A light-weight self-organization solution for IoT devices, which enables them to adapt themselves to the environment, especially when they communicate over shared radio resources (powered by reinforcement learning). • Radio resource management algorithms for IoT serving over cellular networks, including LTE-A, LTE-M and NB-IoT, which significantly enhance network battery lifetime by allocating radio resources based on the expected battery lifetime of devices, as well as their delay budgets and other priorities.

1.6.2

Detailed contributions

This thesis summarizes the contributions presented in the following publications, which have been also appended to the second part of the thesis. • A. Azari and A. Bria, System and method for providing communication rules based on a status associated with a battery of a device, Nov. 2017, PCT/SE2017/050933 (filed). • A. Azari, G. Miao, C. Stefanovic, and P. Popovski, Latency-Energy Tradeoff based on Channel Scheduling and Repetitions in NB-IoT Systems, IEEE Global Communications Conference, Dec. 2018. • A. Azari, M. Masoudi, and C. Cavdar, Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks, submitted to IEEE Transactions on Communications, 2018. • A. Azari, P. Popovski, C. Stefanovic, and C. Cavdar, Grant-Free Radio Access for Cellular IoT, submitted to IEEE Transactions on Wireless Communications, 2018. • A. Azari and C. Cavdar, Self-organized Low-power IoT Networks: A Distributed Learning Approach, in IEEE Global Communications Conference, Dec. 2018. • A. Azari, M. Ozger, and C. Cavdar, Serving Non-Scheduled URLLC Traffic: Challenges and Learning-Powered Strategies, submitted to IEEE Communications Magazine, Aug. 2018.

1.6. CONTRIBUTIONS

21

For a more complete list of our publications, which have been leveraged in writing this thesis, one may refer to Appendix A. In the following, we categorize our contributions with regard to the research questions and list their related publications. 1.6.2.1

Energy efficient serving of mIoT: enhancing the legacy grant-based approach

In this subsection, we describe our contributions in answering RQ1. The existing 2G-4G cellular networks, as well as the newly standardized cellular-IoT networks like NB-IoT and LTE-M, follow the legacy grant-based radio access for communications [23, 60]. Connection establishment The first step in IoT connectivity over cellular networks is the connection establishment procedure, which is carried out over the physical random access channel. Then, we have investigated coexistence of IoT and non-IoT traffic over cellular networks, and investigate access rate and energy efficiency performance of existing random access procedures in dealing with the IoT traffic. Furthermore, random access resource provisioning for achieving a certain level of QoS is carried out using the presented analytical model, and validated using the system level simulations. Moreover, in cooperation with co-authors of [40, 74], a novel random access procedure with backward compatibility, called DERA, is proposed, and its performance in comparison with the legacy RA procedure is investigated. Finally, we investigate clustered-access, also known as capillary networking, to decrease the overall devices’ energy consumptions in resource reservation and scheduled data transmission. In this solution, local clusters of nodes are formed, devices use grant-free radio access to send data to the cluster-heads (CHs) with much lower energy than the case of direct communications with the BS, and CHs relay data to the BS. Our contributions here consist of battery lifetime optimized cluster forming, CH (re)selection, and inter-, intra-cluster communications. The following papers9 summarize our contributions in this field. • Paper 1: A. Azari, Energy-efficient scheduling and grouping for machinetype communications over cellular networks, Ad Hoc Networks, vol. 43, pp. 16–29, June 2016. • Paper 2: G. Miao, A. Azari, and T. Hwang, E 2 -MAC: Energy efficient medium access for massive M2M communications, IEEE Transactions on Communications, vol. 64, no. 11, pp. 4720–4735, Sept. 2016. 9 For paper 2, the thesis author has contributed in extending the initial ideas, writing of the paper, developing analytical models, and deriving numerical and simulation results. The thesis author acknowledges X. Chen and P. Zhang’s efforts on investigation of feasibility of paper 3, during their master thesis projects in 2013. The research in paper 4 and writing of the paper have been led by the thesis author. In paper 5 and 6, the author of this thesis has contributed in analytical analysis and proofreading.

22

CHAPTER 1. INTRODUCTION • Paper 3: A. Azari and G. Miao, Energy efficient MAC for cellular-based M2M communications, in IEEE Global Conference on Signal and Information Processing (GlobalSIP), Dec 2014. • Paper 4: A. Azari, M. I. Hossain, and J. I. Markendahl, RACH dimensioning for reliable MTC over cellular networks, in IEEE 85th Vehicular Technology Conference (Spring), May 2017. • Paper 5: M. I. Hossain, A. Azari, and J. Zander, DERA: Augmented Random Access for Cellular Networks with Dense H2H-MTC Mixed Traffic, in IEEE Global Communications Workshops, Dec. 2016. • Paper 6: M. I. Hossain, A. Azari, J. Markendahl, and J. Zander, Enhanced Random Access: Initial access load balance in highly dense LTE-A networks for multiservice (H2H-MTC) traffic, in IEEE International Conference on Communications, May 2017.

Resource provisioning and scheduling Towards enhancing the resource provisioning and scheduling procedures, we analytically model the battery lifetime of IoT devices in terms of the amount of allocated resources to IoT, and the resource scheduling policy over the allocated resources. This model is subsequently employed in devising uplink scheduling solutions for IoT, when single-carrier frequency division multiple access (SC-FDMA10 ) is used for multiple access. Also, low-complexity scheduling solutions with limited feedback requirement are also presented. Furthermore, we investigate the performance impact of serving IoT communications on QoS of non-IoT communications, and also on energy consumption of the access network. Our results indicate that improper resource provisioning for IoT communications not only decreases reliability and energy efficiency of these services, but also degrades experienced delay by other services, and increases energy consumptions of the access network. Finally, we investigate physical channel scheduling of NB-IoT systems, and analytically model how it impacts latency and energy efficiency of communications. We further leverage these models, and show that it is possible to design intelligent channel scheduling schemes for NB-IoT, in order to prevent performance degradation for devices close to the BS because of coverage extension for far away located devices. Our contributions in this field have been presented in the following patent and papers: • Patent (included in the appendix): A. Azari and A. Bria, System and method for providing communication rules based on a status associated with a battery of a device, Nov. 2017, PCT/SE2017/050933 (filed). • Paper 7 (paper A in the appendix): A. Azari and G. Miao, Network lifetime maximization for cellular-based M2M networks, IEEE Access, vol. 5, pp. 18927–18940, Sept. 2017. 10 SC-FDMA

is used in LTE, LTE-A, and LTE-M system for uplink transmission [75].

1.6. CONTRIBUTIONS

23

• Paper 8: A. Azari and G. Miao, Lifetime-Aware Scheduling and Power Control for M2M Communications in LTE Networks, in IEEE Vehicular Technology Conference (Spring), May 2015. • Paper 9: A. Azari and G. Miao, Lifetime-aware scheduling and power control for cellular-based M2M communications, in IEEE Wireless Communications and Networking Conference, March 2015. • Paper 10: A. Azari and G. Miao, Battery Lifetime-Aware Base Station Sleeping Control with M2M/H2H Coexistence, in IEEE Global Communications Conference, Dec. 2016. • Paper 11: A. Azari and G. Miao, Fundamental tradeoffs in resource provisioning for IoT services over cellular networks, in IEEE International Conference on Communications, May 2017. • Paper 12 (paper B in the appendix): A. Azari, G. Miao, C. Stefanovic, and P. Popovski, Latency-Energy Tradeoff based on Channel Scheduling and Repetitions in NB-IoT Systems, IEEE Global Communications Conference, Dec. 2018. 1.6.2.2

Ultra-energy-efficient serving of mIoT: a grant-free access approach

In this subsection, we describe our contributions in answering RQ2. Network analysis/design with grant-free access In answering RQ2, which seeks a close-to-zero energy consumption profile for IoT communications, we investigate the strengths and shortcomings of the grant-free radio access. Grant-free radio access was first introduced for IoT communications over the unlicensed spectrum [18]. Regarding the fact that there is no prior research presenting a comprehensive analysis of the grant-free radio access IoT systems, we provide a rigorous performance analysis of these systems. This analysis assumes large-scale IoT networks serving multiple IoT traffic types with heterogeneous communications’ characteristics, and heterogeneous location distribution processes. In this model, for the first time in literature, the partial overlapping of short IoT packets in time, frequency and code domains has been considered. The simulation results confirm tightness of the derived performance expressions from the analytical analysis. For enabling the grant-free radio access IoT connectivity over the licensed spectrum, we leverage the derived expressions and present reliability-constrained cost-optimized resource provisioning strategies. This resource provisioning problem includes both the time-frequency radio resources and density of the BSs. Furthermore, we present reliability-constrained battery lifetime-optimized operation control strategies for IoT devices. Next, we investigate the scalability of the access

24


network, i.e. how the amount of provisioned radio and BS resources, which determine the cost of the access network, must be scaled with scaling the number of active devices, amount of external interference, and the required level of reliability in communications. Finally, we investigate the point up to which, IoT devices can adapt themselves to an increase in the level of external interference by increasing their transmit power and number of replica transmissions. Our contributions in answering to RQ2 have been summarized in the following papers11 : • Paper 13 (paper C in the appendix): A. Azari, M. Masoudi, and C. Cavdar, Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks, submitted to IEEE Transactions on Communications, 2018. • Paper 14: A. Azari and C. Cavdar, Performance Evaluation and Optimization of LPWA IoT Networks: A Stochastic Geometry Approach, in IEEE Global Communications Conference, Dec. 2018. • Paper 15: M. Masoudi, A. Azari, E. A. Yavuz, and C. Cavdar, Grant-free Radio Access IoT Networks: Scalability Analysis in Coexistence Scenarios, IEEE International Conference on Communications, Dec. 2018. Enhanced transmission/reception protocols In order to resolve the main drawback of the grant-free access scheme, i.e. collision in the high traffic load regime, we enhance the basic grant-free radio access scheme. The proposed scheme benefits from a path-loss adaptive replica transmission control, and outperforms both the grant-based and basic grant-free data transmission schemes in a wide range of traffic loads. We further develop a SIC-based asynchronous receiver which benefits from the CFO of low-cost IoT devices for further collision resolution. Finally, we identify the operating regions, in terms of traffic load, in which our proposed scheme outperforms the competing grant-based data transmission schemes. These contributions have been also presented in the following papers: • Paper 16 (paper D in the appendix): A. Azari, P. Popovski, C. Stefanovic, and C. Cavdar, Grant-Free Radio Access for Cellular IoT, submitted to IEEE Transactions on Wireless Communications, October 2018. • Paper 17: A. Azari, P. Popovski, G. Miao, and C. Stefanovic, Grant-free radio access for short-packet communications over 5G networks, in IEEE Global Communications Conference, Dec. 2017. 11 The research and writing for paper 13 have been led by the thesis author. In paper 14, the thesis author has partially contributed in development of the baseline knowledge, analytical analysis, and writing of the paper.

1.6. CONTRIBUTIONS

25

Distributed coordination of grant-free access IoT networks In the absence of control over the spectrum, the probability of success in data transmission decreases by an increase in the traffic load due to overlapping of packets in time/frequency/code domain. This problem is more critical when we consider adversarial interference scenarios, in which a channel becomes blocked for a period of time, e.g. due to misbehaving of another node. Furthermore, in case of a disruptive change in the environment, e.g. parking a car in the basement with extreme path-loss to the BS, the established IoT connectivity service becomes blocked. By considering the above-mentioned problems, one sees that in absence of centralized assistance from the network, we need intelligence in IoT devices to let them adapt themselves to the changes in the environment. While the literature in machine learning is mature, we need to develop a learning approach to bring more reliability to communication at the cost of little increase in complexity, expense, and energy consumption of IoT devices. Towards this end, we follow the multi-arm bandit (MAB) learning and modify its reward function to incorporate both success rate in communications and the consumed energy for achieving such a success rate. To analyze the usefulness of the proposed learning-powered scheme, we apply it to the LoRaWAN in which, transmit power, sub-channel, and spreading factor are the tunable communications parameters. Our performance evaluations show that the proposed scheme can significantly improve performance of IoT devices, and after a limited number of trials, IoT devices will achieve a performance level close to the centralized solution. Furthermore, we investigate the case in which the feedback channel is also vulnerable to intended or unintended interference, and devise a respective learning algorithm which trades the convergence time for achieving reliability against interference. Our contributions in answering to the need for distributed coordination of grantfree access IoT networks have been summarized in the following paper: • Paper 18 (paper E in the appendix): A. Azari and C. Cavdar, Selforganized Low-power IoT Networks: A Distributed Learning Approach, in IEEE Global Communications Conference, Dec. 2018. 1.6.2.3

Ultra-reliable serving of cIoT: a hybrid access approach

In this subsection, we describe our contributions in answering RQ3. In description of RQ3, we observed that the need for a large amount of radio resources for assuring reliability of infrequent non-scheduled cIoT communications motivates us to consider an overlay transmission approach for them. Then, we investigate usefulness of a hybrid orthogonal/non-orthogonal multiple access (HMA) for serving non-scheduled and scheduled traffic together. But RRM for the orthogonal RRM is itself a complex problem, and hence, the RRM for HMA, which requires specifying the shared/dedicated resources and the scheduling rules over the shared ones, will be much more complex. Furthermore, to comply with the cIoT delay requirements, the RRM problem is needed to be solved in very short timescales. Motivated by

26


the recent advances in ML, we investigate feasibility of solving this problem using the state-of-the-art ML approaches. The main challenge in utilizing ML in design of RRM consists in large dimensionality and complexity of the RRM problem. The heterogeneity of services, QoS requirements, and diverse set of available resources make the RRM a very complex problem. This complexity has a significant impact on the convergence time of ML algorithms. Towards addressing these issues, we investigate a distributed ML-powered RRM solution, including intelligent resource provisioning, scheduling, and utilizing modules in the RAN control center, edge, and device, respectively. The proposed solution enables spectrum efficient yet reliable coexistence management of scheduled and non-scheduled traffic. We further investigate different sources of uncertainty in proactive resource scheduling for the non-scheduled traffic, how they affect our decisions, and the amount of radio resource which must be sacrificed for compensating the performance loss due to such uncertainties. Our contributions in answering to RQ3 are summarized in the following paper: • Paper 19 (paper F in the appendix): A. Azari, M. Ozger, and C. Cavdar, Serving Non-Scheduled URLLC Traffic: Challenges and LearningPowered Strategies, submitted to IEEE Communications Magazine, Aug. 2018.

1.7

Thesis Outline

Chapter 2 and chapter 3 are dedicated to serving mIoT traffic, in which our aim is providing scalable IoT connectivity for a massive number of devices in an ultraenergy-efficient way. More specifically, chapter 2 presents our evolutionary solutions, including enhanced connection establishment and scheduling procedures, for serving mIoT traffic over cellular networks. Chapter 3 presents our revolutionary solutions towards serving mIoT traffic over cellular networks, including the grantfree radio access scheme. Chapter 4 is dedicated to presentation of state-of-the-art and our contributions towards serving cIoT traffic with stringent delay-reliability requirements over cellular networks. Concluding remarks and future research directions are given in chapter 5. Finally, the included papers are appended to the end of the dissertation.

Chapter 2

IoT Connectivity over Cellular Networks: Challenges, solutions, and key results The1 continuing growth in demand for IoT connectivity has encouraged different telecommunications industries to investigate evolutionary and revolutionary radio access technologies for IoT communications [20]. To comply with the rapid changes in the IoT world, several research projects have been initiated by the 3GPP LTE in order to support large-scale IoT communications over existing cellular infrastructure. The development of LTE standardization for massive IoT communications has been initiated in release 12, and has continued in release 13 [25]. In LTE release 13, LTE category M (LTE-M) and narrow-band IoT (NB-IoT) systems have been introduced. The existing 2G-4G cellular networks, as well as the newly standardized cellular-IoT networks like NB-IoT and LTE-M, follow a grant-based radio access for communications [23, 60]. In the following, we investigate the IoT connectivity solutions which leverage the legacy grant-based radio access, figure out the challenges, present some of the proposed solutions, and show the key results. The concluding remarks are given at the end of this section.

2.1

Serving IoT Traffic over Shared Systems

In this section, we instigate the LTE-A air interface and analyze its limitations in supporting massive IoT traffic in an energy-efficient way. One must note that this part of the air interface remains in essence the same for new releases, and hence, this investigation is of crucial importance for standardization towards 5G [77]. Based on the LTE-A air interface, IoT devices pass the physical random access channel (PRACH) and send their uplink scheduling requests to the BS over the physical 1 Parts

of this chapter have been previously published in the author’s licentiate thesis in 2016

[1].

27

CHAPTER 2. IOT CONNECTIVITY OVER CELLULAR NETWORKS: CHALLENGES, SOLUTIONS, AND KEY RESULTS

28

IoT device

BS

Core network

(Wake up) Data gathering (Turn radio on)

Cell Info

RA response PRC connection request

Connection establishment

Reporting Period

Random Access Request

PRC connection setup

Authen tication

Acknowledgement

Data transfer

Data transfer

(Sleep) (Wake up)

Figure 2.1: Access protocol exchanges in LTE/LTE-A [76]

uplink control channel (PUCCH) [75]. Next, BS carries the scheduling out and transmits scheduling grants to users over the physical downlink control channel (PDCCH). Scheduling is the process performed by the BS to assign time-frequency resources available in the physical uplink shared channel (PUSCH) to user devices [76]. For the legacy human-oriented communications traffic, the above-described connectivity procedure performs well. This is due to the fact that human-oriented communications consist of a limited number of long-lived communications sessions, such as voice and web streaming. However, by introduction of IoT communications to cellular networks, the performance of this connectivity procedure is questionable, especially when it comes to scalability and energy efficiency of communications. Let us divide the connectivity procedure into two parts, i.e. connection establishment, in which user shows its presence to the BS, and resource scheduling, in which a set of radio resources are allocated to the device. Once the impacts of IoT communications on the connection establishment and resources scheduling procedures are studied, we are able to investigate the mutual impact of serving IoT traffic

2.1. SERVING IOT TRAFFIC OVER SHARED SYSTEMS

29

over a limited set of radio resources on the QoS of non-IoT traffic, as well as on costs of the access network. We call the latter as the resource provisioning problem, as the derived impacts shed light on the optimized resource provisioning problem. Towards answering the above-raised concerns, in the following, we present a list of questions that are needed to be answered, when IoT communications are served through the LTE-A air interface. • Connection Establishment: – RQ1.1: How do the access rate and energy consumption of IoT devices in the connection establishment procedure scale with the number of devices? – RQ1.2: How the connection establishment procedure can be improved in order to increase the access rate and energy efficiency of IoT communications? • Resource Scheduling: – RQ1.3: How the radio resource schedulers of cellular networks could be improved to care about battery lifetime of IoT devices, as well as their priorities and delay budgets? • Resource Provisioning: – RQ1.4: What are the consequences of allocating a part of radio resources to the IoT traffic on (i) QoS of non-IoT traffic, (ii) energy consumption of the access network; and (iii) battery lifetime of IoT devices?

2.1.1 2.1.1.1

Proposed modeling, solutions, and key results Scalability and energy efficiency of the connection establishment procedure

In response to RQ1.1, we need to investigate performance of the random access channel of LTE systems. In this thesis, energy efficiency of an IoT solution is investigated by analyzing the achieved battery lifetime for IoT devices served by that solution. Furthermore, scalability is investigated by evaluating the access rate of IoT devices, which is defined as the rate of success in connection establishment after K trials. Let us start by investigating the battery lifetime. For most reporting IoT applications, the energy consumption of devices could be seen as a regenerative process, where the regeneration point is at the end of each successful data transmission epoch [41]. Then, the expected battery lifetime can be approximated as the product of the average reporting period (TRep ) and the ratio between the remaining energy and the average energy consumption in each reporting period, i.e. Lifetime ≈

E0 TRep . ESleep + NConEst EConEst + EDataTx + ESched + EOthSig

(2.1)

30


In this expression, E0 , ESleep , NConEst , EConEst , ESched , EDataTx , and EOthSig represent the stored energy, average energy consumption in the sleep mode in one reporting period, average number of trials for a successful connection establishment, average energy consumption in a connection establishment trial, average energy consumption in receiving data transmission/reception grant, average energy consumption in data transmission/reception, and average energy consumption in other signaling, e.g. acknowledgment, respectively. One must note that the energy consumption in the transmit mode includes both the transmit power sent to the power amplifier, as well as the power consumed in electronic circuits. The modeling of each energy consumption source is technology-dependent, and is different from LTE, to LTE-A and NB-IoT. The interested reader is referred to [41, 60] for the detailed modelings. From (2.1), it is clear that battery lifetime decreases by an increase in NAccRes ≈ Pr1Suc , i.e. by a decrease in the probability of successful connection establishment over the RA procedure. Then, investigation of the access rate also sheds lights on the energy efficiency and battery lifetime problems. The probability of successful connection establishment over RA is itself a function of traffic arrival model and the amount of time/frequency/code domain resources allocated to the RA. The former, describes the intensity of arrival of connection requests, and can be modeled by a Poisson point process (PP) for semi-regular reporting applications, or Switched Poisson Process (SPP) for mixed event-driven and semi-regular reporting applications. In [78], we have used the latter traffic model, as depicted in Fig. 2.2. In this figure, one sees that a massive arrival incurs some transient states in which, we deal both with unsuccessful nodes which retry for connection establishment and the new nodes which send connection requests for the first time. In this figure, M represents the number of RA resources in frequency/code domain, called preamble, TRA represents the average time between two scheduling of RA opportunities, µh , µt , and µl represent the average durations of high, transient, and low traffic arrival modes, respectively, and λh , λt , and λl represent the average traffic arrival rate in high, transient, and low arrival modes, respectively. Considering this arrival process, our work in [78] presents analytical models which describe the access rate in terms of number of preambles in time/frequency domain, the average time between two consecutive RA epochs, maximum number of allowed retransmissions before outage, and the traffic model. Here, we summarize one part of the numerical results presented in [78]. Fig. 2.3 represents the required resource provisioning for target access rate of 0.99 over the RA procedure. Collision in selecting the RA resource has been considered as the only cause of failure, and collided devices retry with probability 1 in the subsequent RA opportunities. In this figure, the regions covered by the yellow color specify the (M, TRA ) pairs, for which, the access rate requirement is satisfied. One observes that the boundary M and TRA values that satisfy the access rate requirement construct a convex curve. Our further analysis in [78] for different access rate requirements reveals that the slope of this curve is inversely proportional to the access rate requirement. In other words, by an increase in the access rate requirement, the substitution rate of M with TRA increases. Then, one


31 time

/

a) RACH provisioning

RACH resources

!"

PUSCH/PUCCH resources

Transient states

#%

#$

#0

time &' (((((((((((((&) (((((((((((((((&*

(((((((((&*+) (((((((&*+, ((((((((((((&-

.0

.%

.$

b) A traffic arrival model for mixed event-driven and semi-regular reporting applications

TRA (× 1 msec)

Figure 2.2: RA resource provisioning and traffic arrival modeling 40

20

0 0

50

100

150

200

250

300

350

400

1 msec)

M (number of preambles) 40

Figure 2.3: Access rate in terms of number of preambles in each RA, i.e. M , and the time interval between two consecutive RA epochs, i.e. TRA . Simulation parameters could be found in [78]. (Reprinted from [78], ©2017 IEEE, reused with permission.)

can conclude that for scenarios in which a high access rate is required, the target access rate could not be achieved only by decreasing the interval between two successive occurrences of RA resources. Thus, introducing frequency/code domain diversity for achieving ultra-reliability is inevitable. We have further validated our analytical/simulation results by comparing our results against the NS3 LENA simulator [71] in [78, Section VI]. 2.1.1.2

Enhancing the contention resolution capability

While in the previous subsection we have treated collisions over each random access resource as a failure for all devices involved in the collision, one may think of improving the contention resolution capability of receivers. One potential solution consists in finding dimension of collision, i.e. number of involved nodes in the collision, and then, responding to the contending nodes with a batch of resources. Then, each node will randomly select a subset of these resources to decrease the probability of collision in the subsequent transmissions. The performance of such

32

CHAPTER 2. IOT CONNECTIVITY OVER CELLULAR NETWORKS: CHALLENGES, SOLUTIONS, AND KEY RESULTS Periodogram Power Spectral Density Estimate

Power/frequency (dB/Hz)

20 0

∆ f1 =-150 Hz

∆ f3 =100 Hz

-20

∆ f2 =-50 Hz

-40 20 -60

20

0

0

-20

-20

-80 -40

-40 0.1

0.2

0.3

19.7

-100 2

4

6

8

10

12

19.8 14

19.9 16

18

Frequency (kHz)

Figure 2.4: Leveraging CFO for collision detection and resolution. This figure represents the output of periodigram, where the input is a demodulated signal complex containing three frequency components. (Reprinted from [79], submitted to IEEE for peer review.)

a scheme is highly dependent on the accuracy of collision dimension estimation. In [63, 79], we have proposed to leverage the CFOs of involving devices for collision dimension estimation. In this scheme, the received signal content is checked by periodogram, in order to find the remaining CFOs in the signal content. Fig. 2.4 represents the output of the periodogram for a received signal, which itself consists of 3 signals with 3 different CFOs2 . When the involved CFOs are very close to each other, one may also benefit from the different time-of-arrival of each packet at the receiver, as we have leveraged in the DERA3 scheme, as follows. In [40, 74], we have leveraged the fact that different nodes may experience different transmission delays in communication with the BS, or may have different timing offsets with regard to the reference time of the BS. Thus, we have proposed an amendment to the RA procedure in the LTE-A, called DERA, in which, the random access response is transmitted to the involved nodes in the collision, and one of the detected time offsets by the receiver is also included in the response message. Hence, all of the colliding nodes will receive the response, but only the one whose transmission delay matches with the estimated time offset by the receiver will continue the random access procedure. The interested reader may refer to [40, 63] for a detailed description of the proposed random access enhancement schemes.

2 This receiver is also of interest for grant-free IoT communications, and is discussed in details in section 3 of this thesis. 3 DERA: Delay estimation based random access.

2.1. SERVING IOT TRAFFIC OVER SHARED SYSTEMS 2.1.1.3

33

IoT traffic scheduling

Now, we have investigated the connection establishment procedure in the previous subsection, and proceed to the resource scheduling procedure, i.e. RQ1.2. Scheduling, in essence, is not part of the standardization and is implemented by each vendor. However, signaling for scheduling is part of the standardization work, and each scheduling scheme complies with the specified signaling in the standards. Resource scheduling includes prioritizing data transmission/reception requests based on the KPIs of interest, and allocating a set of time/frequency radio resources to each device in order to ensure its reliable communications. In literature, delay budget is the main used KPI for prioritizing traffic, and usually, IoT traffic is served over the remaining resources from non-IoT traffic [24]. Regarding the fact that every joule counts for battery lifetime of IoT devices, and resource scheduling has a vital role in energy consumption of devices, here we investigate battery lifetimeaware scheduling solutions, and evaluate their impacts on the individual, as well as network, battery lifetime. In Paper A, included in the appendix, we analytically explore the uplink IoT traffic scheduling problem over a limited set of radio resources by considering delay and battery lifetime as the KPIs. Then, the impacts of resource pool size, coexisting traffic types, and scheduling policy on delay and battery lifetime performance of IoT devices are investigated. As a special case, we extend our proposed battery lifetime-aware scheduler to the SC-FDMA systems. Finally, we investigate low-complexity scheduling solutions with limited feedback requirement. In the following, we present how complex is the scheduling problem for SCFDMA systems, and how our proposed scheduler in [41] (Paper A) looks like. IoT traffic scheduling over SC-FDMA systems Thanks4 to the reduced peak-to-average-power-ratio (PAPR) property of SC-FDMA, it is a favorite multiple access scheme for energy-efficient uplink communications. In scheduling of SC-FDMA systems, the maximum number of clusters of adjacent subcarriers allocated to a device is denoted by G, where G is usually equal to 1 for IoT communications5 . Furthermore, the transmit power of each device over all granted subcarriers is the same [51], and subcarriers are assigned in units of chunks, each containing M subcarriers [82]. In the following, we consider the scheduling problem at time t, when a set of nodes, denoted by A with cardinally |A|, have passed the connection establishment procedure, and seek grant for the uplink transmission. As the scheduling procedure is needed to be battery lifetime-aware, first we model the expected battery lifetime of individual devices, and the expected network battery lifetime. The remaining battery energy at time t for node i is denoted 4 The contents of this subsection, especially the equations, have been recycled from Section IV of Paper A. 5 The constraint on G is called the contiguity constraint [80]. As PAPR increases in G [81], G = 1 is preferred for IoT applications [41].

34


by Ei (t), the average packet size by Di , and the power consumption in data transmission by ξPi + Pc . In this expression, Pc denotes the circuit power, ξ the inverse of power amplifier (PA) efficiency, and Pi the transmit power for reliable data transmission. Using (2.1), the expected battery lifetime of node i at time t is approximated as follows: Li (t) ≈

Ei (t) Ti . Esi + Edi

(2.2)

In this expression, Edi is the average consumed energy per reporting period in data transmission, and is modeled as Edi = [Pc + ξPi ]Di /Ri , Ri denotes the average data transmission rate, Ti denotes the average time-length of each reporting period, and finally, Esi denotes the average static energy consumption per reporting period for data gathering, processing, and etc. Given the battery lifetime model of devices, one may define the network battery lifetime as a function of individual lifetimes. For example, when even missing one device deteriorates the coverage and performance, the shortest individual lifetime (SIL) can be defined as the battery lifetime, i.e. Lsil net = mini Li . In contrast, the longest individual lifetime (LIL) might be of interest in some reporting IoT applications [41]. In [41], we have defined a broad range of network battery lifetime models, and optimized the scheduling problem for them. Here, we stick to the SIL lifetime definition. Now, let us come back to the scheduling problem. Denote the set and number of available resource chunks as C and |C|, respectively. Each chunk consists of M subcarriers in the frequency domain, and spans over τ seconds in the time-domain. Note that if the number of time/frequency resource elements is not sufficient to schedule A at once, a subset of A is scheduled, and the others are left for the next scheduling epoch, i.e. τ seconds later. One can approximate the effective SINR6 of an SC-FDMA symbol as the average SINR over the set of subcarriers. This is due to the fact that each data symbol is spread over the whole bandwidth in SC-FDMA [51]. Now, we can derive the achievable data rate for node i, as R(Ci , Pi ) = |Ci |M Sv (ηi ). In this expression, Ci represents the set of allocated chunks P hei to node i, |Ci | represents the cardinality of Ci , ηi is the SINR and is equal to |Cii|M , and Sv (x) represents the achieved data rate over one subcarrier for a SINR level of ηi , as described in [41]. Moreover, hei is defined as [51] [ ∑ [N0 + Ij ]M w ] hei = |Ci |/ , hji j∈Ci

and is interpreted as the effective channel gain-to-noise-plus-interference ratio for node i. Here, hji is the channel gain of node i over chunk j, which is assumed to be constant in (t, t + τ ) [51], Ij is the PSD7 of interference on chunk j, N0 is the PSD of noise, and w is the bandwidth of each subcarrier. Thus, using (2.2), the 6 SINR: 7 PSD:

Signal to interference and noise ratio. Power spectral density.


35

expected battery lifetime of node i at time t + τ , either to be scheduled at t or not, is approximated as a function of Pi and Ci as follows: Li (t + τ ) =

Ei (t + τ ) − [1 − θi ]Edi Ti , Esi + Edi

(2.3)

where Ei (t + τ ) = Ei (t) − τ Pc [1 − θi ] −

[ ] Di Pc + ξPi θi . R(Ci , Pi )

(2.4)

In this expression, θi , the indicator function, is 1 if node i is scheduled at t, and 0 otherwise. Furthermore, [1 − θi ]Edi models the expected energy consumption from t + τ till the successful transmission, i.e. the regeneration point. Now, we have all the necessary expressions to formulate the scheduling problem at time t as an SIL network battery lifetime maximization problem, as follows: maximizeCi ,Pi ,θi Lsil net (t + τ ) ∑ |Ci | ≤ |C|, subject to: C.2.5.1:

(2.5)

i∈A

C.2.5.2: Ci : contiguous ∀i ∈ A, C.2.5.3: Ci ∩ Cj = ∅ ∀i, j ∈ A, i ̸= j, C.2.5.4: θi Di /R(Ci , Pi ) ≤ τ ∀i ∈ A, C.2.5.5: Pi ≤ Pmax ∀i ∈ A, C.2.5.6: [1 − θi ]Qi = 0 ∀i ∈ A. In this optimization problem, ∅ is the empty set, and Qi ∈ {0, 1} is a binary parameter represents priority of traffic due to the delay budget. The scheduler schedules node i with the highest priority for Qi = 1, and schedules using the remaining resources from the scheduling of high-priority traffic otherwise. Now, let us interpret the constraints. The C.2.5.1 comes from the limitation of radio resources, C.2.5.2 comes from the contiguity requirement, C.2.5.3 ensures that each resource will be allocated to at most one device, C.2.5.4-5 assure the successful transmission of data over granted resources, and C.2.5.6 assures priority of high-priority demanding traffic. It is clear that the above optimization problem is not a convex optimization problem, especially due to the contiguity constraint [51]. One may reformulate this problem as a pure binary-integer programing problem. However, using the results from [82], the complexity of search over all feasible resource allocations for an SC∑|A| (|A| ) (|C|−1 ) FDMA system will be i=1 i i! i−1 . Thus, this approach is clearly complex, especially for scheduling massive IoT communications. In the sequel, we present a low-complexity suboptimal solution for the above scheduling problem. The proposed scheduler We propose to break the scheduling problem in (2.5) to two subproblems based on the priority of devices. Then, the first subproblem satisfies the minimum resource

36


requirement for the set of high-priority nodes, called Ad , while the second subproblem allocates resources to all other nodes based on their impacts on the network lifetime. It is clear that the former is a frequency-domain scheduling problem, i.e., it consists of frequency-domain scheduling for nodes which cannot wait anymore. Furthermore, the second subproblem is a time/frequency scheduling problem, i.e., the remaining nodes might be scheduled using the remaining resources or be postponed to future scheduling epochs. In the second subproblem, the scheduler selects node with the highest impact on the network lifetime, allocates one unit of remaining resources to it, and continues this procedure until radio resources are finished. The proposed scheduler, presented in Algorithm 1, contains two steps, each of them are responsible for one of the above-described subproblems. In Algorithm 1, we call F(·), P(·), and ExpAlg functions, which have been defined in [41, Section IV] in details. The following section presents some key results achieved in the simulation of the proposed scheduler for LTE-A networks. In the simulation setup, the deployment of IoT devices and their traffic model follow the proposed models in [81, Annex A] for smart metering applications, and have been reflected in [41] (Paper A). Fig. 2.5 represents the probability density function (PDF) of battery lifetimes of IoT devices using the following scheduling schemes: Scheme 1, in which time- and frequency-domain schedulers aim at maximizing the SIL network lifetime; Scheme 2, which consists of a round-robin (RR) scheduler for time-domain scheduling and a low-complexity SIL network lifetime maximizing scheduler for frequency-domain scheduling; Scheme 3, in which time- and frequency-domain schedulers aim at maximizing the LIL network lifetime; Scheme 4, which consists of round-robin schedulers for time- and frequency-domain scheduling; and Scheme 5, which consists of a channel-aware scheduler for time-domain scheduling and a round-robin scheduler for frequency-domain scheduling. One sees in Fig. 2.5 that the first battery drain for scheme 1, i.e. the SILnetwork-lifetime maximizing scheduler, happens much later than the ones of the benchmarks, i.e. scheme 4 and 5. In a similar way, one sees that the last battery drain for scheme 3, i.e. the LIL-network-lifetime maximizing scheduler, happens much later than the benchmarks. Furthermore, one observes that scheme 1 benefits from a compact shape, which, in turn, indicates that the individual lifetimes of devices have been distributed in a limited time interval. In other words, the variance of individual battery lifetimes using the SIL-network-lifetime scheduling is the least among other schemes. Then, we can conclude that the maintenance costs for IoT networks may be reduced using SIL-network-lifetime scheduling, as their batteries can be replaced almost at the same time [12, 41]. Fig. 2.6 represents the detailed network lifetime comparisons of the investigated scheduling schemes. Here, it is evident that scheme 1 achieves a SIL network lifetime which is 2.4 times higher than scheme 4, and 3 times higher than scheme 5. Also, we observe that scheme 2, i.e. the low-complexity sub-optimal SIL-network-lifetime maximizing scheduler, outperforms the baseline schemes 4 and 5.


37

Algorithm 1: The SIL-network-lifetime maximizing scheduler for SC-FDMA. (Reprinted from [41], ©2017 IEEE, reused with permission.)

1

2

Initialization; - Define Ad , where i ∈ Ad if Qi = 1; - Define Acd as A \ Ad ; - 0 → θi and ∅ → Ci , ∀i ∈ A ; Step 1; - A → Atd ; - while Atd is non-empty do - arg minj∈Atd Lj (t) → j ∗ ; - Atd \ j ∗ → Atd , 1 → θi ; ( ) - while Dj ∗ /R Cj ∗ , Pmax ) > τ do - ExpAlg(C, Cj ∗ , G) → c∗ , c∗ ∪ Cj ∗ → Cj ∗ , C \ c∗ → C; - If c∗ = ∅, then 0 → θi , Ad \ j ∗ → Ad , Cj ∗ ∪ C → C, exit the loop; Step 2; - Ad ∪ Acd → H; - while C and H are non-empty do - j ∗ = arg minj∈H F(Ei (t), Edi , Esi , Ti , Ci , θi ); - if θj ∗ ̸= 1 then - 1 → θj ∗ ; ( ) - while Dj ∗ /R Cj ∗ , Pmax > τ do - ExpAlg(C, Cj ∗ , G) → c∗ , c∗ ∪ Cj ∗ → Cj ∗ , C \ c∗ → C; - If c∗ = ∅, then 0 → θi , Cj ∗ ∪ C → C, H \ j ∗ → H, exit the loop; else -

Di [Pc +ξP (Cj ∗ )] R(Cj ∗ ,P (Cj ∗ ))

→ B;

- ExpAlg(C, Cj ∗ , G) → c∗ ; - if c∗ = ∅ then - H \ j ∗ → H; else - c∗ ∪ Cj ∗ → Cj ∗ , C \ c∗ → C; - If

Di [ξP(Cj ∗ )+Pc ] R(Cj ∗ ,P(Cj ∗ )) ∗

> B, then Cj ∗ \ c∗ → Cj ∗ , C ∪ c∗ → C,

H \ j → H; 3 4

P(Ci ) → Pi , ∀i ∈ A; return Pi , θi , Ci , ∀i ∈ A

2.1.1.4

Radio resource provisioning for IoT traffic

After investigating the connection establishment and resource scheduling procedures, we are ready to investigate the resource provisioning problem, i.e. RQ1.3,

CHAPTER 2. IOT CONNECTIVITY OVER CELLULAR NETWORKS: CHALLENGES, SOLUTIONS, AND KEY RESULTS PDF

38

0.11

scheme 1

PDF

0 0.11

1e4

1e5

1e6

1e7

1e8

1e4

1e5

1e6

1e7

1e8

1e4

1e5

1e6

1e7

1e8

1e4

1e5

1e6

1e7

1e8

1e4

1e5

1e6

1e7

1e8

scheme 2

PDF

0 0.11

scheme 3

PDF

0 0.11

scheme 4

PDF

0 0.11

scheme 5

0

Time (× Ti)

7

104

107

3.5

6

3

LIL network lifetime SIL network lifetime

5

2.5

4

2

3

1.5

2

1

1

0.5

0

Sch. 1

0 Sch. 2

Sch. 3

Sch. 4

Sch. 5

SIL network lifetime ( Ti)

LIL network lifetime ( Ti)

Figure 2.5: Empirical PDF of individual lifetimes using different scheduling schemes. (Reprinted from [41], ©2017 IEEE, reused with permission.)

6

Scheduling Schemes

Figure 2.6: SIL/LIL network lifetime comparison. (Reprinted from [41], ©2017 IEEE, reused with permission.)

when IoT and non-IoT traffic are assigned separate sets of resources over PRACH and PUSCH resources, see Fig. 2.7. In [83], we develop an analytical model which involves access rate and energy consumptions analysis of IoT devices, experienced delay and spectral efficiency analysis of non-IoT traffic, and energy consumption of the access network in serving the mixed IoT/non-IoT traffic. Our analyses figure out the ways in which spectral efficiency, energy saving for the access network, and QoS of the non-IoT traffic could be traded to extend battery lifetimes of IoT

frequency


‫( ܯ‬sec) ‫ܯ‬ூ

(sec)

39

(Hz) (Hz)

time

PRACH resource (ŶŽŶͲ/Žd) PRACH resource (IoT)

PUSCH resource (ŶŽŶͲ/Žd) PUSCH resource (IoT)

Figure 2.7: Resource provisioning for IoT

devices by compromising on the level of provisioned radio resources. For a full list of insightful observations derived from the performance tradeoff analysis, one may refer to [83]. In the following, one of the key findings of this paper is discussed. To ease understanding of couplings among lifetime for IoT, delay for non-IoT, consumed energy for the BS (in the single-cell scenario), and spectral efficiency of radio resources, in Fig. 2.8 the optimized operation points have been depicted over a 2D view of the energy consumption of the BS. The design parameters are the ratio of allocated PRACH and PUSCH resources to the IoT traffic. In the background, different colors refer to different energy consumption levels. The colors, ranged from yellow to dark blue, are indicating high to low energy consumption regimes respectively. Let us consider the minimum energy consumption for the BS as the reference operation point. Fig. 2.8 shows that the average consumed energy by the BS, energy efficiency of IoT communications, and delay of non-IoT communications increase by an increase in the amount of allocated radio resources to the IoT traffic. The improvement in energy efficiency of IoT communications comes from the fact that the access rate over PRACH and success probability in data transmission over PUSCH increase in the amount of provisioned radio resources. However, this improvement for IoT communications is achieved at the cost of increasing energy consumption of the access network, as BSs need to stay longer in the non-sleep mode for serving potential IoT traffic. 2.1.1.5

Access network provisioning for IoT traffic

After investigating the radio resource provisioning problem for IoT communications, here we proceed by investigating the impact of density/activity of BSs of the access network on QoS of IoT communications, i.e. RQ1.4. While in the path from 3G to 4G wireless networks, the main driver was delivering higher data rates to users, for the first time energy efficiency in deployment and operation of cellular networks was also taken into account seriously as a design objective [55]. BS sleeping is a major proposed/standardized energy saving technique for the wireless access networks [55]. In this scheme, once there is no user to be served or the request demand is low, the BS may go to the low energy consumption mode. Furthermore, by a drastic change in the traffic arrival, e.g. from midnight to morning, some micro

40


β: PUSCH alloc. to IoT

1

Max energy efficiency for IoT

0.8 EE, EC, and ED increase; SE decrease

0.6

Min BS energy consumption

10 -2

EC and SE increase

EC increases; EE and ED decrease

0

0.1

0.2

0.3

0.4

0.5

0.6

α : RACH alloc. to IoT Min Delay for non-IoT

0.7

0.8

0.9

1

Max uplink spectral efficiency (non-IoT+IoT)

Figure 2.8: Performance tradeoffs. EC, EE, ED, and SE refer to energy consumption of BS, energy efficiency of IoT communications, experienced delay of non-IoT communications, and spectral efficiency of radio resources. (Reprinted from [83], ©2017 IEEE, reused with permission.)

BSs might be turned off for energy saving. In this case, their coverage areas must be covered by the neighboring BSs, i.e. cell zooming occurs. This means that in the low traffic arrival time periods, the density of active BSs decreases, and hence, the communications distances increase. The legacy UEs in cellular networks, i.e. smart-phones, have a semi-daily recharging pattern. Hence, the impact of energy saving for the access network on their performance would be negligible. On the other hand, IoT devices are required to last for years, and hence, an increase in the communications distance may affect their battery lifetime performance significantly. In [84], we develop a tractable framework to model the operation of a sleepingmode enabled BS, which serves mixed IoT/non-IoT traffic. We have derived closedform expressions for energy consumption of the BS, access delay of UEs, and battery lifetime of IoT devices in terms of traffic characteristics and BS operation parameters. These expressions confirm existence of a strong tradeoff between energy saving for the BS and battery lifetime of IoT devices. Then, we explore the impact of system and traffic parameters on the introduced tradeoff. We also extend our solutions to the multi-cell scenario, derive closed-form expressions for the energylifetime tradeoff, and investigate the performance impact of the sleeping control parameters on the tradeoff. Some of the presented simulation results in [84] for the single-cell scenario are summarized in the following. In the single cell scenario, we assume that after serving the last traffic request in the queue, BS listens to the control channel for a period of length TLis seconds. If no request arrives, it sleeps for a period of length TSleep seconds. Also, we consider two traffic types, including non-IoT traffic, denoted by P1 , and IoT traffic, denoted by P2 . The detailed distribution of TLis , TSleep , and characteristics of P1 and P2 could be found in [84].

41

110

70

100

56 E bcons , simulation

90

E bcons , analytic

42

D2 , simulation

80

D2 , analytic

28

Delay (sec)

Energy (Joule)

2.2. SERVING IOT OVER DEDICATED SYSTEMS

D1 , simulation D1 , analytic

70

60 1

10 100 Mean listening time (sec)

14

0 1000

Figure 2.9: Tradeoff between average energy consumption of the BS per unit time, b i.e. Econs , and the experienced delay for P1 and P2 traffic. (Reprinted from [84], ©2016 IEEE, reused with permission.)

The tradeoff between energy saving for the BS and the delay in data transmission for P1 and P2 devices has been depicted in Fig. 2.9. The x-axis in this figure represents the mean listening time, which indicates the average amount of time BS spends in the idle listening before going to the sleep mode, i.e. E{TLis }. Note that if any service request arrives during the listening time, BS serves it immediately and the listening counter is restarted. The solid lines represent the simulation results, while the dashed lines represent the analytical results, derived from the analytical expressions in [84, Section 4]. It is clear that the energy consumption (operational costs) of the BS increases by an increase in the listening time, while the communication delay of users decreases by an increase in the listening time. The tradeoff between energy saving for the BS and energy efficiency in IoT communications has been depicted in Fig. 2.10. One observes that energy efficiency of IoT communications, which determines the battery lifetime, and energy consumption of the BS, which determines the operational costs, increase by an increase in the idle listening time. The interesting point of this tradeoff consists in the fact that MNOs pay the electricity bills of the access network and benefit from further energy saving for the access network. On the other hand, IoT-devices’ owners, e.g. a company which is responsible for temperature monitoring of a building, pay for the maintenance of IoT devices and benefit from maximizing devices’ battery lifetimes. One must note that while investment in the access network by MNOs increases their expenses and benefits IoT-service consumers, but in return, it increases the interest in cellular networks for enabling IoT, and hence, increases their revenues as well.

Energy (Joule)

110

×10 7 4

100

3.3

90

2.6

80

1.9 E bcons

70 60 10 0

for the BS

Energy efiiciency for P2 devices

10 1

10 2

1.2

Energy Efficiency (bpj)


42

0.5 10 3

Mean listening time (sec)

Figure 2.10: Tradeoff between energy consumption of the BS per unit time and energy efficiency of P2 devices. (Reprinted from [84], ©2016 IEEE, reused with permission.)

2.2

Serving IoT over Dedicated Systems

NB-IoT is the state-of-the-art cellular technology introduced by the 3GPP for serving IoT traffic over cellular networks in an energy efficient way. As mentioned in the background section, in this technology different coverage classes have been defined to enable faraway located devices experiencing large path-loss values to become directly connected to the network. However, there is a lack of research on mutual impacts of coexisting coverage classes on each other. This concern originates from the fact that in NB-IoT systems, the data, control, random access, and broadcast channels are multiplexed on the same set of radio resources, as depicted in Fig. 2.11. Then, if we consider a scenario in which uplink channel is mainly occupied by the random access and data transmission of faraway located devices with poor coverage and a high repetition order, the radio resources for other classes cannot be scheduled frequently. Thus, their service delay and consumed energy in communications will increase. Then, it is important to adapt scheduling of random access, shared data, broadcast, and control channels in the uplink and downlink directions based on the coexisting coverage classes in the service area, as well as the number of connected devices in each coverage class. In order to fill this research gap and improve IoT connectivity over NB-IoT networks, the following research questions are studied in this work: • Scheduling of Physical Channels: – RQ1.5: How does scheduling of random access, shared data, broadcast, and control channels in the uplink and downlink directions affect battery lifetime and service delay of IoT devices? • Coexistence of Coverage Classes:


43

Downlink frame (10 subframes) 1

2

NPBCH

3

4

5

6

NPSS

NPDCCH or NPDSCH

7

8

9

NPDCCH or NPDSCH

10 NSSS/NPDCCH or NPDSCH

Uplink frame (10 subframes) 1

2

3

4

5

6

7

8

9

10

NPUSCH or NPRACH

Figure 2.11: NB-IoT technology features frequency-division duplex for uplink and downlink communications [28]. Downlink/uplink channels and signals are multiplexed in the time domain. (Reprinted from [60], ©2018 IEEE, reused with permission.)

– RQ1.6. What are the consequences of serving devices requiring extreme coverage support, i.e. experiencing large path-loss values, on the performance of other IoT devices? How can we compensate for the consequences?

2.2.1

Proposed modeling, solutions, and key results

In response to RQ1.5., we study the physical channel scheduling for NB-IoT, i.e. when and how much resources must be allocated to NPRACH, NPUSCH, NPDCCH, and NPDSCH. We investigate this problem for the case in which, multiple coverage classes coexist in the service area. Let us first exemplify the interactions between scheduling of different physical channels in order to get insights on how a scheduling policy can affect battery lifetime and delay performance of devices. From Fig. 2.11, it is clear that {NPRACH, NPUSCH}, and {NPDCCH, NPDSCH} are multiplexed in time over uplink and downlink radio resources respectively. If NPRACH scheduling occurs frequently, the NPUSCH will suffer from a lack of resources for data transmission, and hence, latency in data transmissions increases. On the other hand, if the scheduling of NPRACH occurs infrequently, latency and energy consumption in connection establishment increase because of the extra delay in starting the RA procedure, and increase in probability of collision over NPRACH. Furthermore, infrequent scheduling of NPDCCH leads to wastage of uplink/downlink resources because access to up/downlink resources is granted through the NPDCCH. Conversely, if NPDCCH is scheduled frequently, NPUSCH will suffer from the lack of resources for data transmissions, and hence, latency and energy consumption of devices over NPUSCH will increase. While the interactions amongst scheduling of different physical channels are complicated, the introduction of coverage classes and coexistence of heterogeneous services over the same system make the problem more challenging, and harder to be solved. In [60] (Paper B), we investigate the NB-IoT channel multiplexing problem in coexistence of devices from a diverse set of coverage classes in the same service area. Towards answering the RQ1.5 and RQ1.6, we first model the access pro-

44

CHAPTER 2. IOT CONNECTIVITY OVER CELLULAR NETWORKS: CHALLENGES, SOLUTIONS, AND KEY RESULTS RA

Uplink

Data RAR

Power

Downlink

Transmit data PC

NRS, NPSS, NSSS, master information block (MIB)

Time $%$

%$!"#

&'

Deep sleep

a) Device demanding uplink service Uplink

RA Data

RAR

Power

Downlink

$'

Time Synch. PC

idle PC

RACH PC

Waiting for RAR PC

Receive data PC

b) Device demanding downlink service

NPBCH

NPSS

NSSS

Optional check for data from BS

NPRACH NPDCCH NPUSCH NPDSCH

Figure 2.12: Data exchanges and their respective power consumptions. The legend shows each data exchange happens over which physical channel. The reference signals, including NRS, NPSS, NSSS, and master information block (MIB), are broadcasted regularly, while one realization has been depicted here for brevity. (Reprinted from [60], ©2018 IEEE, reused with permission.) #$% &/ ()*+", -% .

NPUSCH

NPDSCH

NPRACH

!"

BS initiated control signals

NPDCCH

#$% &' ()*+", -% .

Cell refere-nce signals

IoT devices

Figure 2.13: Queuing model of the NB-IoT access networking. The yellow and gray circles represent servers for downlink and uplink channels, respectively (Reprinted from [60], ©2018 IEEE, reused with permission.).


45

tocol exchanges of NB-IoT systems, based on the description in [21, Section 7]. Fig. 2.12 represents the access protocol exchanges, and their respective power consumption levels. In order to shed light on the interconnections of scheduling of different physical channels, we propose a queuing model of NB-IoT connectivity. This model, depicted in Fig. 2.13, takes the operations of random access, control, and data channels into account. In Fig. 2.13, the yellow circle represents the downlink channel, which serves three channel queues, i.e. NPDCCH, NPDSCH, and reference signals. Also, the gray circle represents the uplink channel, which serves two channel queues, i.e. NPRACH and NPUSCH. By investigating the proposed queuing model, and considering multi-type traffic arrival from different coverage classes, we derive tractable analytical models of service latency, energy consumption, and expected battery lifetime that take into account message exchanges on both downlink/uplink channels, from synchronization to the service completion. These expressions, validated by simulation results, enable us to investigate the latency-energy tradeoff in channel scheduling, and study the interaction among the coverage classes coexisting in the system. In the following, we summarize parts of the performance evaluation results presented in [60] (Paper B). First, Fig. 2.14a represents the coexistence impact of two coverage classes on each other. Here, traffic from class 1 and class 2 devices are served using the same set of radio resources. The fraction of devices belonging to class i, i ∈ {1, 2}, is denoted by fi . For this figure, t and d, the time intervals between two scheduling of NPRACH and NPDCCH, are fixed to 10 ms and 65 ms, respectively. The y-axis in Fig. 2.14a represents the expected battery lifetime, and the x-axis represents the repetition order of class 2 devices, i.e., c2 . Note that c1 has been fixed to 1. The other simulation parameters could be found in [60] (Paper B). In this figure, it is clear that by an increase in c2 , the amount of radio resources used for data transmissions of class 2 devices increases. This, in turn, increases the latency of communications service for class 1 devices, and hence, increases their energy consumptions in the idle waiting mode. Furthermore, one observes that the battery lifetime performance for class 1 devices decreases in the fraction of devices belonging to class 2. For example, by increasing the repetition order of the second class from 11 to 13, the average battery lifetime of class 1 nodes decreases by 6 % when f1 = 0.95 (i.e., f2 = 0.05) and by 28 % when f1 = 0.90 (i.e., f2 = 0.1). Then, performance of class-1 devices is traded to achieve a deeper coverage for class-2 devices. Now, let us investigate the impact of channel scheduling on the KPIs. In Fig. 2.14b, the expected battery lifetime has been depicted by changing t and d, i.e., the average time intervals between two scheduling epochs of NPRACH and NPDCCH, respectively. As per Fig. 2.14a, devices belonging to two coverage classes coexist in the service area. One observes that increasing t increases the battery lifetime of devices in both classes up to a certain point. This is due to the fact that increasing t provides more resources for scheduling of NPUSCH, and hence, decreases time spent in data transmission, i.e., Dtx . One further observes that after a certain point, increasing t shortens the battery lifetime. This is mainly due to

46


the fact that the average delay in resource reservation and probability of collision over NPRACH increase in increasing t. Following the same reasoning, increasing d increases the battery lifetime to some point by providing more resources for NPDSCH, i.e. by decreasing the time spent in data/control reception, Drx . After a certain point, increasing d decreases the battery lifetime. This is mainly due to the fact that increasing d increases the expected time in receiving the resource reservation response, and hence, wastes extra energy. The normalized lifetime and latency for class 1 are depicted in Fig. 2.14c, when t is fixed to 1 ms. One observes that the downlink and uplink latencies are minimized at d = 350 ms and d = 40 ms respectively, and battery lifetime is maximized at d = 80 ms.

2.3

Concluding Remarks

In this chapter, we have presented scalability and energy efficiency analyses of cellular networks in serving massive IoT communications. Furthermore, we have presented enhanced RRM procedures, and have shown the extent up to which, they can improve scalability and energy efficiency in cellular-based IoT communications. The results confirm that the massive access issue could be addressed by well-scaled resource provisioning, as well as enhanced collision resolution, strategies. The results further show that while the scalability of IoT communications could be improved in the light of proposed schemes, the signaling overhead is still a major bottleneck for enabling ultra-long battery lifetime in these systems. Especially, when we consider IoT messages containing a payload of several bits, the total consumed energy in synchronization, authentication, and resource reservation is much higher than the consumed energy in the actual data transmission. From this analysis, we conclude that revolutionary MAC schemes breaking the barrier of grant-based communications are needed to simplify the IoT connectivity procedure as much as possible.

2.3. CONCLUDING REMARKS

47

L: Battery lifetime (days)

1500

Class 1, f 1 =0.9

1000

Class 2, f 1 =0.9 Class 1, f 1 =0.95 Class 2, f 1 =0.95

500

Class 1, f 1 =0.98 Class 2, f 1 =0.98 Class 1, f 1 =0.995 Class 2, f 1 =0.995

0 0

5

10

15

20

c2 : Number of repetitions for class 2

(a) Mutual impact between two coexisting classes in a cell versus number of repetitions for the second class (c1 = 1, f2 = 1 − f1 ).


Class 1 Max lifetime@ {d=0.0164, t=0.0989}


1500

1000

500

0 10 0 10 -1

10 -1

10 -2

t (sec)

d (sec)

Normalized lifetime and latency

(b) Battery lifetime L versus t and d. 1 0.8 Max lifetime d=0.08

0.6

Min uplink latency d=0.04

0.4

Battery lifetime Downlink latency Uplink latency Min downlink latency d=0.35

0.2 0 0.05

0.09

0.13

0.23

0.33

0.5

d (sec)

(c) Overall performance analysis for class 1 vs. d; t = 1 ms.

Figure 2.14: Performance evaluation. (Reprinted from [60], ©2018 IEEE, reused with permission.)

Chapter 3

Enabling Ultra-Energy-Efficient IoT Connectivity: Challenges, solutions, and key results In the previous section, we focused on RQ1, and tried to shed light on major bottlenecks of existing cellular infrastructure in offering ultra-energy-efficient IoT connectivity service. The analytical and simulation results indicated that the communications overhead in connection establishment and control signaling consumes as large as, or even higher than, the energy consumption in actual data transmission. In this section, we focus on RQ2, i.e., given the limitations of cellular networks in supporting IoT, how can one develop communications protocols with a close-tozero energy consumption profile? In response to this question, we investigate the possibility of removing the resource reservation procedure for short-packet IoT communications. Towards this end, grant-free radio access is introduced here in which, once a device has a packet to transmit, it sends one/several copies of the packet over the shared resource pool. At first, we investigate challenges associated with the grant-free radio access in large-scale heterogeneous IoT networks. Then, we investigate the reliability of communications and its interactions with battery lifetime, latency, density of BSs, and communications bandwidth. We further develop enhanced transmission/reception solutions for grant-free access IoT networks. Finally, we develop distributed yet energy-efficient radio access coordination schemes, to be implemented in devices, which are able to enhance devices adaptation to the environment significantly.

3.1

Grant-free Access: A performance analysis

The IoT connectivity solutions over the unlicensed spectrum, e.g. SigFox, along with their cellular competitors, e.g. NB-IoT, have targeted the energy-hungry IoT market [18]. In essence, grant-free radio access belongs to the category of IoT 49

CHAPTER 3. ENABLING ULTRA-ENERGY-EFFICIENT IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS

50

solutions over the unlicensed spectrum in which, devices don’t need to reserve radio resources, however, they need to comply with the fair use of radio resources [30]. In recent years, grant-free access has also gained attention in literature and industry for realizing long battery lifetime in cellular networks over the licensed spectrum [85]. Thanks to its simplified connectivity procedure, grant-free access is a promising solution to let IoT devices focus just on data gathering and reporting, and forget about the legacy energy-consuming connection establishment procedure. While the 3GPP has started standardization of the grant-free access over cellular networks [85], there is still a lack of comprehensive research on grant-free access, especially on reliability of communications. In section 1.2.2, we observed that the prior arts on reliability in grant-free access have been mainly focused on success probability analysis in homogeneous scenarios, i.e. when a specific technology like LoRaWAN, or a specific IoT service like smart metering is the sole user of the radio resources. Furthermore, spatial correlation in locations of devices has not been investigated in prior arts. In most prior arts, e.g. [62], the distribution process of locations of IoT devices has been modeled by a PPP, which is infeasible because the cell ranges in wide-area IoT networks can be up to tens of kilometers. Then, there is a high probability that the service area includes hot-spots, as well as regions without any device deployment, as depicted in Fig. 3.1. In response to the lack of research on large-scale, heterogeneous, and grant-free radio access IoT networks, in this section we aim at investigating the reliability of communications over such networks. This problem is defined as follows. • RQ2.1: How can one model the reliability of communications in large-scale, heterogeneous, grant-free access IoT networks?

3.1.1

Proposed modeling and key results

In [65], appended Paper C, we answer RQ2.1 by investigating reliability analysis in grant-free IoT networks. First, we leverage a multi-PCP1 model, for modeling distribution of locations of devices in a wide-area network. Then, we provide a rigorous analytical model of reliability of such a network, powered by stochastic geometry analysis of the 3D interference2 , in terms of network resources and characteristics of the multi-type traffic. In the following, we present a key result of this work in which, probability of success for a network serving K-type IoT devices over a bandwidth of W is investigated. For a broader view of contributions and derived expressions, one may refer to Paper C. 1 PCP: 2 In

Poisson cluster process. time, frequency, and code domains

3.1. GRANT-FREE ACCESS: A PERFORMANCE ANALYSIS

1000

51

BS IoT device (type 1) IoT device (type 2) IoT device (type 3)

y-axis (meter)

500

0

-500

-1000 -1000

-500

0

500

1000

x-axis (meter) (a) Illustration of locations of devices for K = 3. This figure highlights heterogeneity in distribution processes of locations of devices.

(b) A snapshot of the received traffic from 4 transmitting devices (K = 3) at the receiver. This figure highlights differences in communication characteristics.

Figure 3.1: Graphical description of motivation for RQ2.1. Performance analysis of large-scale IoT networks with heterogeneity in communications characteristics and distribution processes of locations of devices is missing in the prior arts [65]. (Reprinted from [65], submitted to IEEE for peer review.)

52 3.1.1.1

CHAPTER 3. ENABLING ULTRA-ENERGY-EFFICIENT IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS Probability of success in a grant-free access IoT network with K-type IoT devices

Let us first describe the system model. A massive number of IoT devices, denoted by set Φ, have been distributed according to different 2D PCPs in a wide service area. ∆ Φ consists of K subsets, denoted by Φk for k ∈ K = {1, · · · , K}, and each subset represents a specific type of IoT devices. Traffic from different subsets, i.e. different IoT devices, differ in communications characteristics. Some of these characteristics are as follows: (i) frequency of packet generation 1/Tk , (ii) signal bandwidth wk , (iii) packet transmission time τk , (iv) number of transmitted replicas per data packet nk , and (v) transmit power Pk . For PCP model of type-k IoT devices, a feature tuple, like (λk , υk , f(x)), represents the distribution process. In this tuple, λk is the density of parent points, i.e. cluster centers, and υk is the average number of daughter points per parent point, i.e. number of cluster members. Also, f(x) characterizes the distribution of cluster members in each cluster, and is described by an isotropic function. For example, in the case of a normal distribution, we have: √ f(x) = exp(−||x − x0 ||2 /(2σ 2 ))/ 2πσ 2 , (3.1) in which, σ denotes the variance of distribution and x0 is the location of the cluster center. A frequency spectrum of W (Hz) is shared amongst nodes for communications. Over the communications bandwidth, the PSD of noise is denoted by U. To account for both IoT connectivity over the licensed and unlicensed spectrum, we consider collecting data from a subset of K, denoted by ϕ, where |ϕ| ≤ |K| = K. The transmissions from other devices are treated as interference. When |ϕ| = K, the modeling works for the licensed spectrum, in which, only authenticated devices can transmit data over the resource pool. When |ϕ| < K, the modeling works for IoT connectivity over the unlicensed spectrum, in which, coexisting technologies may reuse the same resource simultaneously. Finally, we need to take into account different transmission codes that might be used by different device technologies for resilience to interference, e.g. in LoRaWAN [18]. Here, for the brevity of text we neglect use of such codes, however, in Paper C, we have derived the analytical results by considering the existence of such codes. Now, let us come back to our performance analysis problem. Denote by U , IΨ , and γth,i the additive noise at the receiver, the received interference at the receiver, and the threshold SINR for correct decoding of a packet of type-i. Also, denote the average numbers of interfering type-k devices in each cluster as υˆk , which is the product of υk and ρk = ρtk ρfk ρck . The latter, ρk , is the 3D activity factor of devices, and is derived as the product of device’s activity factors in time, frequency, and code domains. These activity factors have been quantified in [65, Section III]. Furthermore, assume the fading between devices and the BS is modeled by the Nakagami-m model, with the shaping and spread parameters of m ∈ Zi and Ω > 0 respectively. Finally, assume the following general path-loss model in communications of devices and the BS: g(x) = 1/(α1 +α2 ||x||δ ), in which α1 and α2

3.1. GRANT-FREE ACCESS: A PERFORMANCE ANALYSIS

53

are the control parameters, and δ represents the path-loss exponent. Now, we have all the necessary background for performance analysis. The probability of success in packet transmission of a type-i device, located at z, to the BS, which has been located at the origin, is derived as: ps (i, z) = Pr(Pi hg(z) ≥ [U + IΨ ]γth,i ) ∫ ∞ γth,i mq ν (a) ∑m-1 1 = exp(− )q dPr(IΨ +U ≥ q) ν=0 ν! 0 ΩPi g(z) ν (b) ∑m-1 (−1) = [LIΨ (s)LU (s)](ν) s= γth,i m . ν=0 ΩPi g(z) ν!

(3.2)

ν

∂ In this expression, [F (s)](ν) = ∂s ν F (s), (a) follows from [86, Appendix C], and ∂n (b) follows from [87, Lemma 3.1] and the fact that L(tn f(t)) = (−1)n ∂s n F (s). Furthermore, LIΨ denotes the Laplace functional of the received interference from potential active nodes, and has been derived in section III of Paper C. Also, LU (s) denotes the Laplace functional of noise. In order to shed light on the coexistence impact of different IoT traffic types, in the following we focus on m = 1, i.e. when the Nakagami-m reduces to the Rayleigh fading.

Theorem 1. For m = 1, the success probability in a packet transmission of type-i devices could be approximated as: ( ) ( ∑ ps (i, z) ≈ exp − Uγth,i /[ΩPi g(z)] exp −

Pk γth,i ) ) k∈K ΩPi ( γth,i ) × exp − υî H(z, f ∗ (x), ) , (3.3) Ω λk υˆk H(z, 1,

( ) where f ∗ (·) = conv f(·), f(·) , and ( H z, f ∗ (x), ω) =

∫ x∈R2

g(x) f ∗ (x)dx. g(x) + g(z)/ω

Proof. Refer to Appendix A of Paper C. The expression in (3.3) consists of three exponential expressions, related to probability of success in presence of noise, and potential inter-cluster and intracluster interference, respectively. Furthermore, the H(z, f ∗ (x), ω) and H(z, 1, ω) functions could be derived in closed-form for most well-known path-loss and user distribution functions. For example, for g(x) = α0 ||x||−δ , we have: 2

H(z, 1, ω) = ||z||2 ω δ 2π 2 csc(2π/δ)/δ. In the following, we investigate tightness of the derived analytical expressions in Theorem 1. Towards this end, consider a coexistence scenario in which, two device types coexist. Communications from type-1 devices are of our interest,

54


p s (1, d): Probability of success

1

0.9

0.8 Simulation Analytic Impact of Type-1 (other clusters) Impact of Type-1 (same cluster) Impact of Type-2 (all) Impact of noise

0.7

0.6

0.5 0

500

1000

1500

2000

d: Distance from the BS (meters)

Figure 3.2: Validation of analytical and simulation results. Device distribution: K=2, λ1 =0.19, λ2 =3.8, υ1 =1200, υ2 =30, P1 =21 dBm, and P2 =25 dBm. Other parameters could be found in [65]. (Reprinted from [65], submitted to IEEE for peer review.)

and hence, type-2 communications are treated as interference. Also, the transmit power of type-2 devices is 4 dB higher than type-1 devices. The probability of success in packet transmission of type-1 devices has been depicted as a function of distance from the BS in Fig. 3.2. One observes that the analytical model, derived in (3.3), matches well with the simulation results. We have further highlighted the contributions of noise, inter-cluster interference, and intra-cluster interference, as derived in (3.3). By comparing the respective curves, it is clear that the interference from type-2 traffic (plus-marked curve) is the most limiting factor. This is due to the fact that transmit power of these interfering devices is 4 dB higher than the devices of interest. It is clear that the presented reliability model could be extended to evaluate performance of any IoT network in coexistence scenarios. For example, we have used this analytical framework in [61] for performance analysis of two private LoRa-based networks.

3.2

Deployment and Operation Strategies for Grant-free Access Networks

In the previous subsection, we highlighted the lack of research on reliability analysis in the large-scale grant-free access IoT networks. Furthermore, one must note that study of the tradeoff between required investment cost in the access network, in terms of BS density and communications bandwidth, and provided quality of service for IoT devices, in terms of reliability, latency, and battery lifetime, is also missing in the literature. To address this open problem, in this section we focus on network

3.2. DEPLOYMENT AND OPERATION STRATEGIES FOR GRANT-FREE ACCESS NETWORKS 55 design and operation control problems in grant-free IoT networks. Enabling widearea IoT connectivity requires deployment of BSs and provisioning of frequency resources, where both impact the network costs. Furthermore, the experienced delay and consumed energy in data transfer for IoT devices have strong couplings with reliability of data transfer, which has been found in (3.3) to be interconnected with the amount of provisioned BS density/radio resources for IoT communications. In order to investigate the couplings and fill this research gap, here we answer the following research question: • RQ2.2: What is the interplay amongst the density of BSs within the service area, communications bandwidth, and QoS for IoT communications? How does this interplay shape the resource provisioning and operation control strategies?

3.2.1

Proposed solutions and key results

In [65], appended Paper C, we answer RQ2.2 by investigating the tradeoff between provisioned network resources and QoS of IoT communications. At first, we provide analytical models of battery lifetime and latency for IoT devices in terms of devices communications parameters and reliability of communications. We further model the required annual investment in the access network in terms of leasing cost for the communications bandwidth, density of BSs, and the annual cost for maintenance and energy consumptions of BSs. Then, we analyze how (i) change of device’s communications parameters, e.g. number of replicas or transmit power, and (ii) change of network resources, e.g. BS density and communications bandwidth, impact the device battery lifetime and latency of communications. This analysis further presents operation regions in which, tuning devices’ communications parameters, e.g. the number of replicas, increases both reliability and battery lifetime, offers a tradeoff between them, or decreases both of them. Backed to the derived expressions, we further present cost-optimized resource provisioning strategies for the access network in which, the required density of BSs and/or volume of radio resources for achieving a required level of QoS are determined. The analysis is further extended to the centralized operation control of IoT devices. Then, battery lifetime-optimized operation control strategies, subject to a given reliability level, are derived for IoT devices. We also leverage the developed analytical framework and carry out the scalability analysis. For example, we present scalability of the access network by investigating the point up to which, the amount of provisioned network resources should be scaled to compensate for the increased traffic load, interference, or QoS requirement. Finally, for a given access network, we investigate the scalability of device resources, and derive the bounds up to which, energy consumption per packet must be scaled to compensate for the increased traffic load, interference, or QoS requirement.

56

CHAPTER 3. ENABLING ULTRA-ENERGY-EFFICIENT IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS IoT devices: Types, density, transmit power, replica/packet.

Provisioned resources: Density of BSs, system BW

IoT devices: Types, density, transmit power, replica/packet.

Provisioned resources: Density of BSs, system BW

Analysis

Optimized operation control

Analysis Optimized resource provisioning

Eco. viability of access network

Probability of success in transmission

Energy consumption per reporting period

Reliability of communications

Durability of communications

Probability of success in a transmission




(a)

(b)

8000

Battery lifetime ( × reporting period) Probability of collision (1-p s(1,d eg))

6000

Cost, Lifetime

1

Network cost ( × 0.1 c 1 )

0.8 0.6

4000 0.4 2000

0 10 -2

B1 = 8

0.2 0 10 -1

10 0

Probability of collision in transmission

Figure 3.3: Contributions related to RQ2.2: (a) optimized resource provisioning for the access networks in serving IoT traffic; (b) optimized operation control for IoT devices. (Reprinted from [65], submitted to IEEE for peer review.)

λa : Density of BSs ( × Km 2 )

Figure 3.4: Tradeoff between network cost, reliability, and battery lifetime (K = 1, λ1 =6.4). c1 represents the average annual cost for maintenance of one BS, and √ deg = 1/(πλa ). (Reprinted from [65], submitted to IEEE for peer review.)

The interested reader may refer to Section V and VI of [65] (Paper C) in which, detailed expressions for KPIs and their interactions have been presented. Fig. 3.3 depicts a graphical illustration of the major contributions presented in [65]. In the following, we summarize parts of the performance evaluation results conducted in [65]. The simulations parameters could be found in [65].

1

600

acceptable region

Cost ( × c 1 )

p s (1,d eg): Reliability

3.2. DEPLOYMENT AND OPERATION STRATEGIES FOR GRANT-FREE ACCESS NETWORKS 57

0.5

400

minimum cost in the acceptable region

acceptable region

200 0 20

0 20 10

10

System BW (KHz)

0

0

Bandwidth (KHz) Density of BSs (Km -2 )

10 0

10

10 -1

0

10 -1

Density of BSs (Km -2 )

(a) Specifying the bound beyond which, (b) Finding the minimum-cost investthe reliability constraint is satisfied. ment strategy in the acceptable region.

Figure 3.5: The optimized deployment strategy: How much do we need to invest in the density of BSs and communications bandwidth? (λ1 =4.6, required ps (1, deg )=0.7). (Reprinted from [65], submitted to IEEE for peer review.) The tradeoff between cost of the access network (left y-axis), battery lifetime of devices (left y-axis), and reliability of communications (right y-axis) could be found in Fig. 3.4. Here, the x-axis represents the density of deployed BSs. One observes that an increase in density of deployed BSs incurs a decrease in the probability of collision, and hence, offers a lower required number of retransmissions for achieving a certain reliability level. This, on the one hand, confirms that battery lifetime could be prolonged by provisioning more network resources for the IoT traffic, and hence, at the cost of higher investment in the access network. The resource provisioning problem has been investigated in Fig. 3.5. First, Fig. 3.5a depicts the reliability of communications by changing the BS density and communications bandwidth. By considering 0.7 as the required success probability in a packet transmission, the (BS density, communications bandwidth) pairs for which the reliability constraint have been satisfied has been characterized, as depicted in Fig. 3.5a. Next, in Fig. 3.5b, network cost has been depicted by changing the provisioned network resources, i.e. BS density and communications bandwidth. By searching over the acceptable region, found by mapping Fig. 3.5a to Fig. 3.5b, one can find the (BS density, communications BS) pair which minimizes the network costs, as marked in Fig. 3.5b. The centralized device operation control problem has been investigated in Fig. 3.6. This figure characterizes the expected battery lifetime in terms of devices’ energy consumption in a packet transmission, i.e. transmit power (P1 ), and the number of replicas per packet (n1 ). One sees that by increasing both P1 and n1 , the expected battery lifetime increases up to some point, and then starts decreasing. This is due to the fact that by increasing the number of replicas and transmit power, the performance against noise will be improved, but the system will become interference-limited after some point. Finally, the rate at which, the energy consumption of IoT devices or the amount

CHAPTER 3. ENABLING ULTRA-ENERGY-EFFICIENT IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS Type-1 battery lifetime ( × reporting period)

58

3000

Lifetime vs. P 1 (Sc1) Lifetime vs. n 1 (Sc1)

2500


2000 1500 optimized operation point for type-1

1000 500 0 1

2

3

n 1 : Nr of replicas,

4

5

6

7

8

9

10

P1 : Transmit power ( × 126 mW)

Figure 3.6: The optimized operation control (K = 2, λ2 =2.4, λ1 =2.4 in Sc1 and λ1 =1.2 in Sc2). In circle-marked curves, n1 = 1 and P1 is varying. In plus-marked curves, P1 = 126 mW and n1 is varying. (Reprinted from [65], submitted to IEEE for peer review.) of provisioned network resources, should be scaled to comply with the increase in the required QoS has been depicted in Fig. 3.7. One sees that the transmit power of devices could be increased up to a certain point, just to combat the noise effect. However, beyond that point, an increase in the transmit power cannot incur an increase in the success probability because the network becomes interferencelimited. While in this scenario, increasing number of replicas per packet is an effective tool for increasing reliability, there is a saturation point in scenarios with higher densities of nodes, where increasing number of replicas increases the traffic load, and hence, it becomes a foe.

3.3

Enhanced Transmission and Reception Schemes for Grant-free Access

In response to RQ2.1 and RQ2.2, we found that probability of success in grantfree data transmission decreases by an increase in the traffic load, and hence, the grant-free radio access scheme loses its superior performance in prolonging battery lifetime. Let us denote by basic grant-free access scheme, the data transmission and reception schemes investigated in response to RQ2.1 and Rq2.2. In this scheme, once a device has a packet to transmit, it sends a pre-determined number of replicas of the packet. The BS receives the replicas and combines them for data recovery. In this section, we aim to enhance the transmission and reception procedures, by answering the following research question: • RQ2.3: Given the bottlenecks of grant-free data transmission in RQ2.1 and

10 1

20

Required AP density (× Km-2 ) Required BW (× 10 KHz) Required n 1

15

Required P 1

10 0

10 10 -1 5 10 -2 0.3

59

Required BW, n 1 , P1

Required BS density ( × Km -2 )

3.3. ENHANCED TRANSMISSION AND RECEPTION SCHEMES FOR GRANT-FREE ACCESS

0 0.4

0.5

0.6

0.7

0.8

0.9

1

p s (1,d eg): Required reliability

Figure 3.7: Scalability analysis (K = 1, λ1 =3.2). (Reprinted from [65], submitted to IEEE for peer review.)

RQ2.2, how can one enhance the transmission and reception schemes to compensate for the bottlenecks?

3.3.1


In [79], appended Paper D, we answer RQ2.3 by enhancing both the transmission and reception schemes. In order to improve the replica transmission used in the basic grant-free access scheme, in which each device sends its data by sending a constant number of replica transmissions, we develop a path-loss dependent replica transmission control scheme. In this scheme, each device selects its replica transmission order based on its average communications path-loss with the BS. Then, we extend the reliability expressions derived in response to RQ2.1 and RQ2.2 to cover the case in which, the replica transmission order differs from one device to another. Furthermore, we formulate the operation control problem as a lifetime maximization problem in which, the optimized threshold path-loss values and number of replicas in each path-loss region are to be found. Fig. 3.9 represents a system in which, different devices send different numbers of replicas per packet. In this system, once a device has a packet to transmit, it assumes a virtual frame (VF) consisting of Mv virtual slots, each equal to the time duration of its packet transmission. Then, it sends the first replica immediately and sends the remaining Nv −1 replicas in Nv − 1 randomly selected slots afterwards. The VFs across different devices are asynchronous in time and frequency. Also, the Nv and Mv are parameters for optimization, can differ from one device to the other, and are investigated in our enhanced packet transmission protocol in [79]. We further investigate the packet reception procedure and propose an advanced receiver, which leverages the random timing and frequency offsets of received packets in order to carry out successive interference cancellation. The proposed receiver

60


design has been depicted in Fig. A. In this figure, DM(f ) represents demodulation with carrier frequency f , and ∆fk represents frequency drift of the k-th signal involved in the collision. The detailed description of this receiver and design procedures of its modules, e.g. the peak detection and decision making modules, has been presented in Paper D. In short, the proposed receiver, receives a signal complex containing an unknown number of involved signals, determines the collision order from time and frequency offsets, and tries to decode each involved signal. For the signals it is successful in decoding them, the positions of their related replicas are also revealed3 . Thus, the other replicas are also removed for further interference cancellation. The signals involved in the collision complex which cannot be decoded are stored and their CFOs are used as their signatures. Then, the packets with almost the same CFOs are combined together, e.g. using selection combining, for potential data recovery. The reason behind using CFO as the signature consists in the fact that during transmission of replicas related to a data packet, the device’s CFO could be assumed to be almost the same [88], as depicted in Fig. 3.9. The performance evaluation results in Paper D, which is partially represented in the following, confirm that this CFO-triggered SIC-based receiver can enhance the collision resolution performance significantly. In the following figures, N =x;TiAs;FrAs represents the basic grant-free access, in which, x denotes the number of replicas per data packet. Furthermore, ADP(N1 , N2 , dth ) represents the advanced transmission scheme in which, the threshold for replica transmission control is adjusted by the distance from the BS, i.e. dth . Also, N1 and N2 represent the number of replicas for devices before and beyond the dth , respectively. Finally, one must note that the advanced receiver, introduced in the above, is used for packet reception of the proposed and benchmark data transmission schemes. The battery lifetime and delay analysis of the basic grant-free, advanced grantfree, and grant-based radio access schemes have been presented in Fig. 3.10. For a circular service area of 3 radius Km, Fig. 3.10a shows that the ADP(1,2,2500) outperforms the others in battery lifetime. Especially, it is clear that the bottleneck of the basic grant-free scheme with fixed N in medium to high traffic load regimes, has been addressed using the advanced packet transmission scheme. Similar results are observed in Fig. 3.10b for the service delay. In this figure, one observes that the ADP(1,2,2500) outperforms the grant-based approach, as well as the basic grant-free one. Now, we investigate the performance impacts of dth and N2 on the ADP(1,N2 ,dth ), using the results presented in Fig. 3.11. First, the battery lifetime performance by changing the threshold distance has been depicted in Fig. 3.11a. One sees that dth =2750 m, achieves the highest battery lifetime, i.e. when 84% of the coverage area (outer ring) is covered by N1 = 1, and 16% of the coverage area is covered by N2 = 2, we achieve the highest battery lifetime. Also, Fig. 3.11b represents the battery lifetime performance by changing the value of N2 . It is clear that N2 = 2 achieves the highest battery lifetime for different traffic load regimes. 3 Each packet contains the packet index and information about the position of the packet relative to the other replicas.

3.3. ENHANCED TRANSMISSION AND RECEPTION SCHEMES FOR GRANT-FREE ACCESS (!)

" (#)

Sampling Windowing

Peak Detection

Periodogram

DM(*)

DM(* 0 1*' )

DM(* 0 1*% )

DM(* 0 1*$ )

/*$ /*% /*'

&$ (#)

",' (#)

",% (#)

",$ (#)

Correlation &' (#)

Peaks Detection

Decision Making

Preamble(s)

(-",$ # , *",$ , +$,. )

(*",$ , +",$ )

Cut sequence

SIC Module

61

(*",', , +",' )

(-",' # , *",' , +$,' )

Figure 3.8: The proposed receiver design for triggering SIC by CFO. (Reprinted from [79], submitted to IEEE for peer review.)

Frequency

VF VS

Time

12000

20 N=1; TiAs; FrAs N=2; TiAs; FrAs Grant-based ADP(1,2,1000) ADP(1,2,1500) ADP(1,2,1950) ADP(1,2,2500)

10000 8000

N=1; TiAs; FrAs N=2; TiAs; FrAs Grant-based ADP(1,2,1000) ADP(1,2,1500) ADP(1,2,1950) ADP(1,2,2500)

15

Delay (Sec)

Battery Lifetime ( × reporting period)

Figure 3.9: Received packets at the receiver. VFs have different carrier frequencies, start times, and slot sizes.

6000

10

5

4000 2000

0 0

1

2

3

4

Λ : Aggregated packet arrival rate (Hz)

(a) Battery lifetime analysis

5

6

0

1

2

3

4

5

6


(b) Delay analysis

Figure 3.10: Battery lifetime and latency analysis. (Reprinted from [79], submitted to IEEE for peer review.)

Battery lifetime (×1000 reporting period)

10 9 8 7 6

Λ=1.7 Λ=2.5 Λ=3.3 Λ=4.9

5 4 1000

1500

2000

2500

dth : Threshold distance

2750

3000

Battery lifetime ( × reporting period)


62

9000 8000 7000 6000 5000

Λ=2.5 Λ=3.3 Λ=4.5 Λ=5.7

4000 3000 1

2

3

4

5

N2 : Number of replica transmissions beyond dth

(a) Battery lifetime vs. threshold dis- (b) Battery lifetime vs. number of replitance for ADP (1,2,dth ) cas per packet for ADP (1,N2 ,2500)

Figure 3.11: Tuning design parameters for ADP(N1 , N2 , dth ). Λ denotes the aggregated traffic arrival rate. (Reprinted from [79], submitted to IEEE for peer review.)

The degradation in the battery lifetime by further increase in N2 is due to the fact that by increasing the number of replicas, the system becomes interference limited.

3.4

Distributed Coordination of Grant-free Access Networks

One main drawback with existing grant-free IoT networks, e.g. LoRaWAN and SigFox, is the lack of coordination in spectrum usage. Fig. 1.7 represents the interference measurement results in the ISM band, and confirms that some subbands could be blocked in some use-cases for a short/long period of time due to misbehaving devices or the high density of deployed devices. In this case, the reliability of IoT communication will be low. In order to tackle this problem, we either need coordination from the access networks or device intelligence to adapt itself to the environment. The coordination by the access network requires an active connection and also signaling. The former is challenging in scenarios where connectivity is blocked due to sudden changes, e.g. once a car containing the IoT device enters the basement garage. The latter, i.e. signaling, requires frequent data transmission/reception to/from the BS in order to get up-to-date guidance for communications, and hence, is energy consuming. These shortcomings motivates us to investigate other ways in which, an energy-limited IoT device can adapt itself to the environment. In response to these concerns, we answer the following question:

• RQ2.4: How can one enable IoT devices served over grant-free access networks to adapt themselves to the changes in reliability of their communications?

3.4. DISTRIBUTED COORDINATION OF GRANT-FREE ACCESS NETWORKS

3.4.1

63


In order to answer RQ2.4, we need to enable IoT devices to adapt themselves to the environment. Consider a service area in which a BS is located at the center, and plenty of IoT devices have been distributed in the service area. Also, a communications bandwidth of W , containing |S| subchannels, is shared among devices for uplink communications. The communications parameters of the ith device, which should be adapted, include transmit power (Pi ), data rate (Ri ), frequency subchannel (Si ), and number of replicas per packet (Ni ). Assume at time t, a tagged device, say device with index i, has data to transmit. Then, the operation control problem could be written as follows: max

Pi ,Ri ,Si ,Ni

F (Reli , EEi )

s.t :Pi ∈ P, Ri ∈ R, Si ∈ S, Ni ∈ N,

(3.4)

in which F (·) represents the objective function as a function of reliability (Rel) and energy-efficiency (EE) of communications. Definition of F (·) may differ from one application to the others. Here, we focus on a weighted sum of objectives, i.e. F (Reli , EEi ) = (1 − β)Reli + βEEi . In this expression, 0 ≤ β ≤ 1 tunes a tradeoff between reliability and energy efficiency of communications. Furthermore, X denotes the set of available values for Xi . A straightforward solution to (3.4) comes from solving it as an integer programming problem by receiving information about the environment and coexisting devices from the BS. However, this solution is not applicable to energy-limited IoT devices because requires frequent signaling with the BS, and hence, is in contrast with our aim in design of ultra-energy-efficient IoT connectivity solutions. Instead of solving the problem by receiving information from the BS, let us leverage the distributed online learning solution in which, each device4 maximizes its objective function F (Reli , EEi ) by choosing the best action subset, i.e. Ai = {Pi , Ri , Si , Ni }, out of the action set, i.e. A. After taking an action, the device may receive a downlink message, representing success/failure of its previous transmission(s). This downlink message is called the reward, and itself is vulnerable to intended/unintended interference and noise. Let us denote the reward after taking Ai (t) by ξ(t) ∈ {1, 0}, where 1 and 0 represent acknowledgment and no acknowledgment respectively. This type of machine learning problem is commonly described as multi-agent multi-arm bandit (MAB) in the machine learning literature [89]. In the original MAB learning problem, the agent chooses one of the K arms at each time and receives a reward in response. The agent aims at maximizing its self-accumulative rewards. In dealing with a MAB learning problem, we leverage exploration and exploitation, where the former indicates decision epochs in which device explores different actions randomly, and the latter indicates decision epochs at which agent decides greedy based on the previous actions/rewards. In 4 Hereafter,

we also call it the agent.


64

[72] (Paper E), we follow the multi-arm bandit learning problem, but we modify the reward function to consider both external reward for success in transmission and internal reward for energy saving in successful data transmission, as follows. Denote by Emin and Ej , the minimum consumed energy amongst actions which have resulted in a successful transmission, and the consumed energy in data transmission using action j. We define our updated reward function function as: ˆ = ξ(t)(1 − β) + ξ(t)βEj /Emin . ξ(t)

(3.5)

In this expression, ξ(t) ∈ {0, 1} represents the acknowledgment, β is a design parameter tunning the tradeoff between reliability and energy efficiency, and t represents the time. Leveraging the modified reward function, we propose two algorithms in Paper E, as depicted in Algorithm 2 and Algorithm 3. The former has been intended for operation control in environments dealing with interference over the data channels and no interference over the feedback channel, also called stochastic MAB. The latter has been intended for environments dealing with interference over the data and feedback channels, also called adversary MAB. In Algorithm 2, ˆ represent the number of times action k has Tk (t), Zk (t), bk (t), A(t), ξ(t), and ξ(t) been chosen, the accumulated updated reward for action k, the index5 of action k, the selected action, the received reward, and the modified reward. In Algorithm 3, Wk and pk represent the weight and probability of selection of action k, respectively. Also, Ω and ρ tune the exploration-exploitation tradeoff. In order to investigate the usefulness of these algorithms in practice, we apply the proposed distributed learning approach for operation control in the context of LoRa technology. Then, the results from distributed learning are compared against the results from solving the equivalent centralized optimization problem, where the latter is derived by leveraging tools from stochastic geometry. The performance evaluation results indicate a significant decrease in energy consumption, as well as probability of failure in communications. Furthermore, the results indicate a significant tradeoff between reliability of communications against unintended/adversarial interference and energy efficiency. In other words, increasing the former, do decrease the latter, and vice versa. In the following, we represent parts of the simulations results of Paper E. Consider a massive number of LoRa nodes distributed according to a PPP in a 2D service area of radius 2 Km. Our aim is to assign 6 spreading factors, i.e. {7, 8, 9, 10, 11, 12}, and two transmit power levels, i.e. {8,14} dBm, to the devices. The simulation parameters have been presented in Table 1 of Paper E. In the following figures, Alg1 and Alg2 denote Algorithm 2 and Algorithm 3, respectively. Furthermore, EqLoad refers to a centralized algorithm for assigning spreading factors, as proposed in [90]. In the EqLoad algorithm, the number of devices using a specific SF is proportional to the data rate associated with that SF. Furthermore, RandSel refers to the algorithm in which SFs are selected randomly by devices. The Alg1, Alg2, and the benchmark schemes choose Pt = 14 dBm 5 This

index represents the merit of each action in the action selection step.

3.4. DISTRIBUTED COORDINATION OF GRANT-FREE ACCESS NETWORKS

65

Algorithm 2: Pseudo-code of UUCB1 for stochastic MAB. (Reprinted from [72], ©IEEE 2018, reused with permission.) 1 2

Initialization: Zk (1)=0, Tk (1)=1, ∀k ∈ A; for t = 1, 2, · · · do √ - Update index: bk (t) = Zk (t) + Ω log(t)/Tk (t); - Take action: arg maxk∈A bk (t) → A(t); - Receive reward: ξ(t) ∈ {0, 1}; - Update reward: Zk (t+1)=Zk (t), ∀k ∈ A\A(t); ˆ ZA(t) (t+1)=ZA(t) (t)+ξ(t); - Update counter: TA(t) (t+1)=TA(t) (t)+1; Tk (t+1)=Tk (t), ∀j ∈ A\A(t); - return A(t);

Algorithm 3: Pseudo-code of UEXP3 for adversary MAB. (Reprinted from [72], ©IEEE 2018, reused with permission.) 1 2

Initialization: Wk (1) = 1, ∀k ∈ A; for t = 1, 2, · · · do Wk (t) - Define Dist.: pk (t)=(1-ρ) ∑|A| j=1

Wj (t)

ρ + |A| , ∀k ∈ A;

- Take action: A(t) ∼ {p1 (t), · · · , p|A| (t)}; - Receive reward: ξ(t) ∈ {0, 1}; - Update weight: Wk (t+1)=Wk (t), ∀k ∈ A\A(t); ˆ

WA(t) (t+1)=WA(t) (t) exp( |A|pρξ(t) (t) ); A(t)

0.8

0.7

0.6 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel Alg.1(PC), β=0.01

0.5

0.4 0

20

40

60

Index of transmitted packets

80

Consumed energy per packet (J)


- return A(t);

12

×10 -3 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel Alg.1(PC), β=0.01

10 8 6 4 2 0

20

40

60

80


Figure 3.12: Learning for power and SF selection control with stochastic external interference (Sc2). (Reprinted from [72], ©IEEE 2018, reused with permission.)


66

2000

2000

1500

1500 1000 SF7 SF8 SF9 SF10 SF11 SF12

500 0 -500

y-axis (meter)

y-axis (meter)

1000

0 -500

-1000

-1000

-1500

-1500

-2000 -2000

-1000

0

x-axis (meter)

1000

2000

SF7 SF8 SF9 SF10 SF11 SF12

500

-2000 -2000

-1000

0

1000

2000

x-axis (meter)

Figure 3.13: The constellation of selected SFs for Alg1 (left) and EqLoad Algorithm (right). (Reprinted from [72], ©IEEE 2018, reused with permission.) as the transmit power. In contrast, Alg1(PC) represents a version of Algorithm 2 in which, not only the SF, but also the transmit power could be selected by each device out of {8, 14} dBm. The reliability and energy consumption analyses have been presented in Fig. 3.12. In this setup, we have external interference along with the interference from the LoRa devices over the data channels. The intensity of interference over each SF differs from another. One observes a significant increase in success probability and a significant decrease in energy consumption have been achieved by leveraging the proposed algorithms. Here, one can also observe the tradeoff between reliability and energy efficiency of communications. This tradeoff is tuned by the design parameter, β. We observe that Alg1(PC), which controls transmit power as well as the SF, achieves a good success probability at the cost of ultra-low energy consumption, when β is fixed to 0.5 (the solid green line). On the other hand, by setting β to 0.01, one observes that the success probability significantly improves at the cost of an increase in the energy consumption in comparison with the Alg1(PC). One can further observe the constellation of allocated SFs to devices in different regions of the cell, by following Alg1 (left) and the EqLoad scheme (right), in Fig. 3.13. The setup in Fig. 3.12 has been repeated again in Fig. 3.14, with the only difference that here we deal with adversary interference over the feedback channel. In this setting, the adversary inverts 50% of the feedback messages sent by the BS, i.e. with a probability of 0.5, an ACK message is substituted by a NACK message and vice versa. Here, we observe that Alg2, outperforms the others in reliability and energy consumption. One must note that while the depicted energy consumption per packet transmission for Alg1 is lower than Alg2 in Fig. 3.14, but because of the increased number of required retransmissions in Alg1, the total energy consumption of Alg1 will be much higher than the Alg2.

67 ×10 -3

0.5




0.45

0.4 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

0.35

0.3 0

20

40

60


80

12 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

10 8 6 4 2 0

20

40

60

80


Figure 3.14: Learning for transmit power and SF selection control in an adversarial setting with stochastic external interference (Sc2). (Reprinted from [72], ©IEEE 2018, reused with permission.)

3.5

Concluding Remarks

In this chapter, we have evaluated the performance of grant-free access IoT networks. This analysis is beneficial in determining the strengths and shortcomings of grant-free radio access in different traffic load regimes and coexistence scenarios. Our analysis confirms that for low to medium traffic load regimes, the grant-free access can decrease both latency and energy consumption in data transmission significantly. By an increase in the traffic load; including an increase in the number of devices, message length, or packet generation frequency; the grant-free access starts losing its superior performance. To partially compensate for this problem, we have also developed an advanced replica transmission control, as well as a SICbased receiver. Furthermore, for distributed control of grant-free IoT networks, we have leveraged on-device intelligence and devised a light-weight learning scheme to send data over interference-prone channels. The simulation results have verified the performance of the proposed solutions, and promote integration of them in future cellular networks, along with the legacy grant-based schemes in LTE and NB-IoT systems. Finally, we note that the proposed approach has the potential to be used in other asynchronous communications systems, e.g., in satellite communications.

Chapter 4

Enabling Ultra-Reliable IoT Connectivity: Challenges, solutions, and key results In this chapter, we shift our focus from serving traffic from energy-limited IoT devices to serving reliability- and delay-constrained (cIoT) traffic. Supporting cIoT communications is a major challenge of next-generation wireless networks. The cIoT is indeed considered as the crucial prerequisite of a new wave of services, including smart factory, remote control, and intelligent transportation systems (ITS) [19, 66]. The latter itself could be seen as a promising solution for many challenges, as it can enable safer and greener transportations. Whilst enabling cIoT is essential for realizing such promising applications, it is never an easy task. Literature study, including the 3GPP standardization reports, reveals that serving cIoT communications over cellular networks is in its infancy [66]. The main reason behind such a gap is the stringent reliability requirement in cIoT, which should be addressed within a limited number of retransmissions, or even without any retransmission [19]. Here, we mainly focus on radio resource management (RRM) for cIoT.

4.1

RRM for Non-Scheduled Critical IoT Communications

In the uplink direction, enabling cIoT requires satisfying delay and reliability requirements not only for scheduled cIoT transmissions but also for non-scheduled (NS) transmissions. The non-scheduled transmissions can originate from connected/disconnected devices which have urgent data, or devices which ask resource reservation for subsequent transmissions. A graphical illustration of the scheduled/non-scheduled traffic coexistence has been depicted in Fig. 4.1(a). Here, an intelligent vehicle communicates with other vehicles (V2V) as well as with the infrastructure (V2I) for traffic efficiency and safety. The depicted non-scheduled communications in this figure include (i) first packet transmission (FPT) of devices 69

CHAPTER 4. ENABLING ULTRA-RELIABLE IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS

70

which transmit critical data and reserve resource for further packets [19] (depicted as FPT-NS in Fig. 4.1(b)); (ii) urgent packet from a device to the network [70] (depicted as V2I-NS in Fig. 4.1(b)); and (iii) urgent V2V communications [91] (depicted as V2V-NS in Fig. 4.1(b)). As discussed in the background section, time-diversity, and hence retransmission, could not be leveraged in assuring reliability within a strictly bounded time interval. Thus, reliable transmission of data must be carried out using frequency diversity. Regarding the fact that frequency resources are limited, the main challenge consists in jointly serving non-scheduled and scheduled traffic over a limited set of frequency resources, while their reliability constraints are satisfied. Towards addressing this concern, the third research question is defined as follows.

• RQ3: What kinds of RRM schemes are suitable for serving scheduled and non-scheduled uplink cIoT traffic over a limited set of radio resources?

!

BS

downlink data /control/ACK

$

!

!

!

signaling FPT-NS

V2V-NS

scheduled transmission scheduled transmission BS!

signaling

V2I-NS V2V-NS

# !

"

(a) V2X communications (ITS use case)

(b) Communications exchanges

Figure 4.1: An illustrative example of the system model in the ITS use case, including scheduled (green dashed arrows) and non-scheduled (red dotted arrows) communications. (Reprinted from [17], submitted to IEEE for peer review.) The straightforward solution to RQ3 comes from the isolation of scheduled and non-scheduled traffic types, i.e., by allocating orthogonal slices of resources to both of them. However, the need for extra frequency resources in cIoT communications for achieving reliability limits scalability of this proposal. This is even more challenging when we consider the fact that non-scheduled traffic occurs infrequently, and hence, continuous allocation of a bunch of radio resources to an infrequent traffic type for achieving reliability is not spectrum efficient. Recently, non-orthogonal serving of coexisting services, including eMBB/cIoT and eMBB/mIoT, has attracted attention [69, 70]. Especially, overlay transmission for cIoT communications which cannot tolerate scheduling delay has been investigated in the 3GPP standardization [85].

4.2. PROPOSED SOLUTIONS AND KEY RESULTS

4.2

71

Proposed Solutions and Key Results

Here, we also follow the non-orthogonal serving approach and develop a hybrid orthogonal/non-orthogonal multiple access (HMA) approach in which, the available radio resources are divided into three parts: (i) dedicated to scheduled transmissions, (ii) dedicated to non-scheduled transmissions, and (iii) shared between them. While on the one hand overlay transmission of critical data can save radio resources, on the other hand it introduces more constraints to the RRM problem and makes it more complex. Furthermore, one must note that in order to comply with the delay requirements of cIoT applications, the RRM is needed to respond to the changes in network/environment in very short timescales. Motivated by the recent advances in artificial intelligence and machine learning (ML) [92], in response to RQ3, we investigate the potential use of ML approaches for devising a spectrum efficient RRM scheme able to dynamically manage resources between scheduled and non-scheduled uplink traffic from different sources. The traditional ML includes a central node which has access to a huge amount of data and makes decisions in a centralized way. In managing radio resources for cellular networks, these approaches are neither scalable nor can comply with the delay/reliability requirements of critical communications. Furthermore, one must take into account the large dimensionality and complexity of the RRM problems, which come from the heterogeneity of resources, services, and QoS requirements. In [17] (Paper F), we investigate leveraging machine learning in design of an RRM solution for serving non-scheduled cIoT traffic along with other services. Towards addressing the challenges faced in solving the RRM problem with traditional ML, we investigate a distributed learning-powered RRM, which includes intelligent radio resource provisioning (RRP) at the RAN control center, intelligent resource scheduling (RRS) at the edge, and intelligent resource utilization (RRU) at the end devices. The proposed solution aims at enabling spectrum efficient yet reliable coexistence of non-scheduled cIoT traffic along with the other traffic, when both of them call for specific delay-reliability constraints. In [17], we also investigate various sources of uncertainty in our proposed resource scheduling, including traffic arrival and channel fading, study their impacts on the RRM decisions, and investigate the amount of radio resources which are sacrificed to compensate for those performance losses. A graphical representation of the proposed hierarchical learning solution has been depicted in Fig. 4.2. In this figure, proactive RRP based on training and prediction is carried out at the RAN control center. This RRP leverages transfer learning, federated learning, and risk-aware decision making. The former includes making decisions based on the temporal/spatial correlations in the service area, i.e. leveraging the knowledge gained in decisions taken in neighboring service areas and/or decisions taken in the same service area but at previous time instances. [93]. Federated learning consists in leveraging the available records of requests/responses at the edges for extracting common rules and policy sets without the need to transfer all raw data to the RAN control center [94]. Here, the edge nodes perform

CHAPTER 4. ENABLING ULTRA-RELIABLE IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS

72

preprocessing of data, and update their corresponding models based on the available data, and send a compressed version of data or model parameters to the RAN control center. Finally, one must note that in contrast with the majority of MLpowered solutions in which, we aim at maximizing the return, e.g. throughput, here we seek for reliability in communications. Thus, one needs to shift the objective of learning from return maximization to risk minimization in which, the risk takes into account the outage probability for cIoT communications. In cellular infrastructure, each BS is responsible for a given part of the service area, and hence, it has full access to local information and limited access to the neighboring cells’ informations through BS-2-BS and BS-2-RAN communications (the partial observability problem). On the other hand, cIoT traffic generating devices could be highly mobile, e.g. autonomous vehicles, and hence, resource provisioning at the BSs for non-scheduled cIoT communications could be risky due to the partial observability of BSs. Distributing the learning task among BSs, devices, and the RAN control center can address the partial observability of BSs. For example, while a BS cannot predict arrival of a vehicle platoon to its service area, the RAN control center can make the RRM at the respective BS ready by adapting the high-level RRP policies1 to the network/user status. In the proposed hierarchical RRM solution, BSs control access of the scheduled devices to the dedicated and shared radio resources. The fact that scheduling decisions are carried out at the edges2 , which are close to devices, enables us to react to instant changes in the environment without suffering much from the partial observability. Finally, intelligent utilization of radio resources for reliable communications is left for the end devices. The intelligent devices, leverage both the network assistance and their former experience in communications in order to consume their internal resources, e.g. the transmit power, as well as the network resources, e.g. the dedicated and shared radio resources, to send their data in a reliable way [72]. The above-described hierarchical RRM solution requires a set of modules for performing the intelligent RRM, e.g. a radio map of the service area, which are omitted here for the brevity of text. The interested reader is referred to [17, section IV] (Paper F) for more details.

A Case Study In this section, we investigate the application of the proposed intelligent RRM for jointly serving of scheduled and non-scheduled traffic. We consider a single-cell scenario, in which the infrequent non-scheduled cIoT traffic is transmitted along with the scheduled traffic. The service area is described by the minimum and maximum experienced path-loss, which are -70 and -120 dB, respectively. The total number of radio resource blocks (RRBs), which should be decided to be shared with the scheduled traffic or to be used exclusively by the non-scheduled traffic, is 1 The 2 The

high-level RRP policies are regularly sent to the BSs. decisions are made with partial assistance from the RAN control center.

4.2. PROPOSED SOLUTIONS AND KEY RESULTS

Action: RRP: (i) inter-service/cell resource allocation, (ii) BSs’ activity management.

Learning: (i) risk-aware online learning for RRP update, (ii) inter-BS transfer learning for RRS, (iii) federated learning of mobility pattern, etc.

Action: RRS: (i) inter and intra service resource scheduling.

Learning: (i) risk-aware online learning for RRS update, (ii) regression-based user/radio map update.

Action: RRU: (i) intelligent resource utilization for data transmission and reception.

Learning: (i) risk-aware online learning for RRU update.

73

RAN control center

Figure 4.2: Hierarchical ML-powered RRM. The solution includes (i) RRP at the radio access network (RAN) control center, (ii) RRS at the BSs, and (iii) RRU at the end-devices. (Reprinted from [17], submitted to IEEE for peer review.)

assumed to be 5. The dedicated resources to the scheduled traffic are not considered here. Here, scheduling is done by leveraging the map of users, their directions of movement, and the radio map of the environment. Based on the outage risk of the most critical device with potential non-scheduled cIoT traffic, and the learned risk of outage in the service area, the ML-powered RRP and RRS modules are trained and used for resource management, as described in the above section. Fig. 4.3(a) represents the coupling between the achieved reliable data rate for scheduled traffic by changing the outage probability requirement of non-scheduled traffic. In this figure, the dashed-curve represents a conservative RRM solution in which, resource allocation is carried out based on the risk to the cell-edge user, i.e. it is static. We have further depicted the decoupled gains of RRM with orthogonal multiple access (OMA) and HMA. The former represents the achieved data rate by an orthogonal allocation of resources to the scheduled and non-scheduled traffic, while the latter represents the achieved data-rate by reuse of resources for the scheduled traffic when possible. From this figure, one can conclude that the higher the outage risk, i.e. the higher need for reservation of resources for infrequent non-scheduled traffic in the conservative approach, the higher room for spectrum efficiency using the proposed intelligent RRM. Moreover, one observes that the HMA represents its benefits when the required reliability level increases. One further observes that uncertainty, both in the wireless channel and traffic arrival, calls for scarifying more radio resources to comply with the reliability constraints, which on the other hand degrades the spectrum efficiency of the system. Finally, we investigate the benefit of intelligent radio resource utilization at the end devices. Let us assume a typical source of non-scheduled cIoT traffic experiences a distance-dependent path-loss of -80 dB. Out of 5 available RRBs, at most 3 RRBs could be used by this device for non-scheduled transmissions. The PSD of scheduled

CHAPTER 4. ENABLING ULTRA-RELIABLE IOT CONNECTIVITY: CHALLENGES, SOLUTIONS, AND KEY RESULTS 2.5

×10 6

Outage probability for non-sched traffic

0.9999-reliable data rate (b/s) for sched.

74

2

1.5

1

0.5

0 1e-1

No-learning (conservative design) Intelligent RRP & RRS with OMA Intelligent RRP & RRS with HMA

1e-2

1e-3

1e-4

1e-5

1e-6

1e-7

Target outage probaility for non-scheduled traffic

(a) Intelligent RRP and RRS: reliable data rate for the scheduled traffic versus the outage probability for the non-scheduled traffic (single-shot transmissions). Increase in the required reliability highlights the merits of intelligent RRM and HMA.

10 -2

10 -3 PSD of Interference= [0,-164,-164] dBm/Hz

10

-4

10 -5 PA= [1,0,0]; OMA PA= [α, 1-α ,0] PA= [α, (1-α)/2, (1-α)/2] PA= [α, 1-α, 0] PA= [α, (1-α)/2, (1-α)/2]

10 -6

10 -7

PSD of Interference= [0,-172,-172] dBm/Hz

0

0.2

0.4

0.6

0.8

1

α : fraction of allocated power to RRB1

(b) Intelligent RRU: outage probability for different power allocation (PA) decisions. PA=[α, x, y] represents a decision in which, α fraction of transmit power is used for the dedicated RRB, and x and y fractions are used for the other two shared RRBs.

Figure 4.3: Performance evaluation of the proposed RRM. (Reprinted from [17], submitted to IEEE for peer review.)

traffic’s interference over these 3 RRBs is denoted by [0, -172, -172] dBm/Hz, i.e. the first RRB is allocated solely to the device’s transmissions, and the other two are shared with the scheduled traffic. In Fig. 4.3(b), the outage probability for non-scheduled transmissions of this device by changing the fraction of its allocated power to the dedicated RRB, i.e. α, has been depicted. One observes that the best action consists in distributing the transmit power equally over all RRBs. This could be reasoned by considering the PSD of interference over shared channels, which is comparable with the PSD of noise, i.e. -174 dBm/Hz. On the other hand, we observe that by an increase in the PSD of interference, i.e. in the [0, -164, -164] dBm/Hz case when interference is much stronger than the noise, the best action consists in consuming half of the power over the dedicated RRB, and dividing the other half amongst the shared ones. One further observes in both cases that with an intelligent selection of α, HMA will outperform OMA (square-marked line), significantly.

4.3

Concluding Remarks

In this chapter, we have investigated the coexistence of scheduled and non-scheduled cIoT traffic, as a challenge to be tackled for the realization of cIoT services. Towards serving cIoT traffic over a limited set of resources, we have introduced a hybrid orthogonal/non-orthogonal multiple access scheme, with dedicated and shared subsets of resources. To tackle the complexity of RRM management for this hybrid access scheme, we have proposed a distributed learning-powered RRM solution.


75

Our preliminary results have confirmed the usefulness of the proposed scheme in spectrum efficient serving of the critical traffic, along with serving the scheduled traffic. This promotes use of learning-powered RRM schemes in future 5G networks, where we are dealing with a heterogeneous set of resources, e.g. sub-GHz and THz frequency resources, a heterogeneous set of users, e.g. IoT devices and human beings, and a heterogeneous set of QoS requirements, e.g. latency requirements of milliseconds and minutes.

Chapter 5

Conclusions and Future Research Directions 5.1

Conclusions

Providing large-scale ultra-durable low-cost IoT connectivity is the key requirement for realizing a networked-society, in which everything that benefits from being connected is connected. Towards this end, this thesis was focused on enabling technologies for IoT connectivity, fundamental shortcomings of them, and our proposed solution for filling the research gaps. To conclude this thesis, we need to come back to our three main research questions and answer them one by one.

Answer to RQ1 • What are the challenges of existing cellular networks in energy-efficient serving of mIoT communications? How can one enhance their scalability and energy efficiency, and up to which extent? RQ1 requires performance evaluation of existing cellular networks in serving mIoT communications, and also investigation of their fundamental limits beyond which, their performance cannot be improved. Towards answering this question, we have deeply investigated the state-of-the-art IoT connectivity solutions as well as the enhanced solutions in the literature, and evaluated the achieved KPIs by each solution. Our study reveals that while the massive IoT traffic arrival problem has attracted profound attention in literature, and sophisticated solutions have been proposed for addressing it, the energy efficiency and battery lifetime of IoT devices have been neglected/sacrificed in most existing solutions. Then, while the massive IoT traffic arrival could be addressed by well-scaled resource provisioning and advanced collision resolution strategies, the battery lifetime is still a major bottleneck of existing IoT connectivity solutions. Especially, when considering IoT messages 77

78 CHAPTER 5. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS containing a payload of several bits, the energy consumption in connection establishment and radio resource scheduling is much more than the energy consumption in actual data transmission. This required signaling also limits the achieved gain by our proposed battery lifetime-aware IoT serving solutions. From these results, we get the insight that leveraging revolutionary grant-free MAC schemes, which break the barrier of grant-based communications, could be a solutions to simplify the IoT connectivity procedure as much as possible.

Answer to RQ2 • RQ2: Given the limitations of cellular networks in supporting mIoT, derived from RQ1, how can one design an energy-efficient MAC for IoT communications, to be integrated into future cellular networks? RQ2 requires resolving the major bottleneck of existing cellular networks, i.e. extra signaling for grant-based radio access. Towards this end, we have evaluated performance of the grant-free radio access scheme, when it is used for uplink transmissions of large-scale, heterogeneous IoT networks. This analysis reveals the strengths and shortcomings of grant-free radio access in different traffic load regimes and coexistence scenarios. Our analysis confirms that for low to medium traffic load regimes, the grant-free access can decrease both latency and energy consumption in data transmission significantly. By an increase in the traffic load; including an increase in the number of devices, packet length, or packet generation frequency; the grant-free access starts losing its superior performance. To partially combat this issue, we have developed an advanced replica transmission control, as well as a CFO-triggered SIC-based receiver. The performance evaluation results show that leveraging the proposed schemes, one can enhance the performance of the grantfree radio access, even in higher traffic load regimes. The results further confirm existence of traffic-load regimes in which, the grant-free radio access outperforms the legacy grant-based radio access schemes, which are used in LTE-A and NB-IoT systems. These results pave the way for enabling intelligent grant-based/free operation mode switching in 5G networks. One must note that, while grant-free access provides ultra-low energy consumption in uplink data transmission, it still suffers from extra energy consumption in receiving the downlink data, e.g. acknowledgment, and hence, achieving reliability. We are currently working on solving this problem by leveraging a secure ultra-low power receiver, which has been excluded from this thesis.

Answer to RQ3 • RQ3: What kinds of RRM schemes are suitable for serving scheduled and non-scheduled uplink cIoT traffic over a limited set of radio resources? Among main challenges in providing ultra-reliable low-latency communications, a major problem consists in the crucial need to huge amounts of frequency resources

5.2. FUTURE RESEARCH DIRECTIONS

79

for enabling frequency diversity. Then, we propose a hybrid orthogonal/non-orthogonal multiple access scheme to serve the coexisting non-scheduled and scheduled traffic. The main challenge with this solution consists in developing an RRM, which distributes the resources to each traffic type in an online way. Motivated by the recent advances in machine learning, we proposed a hierarchical learning-powered RRM solution, including intelligent radio resource provisioning at the RAN control center, intelligent radio resource scheduling at the edges, and radio resource utilization at the device. Our preliminary results promote future works investigating more advanced RRM schemes, including spatial, coding, and frequency domain resources, as well as different sources of the outage, including small and large-scale fading. The complexity of these RRM problems, originating from the heterogeneity of traffic, QoS, resources, and outage causes, call for using more advanced ML solutions in order to achieve ultra-reliability within a bounded time interval.

5.2

Future Research Directions

In this work, we have mainly investigated some key challenges of existing IoT connectivity solutions in serving IoT communications. However, successful realization of IoT in a global scale mandates several other aspects to be investigated. In the following, we introduce 3 important aspects, which have been less investigated in the literature.

Need for low-cost devices An ultra-low cost per IoT device enables widespread use of IoT connectivity in homes, offices, farms, factories, vehicles, and any other place/thing that connectivity can improve it. Furthermore, in many wide-area use cases, e.g. monitoring in agriculture, the cost per device in the deployment phase, and maintenance cost in the operation phase are the major limiting factors. Then, availability of low-cost IoT devices enables deploying IoT devices with a higher density than the required density in order to make the connectivity more reliable against the potential changes in the environment, e.g. loss of nodes. In [63, 79], we have investigated the use of cheap oscillators in IoT devices and devised an advanced receiver design to change the inevitable CFO in communications of such devices from foe to friend. The crucial need for low-cost IoT devices calls for more intense research in this field.

Need for light-weight security solutions Another less-investigated aspect is providing low energy consuming security for IoT communications. Recall from the introduction chapter that the payload size in many IoT applications is in order of bits, e.g. 1-bit indicating occurrence or non-occurrence of an event. On the other hand, the legacy security protocols used in existing wireless communications infrastructures include key exchanges and encryption of data with long keys. Then, there is a crucial need for light-weight

80 CHAPTER 5. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS security solutions enabling secure IoT communications with less added overhead to the payload size. Physical-layer security, in which information theory is employed to enable keyless security, is a promising solution for addressing some security concerns in the IoT field. The interested reader may refer to [95] as a starting point, where the authors investigate extracting physical features of IoT devices by using advanced machine learning approaches, to be used in the subsequent transmissions for physical layer security.

Need for cIoT-specific communications protocols Finally, providing ultra reliability within a bounded time interval is a hot topic at the moment in academia and industry. As the stringent delay-reliability requirements of cIoT traffic are fundamentally different from the typical delay-reliability requirements of the legacy traffic, there is a huge research gap in the design of cIoT-specific communications protocols. As a starting point, one may refer to [17], in which challenges in spectrum-efficient radio resource management for serving heterogeneous cIoT traffic have been investigated.

Bibliography [1]

A. Azari, “Energy efficient machine-type communications over cellular networks: A battery lifetime-aware cellular network design framework,” Licentiate Thesis, Dec. 2016, Royal Institute of Technology (KTH), Stockholm, Sweden.

[2]

P. Friess, Internet of Things-Global Technological and Societal Trends From Smart Environments and Spaces to Green ICT. River Publishers, 2011.

[3]

J. Gubbi et al., “Internet of things (IoT): A vision, architectural elements, and future directions,” Future Generation Computer Systems, vol. 29, no. 7, pp. 1645 – 1660, Sept. 2013.

[4]

S. Andreev et al., “Understanding the IoT connectivity landscape: A contemporary M2M radio technology roadmap,” IEEE Communications Magazine, vol. 53, no. 9, pp. 32–40, Sept. 2015.

[5]

A. Osseiran et al., “Scenarios for 5G mobile and wireless communications: The vision of the METIS project,” IEEE Communications Magazine, vol. 52, no. 5, pp. 26–35, May 2014.

[6]

Ericsson, “More than 50 billion connected devices – taking connected devices to mass market and profitability,” www.akos-rs.si, accessed: 2018-09-27.

[7]

Vodafone, “Vodafone M2M barometer,” www.vodafone.com, accessed: 201803-17.

[8]

H. Blaser and C. Fenger, “M2M service providers and LTE: New revenue opportunities,” Tech. Rep., Nov. 2013.

[9]

S. Abdelwahab et al., “Enabling smart cloud services through remote sensing: An internet of everything enabler,” IEEE Internet of Things Journal, vol. 1, no. 3, pp. 276–288, May 2014.

[10] M. o. Palattella, “Internet of things in the 5G era: Enablers, architecture, and business models,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 3, pp. 510–527, Feb. 2016. 81

82

BIBLIOGRAPHY

[11] A. Nordrum, “Popular internet of things forecast of 50 billion devices by 2020 is outdated,” IEEE Spectrum, vol. 18, Aug. 2016. [12] G. Miao, A. Azari, and T. Hwang, “E 2 MAC: Energy efficient medium access for massive M2M communications,” IEEE Transactions on Communications, vol. 64, no. 11, pp. 4720–4735, Sept. 2016. [13] A. Laya, L. Alonso, and J. Alonso-Zarate, “Is the random access channel of LTE and LTE-A suitable for M2M communications? a survey of alternatives,” IEEE Communications Surveys Tutorials, vol. 16, no. 1, pp. 4–16, Dec. 2014. [14] A. Azari and G. Miao, “Energy efficient MAC for cellular-based M2M communications,” in 2014 IEEE Global Conference on Signal and Information Processing, Dec. 2014, pp. 128–132. [15] G. A. Akpakwu, B. J. Silva, G. P. Hancke, and A. M. Abu-Mahfouz, “A survey on 5G networks for the internet of things: Communication technologies and challenges,” IEEE Access, vol. 6, pp. 3619–3647, Dec. 2017. [16] F. Ghavimi and H. H. Chen, “M2M communications in 3GPP LTE/LTE-A networks: Architectures, service requirements, challenges, and applications,” IEEE Communications Surveys Tutorials, vol. 17, no. 2, pp. 525–549, Oct. 2015. [17] A. Azari, M. Ozger, and C. Cavdar, “Serving non-scheduled URLLC traffic: Challenges and learning-powered strategies,” submitted to IEEE Communications Magazine, Aug. 2018. [18] W. Yang et al., “Narrowband wireless access for low-power massive internet of things: A bandwidth perspective,” IEEE Wireless Communications, vol. 24, no. 3, pp. 138–145, May 2017. [19] I. Parvez et al., “A survey on low latency towards 5G: RAN, core network and caching solutions,” IEEE Communications Surveys & Tutorials, Aug. 2018. [20] Ericsson, Huawei, NSN, and et al., “A choice of future M2M access technologies for mobile network operators,” Tech. Rep., March 2014. [21] 3GPP TR 45.820, “Technical specification group GSM/EDGE radio access network; cellular system support for ultra-low complexity and low throughput internet of things (CIoT),” 2015. [22] Samsung, “Cisco visual networking index: Global mobile data traffic forecast update, 2014–2019,” Tech. Rep., March 2015. [23] A. Gotsis, A. Lioumpas, and A. Alexiou, “M2M scheduling over LTE: Challenges and new perspectives,” IEEE Vehicular Technology Magazine, vol. 7, no. 3, pp. 34–39, Sept. 2012.

BIBLIOGRAPHY

83

[24] M. Mehaseb, Y. Gadallah, A. Elhamy, and H. El-Hennawy, “Classification of LTE uplink scheduling techniques: An M2M perspective,” IEEE Communications Surveys and Tutorials, no. 99, Nov. 2015. [25] Nokia Networks, “LTE-M – optimizing LTE for the Internet of things,” Tech. Rep., 2015. [26] R. Ratasuk, B. Vejlgaard, N. Mangalvedhe, and A. Ghosh, “NB-IoT system for M2M communication,” in IEEE Wireless Communications and Networking Conference, Aug. 2016. [27] Y. P. E. Wang et al., “A primer on 3GPP narrowband internet of things,” IEEE Communications Magazine, vol. 55, no. 3, pp. 117–123, March 2017. [28] J. Schlienz and D. Raddino, “Narrowband internet of things,” Rohde and Schwarz, Tech. Rep., Aug. 2016. [29] P. Andres-Maldonado et al., “Narrowband IoT data transmission procedures for massive machine-type communications,” IEEE Network, vol. 31, no. 6, pp. 8–15, Nov. 2017. [30] É. Morin, M. Maman, R. Guizzetti, and A. Duda, “Comparison of the device lifetime in wireless networks for the internet of things,” IEEE Access, vol. 5, pp. 7097–7114, April 2017. [31] W. Yang et al., “Narrowband wireless access for low-power massive internet of things: A bandwidth perspective,” IEEE Wireless Communications, May 2017. [32] M. Lauridsen et al., “Interference measurements in the european 868 MHz ISM band with focus on LoRa and SigFox,” in IEEE Wireless Communications and Networking Conference, March 2017. [33] M. Centenaro et al., “Long-range communications in unlicensed bands: The rising stars in the IoT and smart city scenarios,” IEEE Wireless Communications, vol. 23, no. 5, pp. 60–67, Nov. 2016. [34] “VT offers $2 sigfox-enabled modules to qualified buyers,” www.vt-iot.com, accessed: 2017-07-26. [35] N. Blenn and F. Kuipers, “LoRaWAN in the wild: Measurements from the things network,” arXiv preprint arXiv:1706.03086, June 2017. [36] Semtech, “LoRa modulation basics,” Tech. Rep., 2014. [Online]. Available: www.semtech.com [37] D. Bankov, E. Khorov, and A. Lyakhov, “Mathematical model of LoRaWAN channel access,” in Proceedings of the IEEE 18th International Symposium on A World of Wireless, Mobile and Multimedia Networks, June 2017.

84

BIBLIOGRAPHY

[38] ——, “On the limits of LoRaWAN channel access,” in IEEE International Conference on Engineering and Telecommunication, Nov. 2016. [39] P. Pinault, “Sigfox downlink in details,” www.disk91.com, accessed: 2018-0319. [40] M. I. Hossain, A. Azari, and J. Zander, “DERA: Augmented random access for cellular networks with dense H2H-MTC mixed traffic,” in 2016 IEEE Global Communications Workshops, Dec. 2016. [41] A. Azari and G. Miao, “Network lifetime maximization for cellular-based M2M networks,” IEEE Access, vol. 5, pp. 18 927–18 940, Sept. 2017. [42] ——, “Lifetime-aware scheduling and power control for M2M communications in LTE networks,” in IEEE Vehicular Technology Conference (Spring), May 2015. [43] ——, “Lifetime-aware scheduling and power control for cellular-based M2M communications,” in IEEE Wireless Communications and Networking Conference, March 2015. [44] 3GPP TR 37.868 V11.0.0, “Study on RAN improvements for machine-type communications,” Tech. Rep., Sept. 2011. [45] S. Duan, V. Shah-Mansouri, and V. W. Wong, “Dynamic access class barring for M2M communications in LTE networks,” in IEEE Global Communications Conference, Dec. 2013, pp. 4747–4752. [46] A. Laya et al., “Goodbye, ALOHA,” IEEE Access, vol. 4, pp. 2029–2044, April 2016. [47] G. Foddis et al., “On RACH preambles separation between human and machine type communication,” in IEEE International Conference on Communications. IEEE, July 2016. [48] S. Y. Lien and K. C. Chen, “Massive access management for QoS guarantees in 3GPP machine-to-machine communications,” IEEE Communications Letters, vol. 15, no. 3, pp. 311–313, March 2011. [49] G. Wunder, M. Kasparick, and P. Jung, “Interference analysis for 5G random access with short message support,” in Proceedings of European Wireless Conference, May 2015. [50] K. Wang, J. Alonso-Zarate, and M. Dohler, “Energy-efficiency of LTE for small data machine-to-machine communications,” in IEEE International Conference on Communications, June 2013, pp. 4120–4124.

BIBLIOGRAPHY

85

[51] M. Kalil et al., “Low-complexity power-efficient schedulers for LTE uplink with delay-sensitive traffic,” IEEE Transactions on Vehicular Technology, vol. 64, no. 10, pp. 4551–4564, Nov. 2015. [52] 3GPP TS 22.368 V13.1.0, “Service requirements for machine-type communications,” Tech. Rep., June 2014. [Online]. Available: www.3gpp.org [53] A. Aijaz et al., “Energy-efficient uplink resource allocation in LTE networks with M2M/H2H co-existence under statistical QoS guarantees,” IEEE Transactions on Communications, vol. 62, no. 7, pp. 2353–2365, July 2014. [54] G. Miao, N. Himayat, G. Li, and S. Talwar, “Low-complexity energy-efficient scheduling for uplink OFDMA,” IEEE Transactions on Communications, vol. 60, no. 1, pp. 112–120, Jan. 2012. [55] G. Miao and G. Song, Energy and Spectrum Efficient Wireless Network Design. Cambridge University Press, 2014. [56] X. Lin, A. Adhikary, and Y. P. E. Wang, “Random access preamble design and detection for 3GPP narrowband IoT systems,” IEEE Wireless Communications Letters, vol. 5, no. 6, pp. 640–643, Dec. 2016. [57] T. Kim, D. M. Kim, N. Pratas, P. Popovski, and D. K. Sung, “An enhanced access reservation protocol with a partial preamble transmission mechanism in NB-IoT systems,” IEEE Communications Letters, June 2017. [58] C. Yu et al., “Uplink scheduling and link adaptation for narrowband internet of things systems,” IEEE Access, vol. 5, pp. 1724–1734, Feb. 2017. [59] A. Adhikary, X. Lin, and Y. P. E. Wang, “Performance evaluation of NB-IoT coverage,” in 2016 IEEE 8 Vehicular Technology Conference (Fall), Sept. 2016. [60] A. Azari, G. Miao, C. Stefanovic, and P. Popovski, “Latency-energy tradeoff based on channel scheduling and repetitions in NB-IoT systems,” IEEE Global Communications Conference, Dec. 2018. [61] M. Masoudi, A. Azari, E. A. Yavuz, and C. Cavdar, “Grant-free radio access IoT networks: Scalability analysis in coexistence scenarios,” IEEE International Conference on Communications, Dec. 2018. [62] Z. Li et al., “2D time-frequency interference modeling using stochastic geometry for performance evaluation in low-power wide-area networks,” arXiv preprint arXiv:1606.04791, May 2016. [63] A. Azari, P. Popovski, G. Miao, and C. Stefanovic, “Grant-free radio access for short-packet communications over 5G networks,” in IEEE Global Communications conference, Dec. 2017.

86

BIBLIOGRAPHY

[64] Y. Zhang, K. Peng, Z. Chen, and J. Song, “SIC vs. JD: Uplink NOMA techniques for M2M random access,” in IEEE International Conference on Communications, May 2017. [65] A. Azari, M. Masoudi, and C. Cavdar, “Optimized resource provisioning and operation control for low-power wide-area IoT networks,” submitted to IEEE Transactions on Communications, 2018. [66] M. Bennis, M. Debbah, and H. V. Poor, “Ultra-reliable and low-latency wireless communication: Tail, risk and scale,” arXiv preprint arXiv:1801.01270, Jan. 2018. [67] E. Cabrera, G. Fang, and R. Vesilo, “Adaptive hybrid ARQ (A-HARQ) for ultra-reliable communication in 5G,” in IEEE Vehicular Technology Conference (Spring), June 2017. [68] A. T. Z. Kasgari and W. Saad, “Stochastic optimization and control framework for 5G network slicing with effective isolation,” in 52nd IEEE Annual Conference on Information Sciences and Systems, Jan. 2018. [69] P. Popovski et al., “5G wireless network slicing for eMBB, URLLC, and mMTC: A communication-theoretic view,” arXiv preprint arXiv:1804.05057, April 2018. [70] H. Ji et al., “Introduction to ultra reliable and low-latency communications in 5G,” arXiv: 1704.05565, April 2017. [71] N. Baldo et al., “An open source product-oriented LTE network simulator based on NS-3,” in Proceedings of the 14th ACM international conference on Modeling, analysis and simulation of wireless and mobile systems. ACM, Nov. 2011. [72] A. Azari and C. Cavdar, “Self-organized low-power IoT networks: A distributed learning approach,” in IEEE Global Communications Conference, Dec. 2018. [73] A. Azari and A. Bria, “System and method for providing communication rules based on a status associated with a battery of a device,” PCT/SE2017/050933 (filed), Nov. 2017. [74] M. I. Hossain, A. Azari, J. Markendahl, and J. Zander, “Enhanced random access: Initial access load balance in highly dense LTE-A networks for multiservice (H2H-MTC) traffic,” in IEEE International Conference on Communications, May 2017. [75] N. Abu-Ali et al., “Uplink scheduling in LTE and LTE-advanced: Tutorial, survey and evaluation framework,” IEEE Communications Surveys Tutorials, vol. 16, no. 3, pp. 1239–1265, Dec. 2014.

BIBLIOGRAPHY

87

[76] C. Cox, An introduction to LTE: LTE, LTE-advanced, SAE and 4G mobile communications. John Wiley & Sons, 2012. [77] A. Laya, “The internet of things in health, social care, and wellbeing,” Ph.D. dissertation, KTH Royal Institute of Technology, Sept. 2017. [78] A. Azari, M. I. Hossain, and J. I. Markendahl, “RACH dimensioning for reliable MTC over cellular networks,” in IEEE 85th Vehicular Technology Conference (Spring), May 2017. [79] A. Azari, P. Popovski, C. Stefanovic, and C. Cavdar, “Grant-free radio access for cellular IoT,” submitted to IEEE IEEE Transactions on Wireless Communications, 2018. [80] 3GPP TS 36.213, “Evolved universal terrestrial radio access, physical layer procedures,” Tech. Rep., May 2016, (Rel. 13). [81] 3GPP, “Study on provision of low-cost machine-type communications (MTC) user equipments (UEs),” 3GPP TR 36.888 V12.0.0, July 2013. [82] I. C. Wong, O. Oteri, and W. McCoy, “Optimal resource allocation in uplink SC-FDMA systems,” IEEE Trans. Wireless Commun., vol. 8, no. 5, pp. 2161– 2165, May 2009. [83] A. Azari and G. Miao, “Fundamental tradeoffs in resource provisioning for IoT services over cellular networks,” in IEEE International Conference on Communications, May 2017. [84] ——, “Battery lifetime-aware base station sleeping control with M2M/H2H coexistence,” in 2016 IEEE Global Communications Conference, Dec. 2016. [85] R1-163510 , “Candidate NR Multiple Access Schemes ,” Tech. Rep., April 2016, 3GPP TSG RAN WG1 Meeting 84, Busan, Korea. [86] A. Azari, “Energy-efficient scheduling and grouping for machine-type communications over cellular networks,” Ad Hoc Networks, vol. 43, pp. 16 – 29, June 2016. [87] F. Baccelli, B. Blaszczyszyn, and P. Muhlethaler, “An Aloha protocol for multihop mobile wireless networks,” IEEE Trans. on Inf. Theory, vol. 52, no. 2, pp. 421–436, Feb. 2006. [88] J. Fang et al., “Fine-grained channel access in wireless LAN,” IEEE/ACM Trans. on Networking, vol. 21, no. 3, pp. 772–787, Aug 2013. [89] M. S. Talebi Mazraeh Shahi, “Minimizing regret in combinatorial bandits and reinforcement learning,” Ph.D. dissertation, KTH Royal Institute of Technology, Dec. 2017.

88

BIBLIOGRAPHY

[90] F. Cuomo et al., “EXPLoRa: extending the performance of LoRa by suitable spreading factor allocations,” in IEEE 13th International Conference on Wireless and Mobile Computing, Networking and Communications, Oct. 2017. [91] H. Ye et al., “Deep reinforcement learning based resource allocation for V2V communications,” arXiv preprint arXiv:1805.07222, May 2018. [92] C. She, C. Yang, and T. Q. Quek, “Radio resource management for ultrareliable and low-latency communications,” IEEE Communications Magazine, vol. 55, no. 6, pp. 72–78, June 2017. [93] B. Spector and S. Belongie, “Sample-efficient reinforcement learning through transfer and architectural priors,” arXiv preprint arXiv:1801.02268, Jan. 2018. [94] S. Samarakoon, M. Bennis, W. Saad, and M. Debbah, “Federated learning for ultra-reliable low-latency V2V communications,” arXiv preprint arXiv:1805.09253, May 2018. [95] A. Ferdowsi and W. Saad, “Deep learning for signal authentication and security in massive internet of things systems,” arXiv preprint arXiv:1803.00916, March 2018. [96] A. Azari and C. Cavdar, “Performance evaluation and optimization of LPWA IoT networks: A stochastic geometry approach,” in IEEE Global Communications Conference, Dec. 2018.

Appendix A

Relevant Publications Excluded Publications In this section, we list publications which have been published during the PhD project, but their contributions have been excluded from the thesis.

Enhancing IoT communications over LTE networks • A. Azari, Energy-efficient scheduling and grouping for machine-type communications over cellular networks, Ad Hoc Networks, vol. 43, pp. 16 – 29, June 2016. • M. I. Hossain, A. Azari, and J. Zander, DERA: Augmented Random Access for Cellular Networks with Dense H2H-MTC Mixed Traffic, in 2016 IEEE Global Communications Workshops, Dec. 2016. • M. I. Hossain, A. Azari, J. Markendahl, and J. Zander, Enhanced Random Access: Initial access load balance in highly dense LTE-A networks for multiservice (H2H-MTC) traffic, in IEEE International Conference on Communications, May 2017.

Grant-free radio access for IoT communications • M. Masoudi, A. Azari, E. A. Yavuz, and C. Cavdar, Grant-free Radio Access IoT Networks: Scalability Analysis in Coexistence Scenarios, IEEE International Conference on Communications, May 2018. • A. Azari et al., Inter-Network Coordination for Coexisting IoT Networks, Deliverable for the SooGreen Project, Aug. 2017. 89

90

APPENDIX A. RELEVANT PUBLICATIONS

Partially Covered Publications In this section, we list relevant publications which have not been appended to the thesis, but (parts) of their contributions have been used in writing of the thesis.

Enhancing IoT communications over LTE networks • G. Miao, A. Azari, and T. Hwang, E 2 -MAC: Energy efficient medium access for massive M2M communications, IEEE Transactions on Communications, vol. 64, no. 11, pp. 4720–4735, Sept. 2016. • A. Azari and G. Miao, Energy efficient MAC for cellular-based M2M communications, in 2014 IEEE Global Conference on Signal and Information Processing, Dec. 2014. • A. Azari and G. Miao, Lifetime-Aware Scheduling and Power Control for M2M Communications in LTE Networks, in IEEE Vehicular Technology Conference (Spring), May 2015. • A. Azari and G. Miao, Lifetime-aware scheduling and power control for cellularbased M2M communications, in IEEE Wireless Communications and Networking Conference, March 2015. • A. Azari and G. Miao, Battery Lifetime-Aware Base Station Sleeping Control with M2M/H2H Coexistence, in 2016 IEEE Global Communications Conference, Dec. 2016. • A. Azari and G. Miao, Fundamental tradeoffs in resource provisioning for IoT services over cellular networks, in IEEE International Conference on Communications, May 2017. • A. Azari, M. I. Hossain, and J. I. Markendahl, RACH dimensioning for reliable MTC over cellular networks, in IEEE 85th Vehicular Technology Conference (Spring), 2017.

Grant-free radio access for IoT communications • A. Azari and C. Cavdar, Performance Evaluation and Optimization of LPWA IoT Networks: A Stochastic Geometry Approach, in IEEE Global Communications Conference, Dec. 2018. • A. Azari, P. Popovski, G. Miao, and C. Stefanovic, Grant-free radio access for short-packet communications over 5G networks, in IEEE Global Communications conference, Dec. 2017.

91

Appended Publications In this section, we list publications which have been appended to the second part of the thesis. • A. Azari and A. Bria, System and method for providing communication rules based on a status associated with a battery of a device, Nov. 2017, PCT/SE2017/050933 (filed). • A. Azari, G. Miao, C. Stefanovic, and P. Popovski, Latency-Energy Tradeoff based on Channel Scheduling and Repetitions in NB-IoT Systems, IEEE Global Communications Conference, Dec. 2018. • A. Azari, M. Masoudi, and C. Cavdar, Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks, submitted to IEEE Transactions on Communications, Sept. 2018. • A. Azari, P. Popovski, C. Stefanovic, and C. Cavdar, Grant-Free Radio Access for Cellular IoT, submitted to IEEE Transactions on Wireless Communications, Aug. 2018. • A. Azari and C. Cavdar, Self-organized Low-power IoT Networks: A Distributed Learning Approach, in IEEE Global Communications Conference, Dec. 2018. • A. Azari, M. Ozger, and C. Cavdar, Serving Non-Scheduled URLLC Traffic: Challenges and Learning-Powered Strategies, submitted to IEEE Communications Magazine, Aug. 2018.

Reprint of Publications

93

Patent: System and Method for Providing Communication Rules Based on a Status Associated with a Battery of a Device Amin Azari and Aurelian Bria (Ericsson) Note: The summary of the invention is presented here for the brevity of thesis. Filed patent: number P72269-WO1.

95

P72269-WO1 SYSTEM AND METHOD FOR PROVIDING COMMUNICATION RULES BASED ON A STATUS ASSOCIATED WITH A BATTERY OF A DEVICE

TECHNICAL FIELD The disclosure relates to communication systems and, more particularly, to a system and method for providing communication rules based on a status associated with a battery of a device.

BACKGROUND The Narrowband Internet of Things (“NB-IoT”) is a narrowband system developed for the cellular Internet of Things by the Third Generation Partnership Program (“3GPP”). The system is based on existing Long Term Evolution (“LTE”) systems, and addresses improved network architecture and coverage for substantial number of devices with characteristics such as lower throughput (e.g., two kilobits per second (“kbps”)), lower delay sensitivity (e.g., 10 seconds), lower cost (e.g., below 5 dollars) and lower power consumption (e.g., battery life of 10 years). It is envisioned that each cell (about one square kilometer) in this system can serve thousands (e.g., 50,000) of devices such as sensors, meters, actuators, and other devices. In order to make use of an existing spectrum such as Global System for Mobile Communications (“GSM”), a fairly narrow bandwidth (e.g., 180 kilohertz (“kHz”), which may be similar to the LTE Physical Resource Block (“PRB”)) has been adopted for NB-IoT technology. The entire (or a substantial amount of) NB-IoT traffic can be contained within 200 kHz or one physical resource block, which may be 12 subcarriers of 15 kHz each (in NB-IoT, this is referred to as one carrier or one PRB). 1

P72269-WO1 Existing cellular infrastructures have been designed for a limited set of services including voice/video calling, and web surfing. Each connected device (e.g., a smartphone) may transmit/receive different types of traffic such as web or voice traffic. To satisfy the quality-ofservice (“QoS”) requirements of each service, the 3GPP LTE has defined some quality class identifiers (“QCIs”), and provide a QCI to each traffic type, as described in the European Telecommunications Standards Institute Technical Specification (“ETSI TS”) 123 203, entitled “Digital cellular telecommunications system (Phase 2+) (GSM); Universal Mobile Telecommunications System (UMTS); LTE; Policy and charging control architecture (3GPP TS 23.203 version 13.6.0 Release 13),” (2016-03), which is incorporated herein by reference. The QCIs (see, e.g., Table 6.1.7: Standardized QCI characteristics, pages 47-48 of the above referenced document) are used in the scheduling procedures for prioritizing different types of traffic for uplink/downlink service. The existing QCIs have been mostly designed either (i) to guarantee a traffic-dependent bitrate for data-hungry services, or (ii) to provide ultra-low latency communications for ultra-high priority services. The existing QCIs have been adapted for downlink, while a large part of IoT communication is on uplink, and consists of a substantial number of short-lived sessions. The existing QCIs are not tailored to support the connection establishment phase (i.e., to associate special random access channel (“RACH”) resources for critical QCIs to reduce energy loss in contentions) and uplink scheduled transmissions (i.e., the amount of allocated resources, priority in resource allocation, and signal-to-noise ratio (“SNR”) budget). Apart from serving machine-type communications (“MTC”) along human-oriented communications (“HoC”) traffic, systems like NB IoT have been addressed in LTE Rel. 13, in which it has been proposed to allocate a narrow frequency band to MTC, divide this band further

2

P72269-WO1 to narrowband sub-channels, and allocate them to MTC devices, as described in the 3GPP Technical Requirements Document (“TR”) 45.820, entitled ‘Technical Specification Group GSM/EDGE Radio Access Network; Cellular system support for ultra-low complexity and low throughput Internet of Things (CIoT),” v13.1.0 (2015-11), which is incorporated herein by reference. MTC scheduling in NB-IoT is done in the same way as in previous LTE releases (i.e., based on the allocated QCIs to the traffic). However, energy consumption in NB-IoT is expected to decrease significantly due to the extra available link budget in narrowband communications.

SUMMARY These and other problems may be generally solved or circumvented, and technical advantages may be generally achieved, by advantageous embodiments that, for example, include a system and method for providing communication rules based on a status associated with a battery of a device such as a user equipment. In one embodiment, an apparatus, and related method, operative to communicate with user equipment in a communication system is configured to receive a status including a remaining charge associated with a battery of the user equipment. The apparatus is also configured to provide communication rules for the user equipment to manage a utilization of the battery based on the status of the battery. In another embodiment, an apparatus, and related method, in a communication system is configured to provide a status including a remaining charge associated with a battery of the apparatus. The apparatus is also configured to execute communication rules to manage a utilization of the battery based on the status of the battery. The foregoing has outlined rather broadly the features and technical advantages of the present examples in order that the detailed description that follows may be better understood.

3

P72269-WO1 Additional features and advantages of various examples will be described hereinafter, which form the subject of the claims of the invention. It should be appreciated by those skilled in the art that the conception and specific embodiments disclosed may be readily utilized as a basis for modifying or designing other structures or processes for carrying out the same purposes of different embodiments. It should also be realized by those skilled in the art that such equivalent constructions do not depart from the spirit and scope of the disclosure as set forth in the appended claims.

4

Paper A: Network Lifetime Maximization for Cellular-Based M2M Networks Amin Azari and Guowang Miao in IEEE Access, vol. 5, 2017, pp. 18927–18940. (©2017 IEEE. Reprint with permission.) The conference versions of this paper have been published in IEEE VTC 2015 [42], and IEEE WCNC 2015 [43].

101

Network Lifetime Maximization for Cellular-Based M2M Networks Amin Azari and Guowang Miao KTH Royal Institute of Technology Email: {aazari, guowang}@kth.se

Abstract—High energy efficiency is critical for enabling massive machine-type communications (MTC) over cellular networks. This work is devoted to energy consumption modeling, battery lifetime analysis, lifetime-aware scheduling and transmit power control for massive MTC over cellular networks. We consider a realistic energy consumption model for MTC and model network battery-lifetime. Analytic expressions are derived to demonstrate the impact of scheduling on both the individual and network battery lifetimes. The derived expressions are subsequently employed in the uplink scheduling and transmit power control for mixed-priority MTC traffic in order to maximize the network lifetime. Besides the main solutions, low-complexity solutions with limited feedback requirement are investigated, and the results are extended to existing LTE networks. Also, the energy efficiency, spectral efficiency, and network lifetime tradeoffs in resource provisioning and scheduling for MTC over cellular networks are investigated. The simulation results show that the proposed solutions can provide substantial network lifetime improvement and network maintenance cost reduction in comparison with the existing scheduling schemes. Index Terms—Internet of Things, Machine to Machine Communications, Scheduling, Energy Efficiency, Resource Allocation.

I. I NTRODUCTION NTERNET of Things (IoT) refers to the ever-growing network of uniquely identifiable smart physical objects that are capable of sensing or acting on their environment. Cellular IoT, IoT embedded in cellular network infrastructure, is expected to play a critical role in the success of IoT because cellular networks provide ubiquitous coverage and roaming [1]. Cellular machine-to-machine (M2M) communications, also known as machine-type communications (MTC), means the communications of machine devices in cellular networks without human intervention and serves as the foundation of cellular IoT. The continuing growth in demand from cellularbased M2M communications encourages mobile network operators to investigate evolutionary and revolutionary radio access technologies for accommodating M2M traffic in cellular networks [1]. M2M communications are generally characterized by the massive number of concurrent active devices, small payload size, and vastly diverse quality-of-service (QoS) requirements [2]. Moreover, in many M2M applications smart devices are battery driven and once deployed, their batteries will never be replaced. Then, long battery lifetime is crucial for them, especially when deployed in remote areas. Based on the 5G envision by Nokia [3], bit-per-joule energy efficiency for machine-type communications must be improved by a factor of ten in order to provide battery lifetimes around 10 years.

I

A. Literature Study 1) MTC over Cellular Networks: Random access channel (RACH) of the LTE-Advanced (LTE-A) is a typical way for machine nodes to directly access the base station (BS). The capacity limit of RACH for serving M2M communications is investigated in [4], and it is shown that RACH is neither a scalable nor an energy-efficient access scheme for massive M2M communications. Clustered-access is investigated in [5] to reduce congestion in an overloaded condition. In [6], access class barring with multiple transmit power levels is introduced in order to reduce congestion using the capture effect at the BS. In [7], energy efficient random access for machine nodes in a multi-cell scenario is investigated, where the choice of serving BS and transmit power level are to be optimized. When a device successfully passes the RACH, it can send scheduling request to the BS through the physical uplink control channel (PUCCH). Then, the BS performs the scheduling and sends back the scheduling grants through the corresponding physical downlink control channel (PDCCH). Now, the granted machine node is able to send data over the granted physical uplink shared channel (PUSCH). This scheduling procedure performs well in existing cellular networks for a limited number of long communications sessions, such as voice and web streaming. However, regarding the fundamental differences in characteristics and QoS requirements of M2M communications, it is evident that the presented scheduling procedure cannot survive with a massive number of short-lived M2M communications sessions. The 3GPP LTE has defined some research projects to support low-cost massive machine-type communications in cellular networks. The development of LTE for low-cost massive MTC has been initiated in release 12, and will be continued in release 13 [8]. The target for LTE release 13 includes LTE category M (LTE-M), and narrow-band LTE-M (NB LTE-M) deployments, which are expected to offer MTC over 1.4 MHz and 200 KHz bandwidths [8]. 2) MTC Scheduling over Cellular Networks: Scheduling is the process performed by the BS to assign radio resources to UEs. In general, scheduling is not part of the standardization work, and is left for vendor implementation. However, signaling is standardized, and hence, any scheduling scheme should comply with the control requirements in the standards. Regarding the limited capacity of PDCCH, the number of UEs that can be served at once are limited. Then, the scheduling problem can be broken into two subproblems: (i) time domain scheduling, in which a subset of devices is chosen to be

scheduled; and (ii) frequency domain scheduling, in which the available resource elements are allocated to the selected subset of UEs. A thorough survey on LTE scheduling algorithms for M2M traffic is presented in [9]. This survey indicates that existing scheduling algorithms could be categorized into 4 main categories with regard to the scheduling metric as follows [9]: (i) channel-based schedulers, in which UEs with the highest signal to noise ratio (SNR) have priority in resource allocation in order to minimize the bit error rate and maximize the system throughput [10]; (ii) delay-based schedulers, in which the delay budget prioritize devices for resource allocation [11, 12]; (iii) fairness-based schedulers, which are designed to guarantee a fair distribution of radio resources among UEs [13]; and (iv) hybrid schedulers, which consider a combination of the aforementioned metrics as well as other metrics like power consumption [14], buffer status, and data arrival rates [9]. 3) Energy-Efficient MTC Scheduling: While providing scalable yet energy efficient communications is considered as the key requirement for successful deployment of MTC over existing cellular networks [3, 4], a limited number of research works has been focused on energy efficient uplink MTC scheduling. Energy efficiency of M2M communications over LTE networks is investigated in [15], and it is shown that LTE physical layer is not optimized for small data communications. Power-efficient uplink scheduling for delay-sensitive traffic over LTE systems is investigated in [16], where the considered traffic and delay models are not consistent with the MTC characteristics [2], and hence, the derived results cannot be used here. Power-optimized resource allocation for time, frequency, and code division multiple access (TDMA, FDMA, CDMA) systems has been investigated in [17]. Uplink scheduling for LTE networks with M2M traffic is investigated in [14], where the ratio between the sum data rates and the power consumptions of all users is maximized. In [14], the authors have considered a simple model for energy consumption considering only the transmit power for reliable data transmission and neglected the other energy consumptions by the operation of electronic circuits which are comparable or more dominant than the energy consumption for reliable data transmission [18]. In [19], a clean slate solution for dense machine deployment scenarios is proposed in which, each communications frame is divided into two subframes. The first subframe is dedicated to the contention of machine nodes for access reservation, and the later is dedicated to scheduled data transmission of successful nodes using TDMA scheme. To the best of our knowledge, accurate modeling of energy consumption in machine-type communications, individual and network battery lifetime models, and corresponding scheduling algorithms are absent in literature. As an extension of [5], which investigates clustered-access for massive M2M, in [20] joint energy efficient clustering and scheduling has been investigated, i.e. the cluster-size, selection of clusterheads, and the amount of scheduled resources to clusterheads have been optimized to prolong the battery lifetime. In [21, 22], preliminary studies on feasibility of battery lifetimeaware scheduling for unclustered M2M communications have been presented, and two exhaustive search algorithms for scheduling over frequency domain resources have been de-

veloped. Substantial extension to [21, 22] has been made in this paper, where sophisticated scheduling algorithms over time/frequency resources along with low-complexity and limited feedback solutions have been developed under different network lifetime definitions. Furthermore, detailed analytical analysis, derivation of closed-form scheduling expressions, and complexity and fairness analysis have been presented in this paper.

B. Contributions The main contributions of this paper include: •

•

•

•

•

Introduce accurate energy consumption, and individual and network lifetime models for machine-type devices deployed in cellular networks by taking both transmission and circuit energy consumptions into account. Present a battery lifetime aware resource allocation framework. Explore MTC scheduling based on the MaxMin lifetime-fairness, and analyze its contribution in reducing the maintenance costs of M2M networks. Present uplink scheduling solutions for MTC over singlecarrier frequency division multiple access (SC-FDMA) systems. Present low-complexity scheduling solutions with limited feedback requirement. Figure out the energy efficiency, spectral efficiency, and network lifetime tradeoffs in uplink MTC resource provisioning and scheduling. Extend the proposed solutions for existing 3GPP LTE networks. Present lifetime-improvement evidence using simulation results in the context of LTE.

The rest of this paper is organized as follows. In the next section, the system model is presented. The battery lifetimeaware scheduling framework, and the coupling between network lifetime and control parameters are presented in section III by investigating M2M scheduling in time domain for narrow band cellular M2M networks. The general scheduling problem with time/frequency domain radio resources is investigated in section IV. Low complexity scheduling solutions with limited feedback requirement are investigated in section V. As an example of lifetime-aware scheduling, in section VI we apply the derived solutions in section IV-V to the 3GPP LTE networks, and provide simulation results in section VII in order to demonstrate the lifetime improvement. Concluding remarks are given in section VIII.

II. S YSTEM M ODEL Consider a single cell with one base station and a massive number of machine nodes, which are uniformly distributed in the cell. The machine nodes are battery driven and once deployed, their batteries won’t be replaced, then long batterylifetime is crucial for them. Consider the uplink scheduling problem at time t, where a set of devices, denoted by A, is to be served using a limited set of resources. As we aim at deriving network lifetime maximizing solutions, both individual and network battery-lifetime metrics are defined.

A. Lifetime Metric For most reporting MTC applications, the packet generation at each device can be modeled as a Poisson process [23], and hence, the energy consumption of a device can be seen as a semi-regenerative process where the regeneration point is at the end of each successful data transmission. For node i, the remaining energy at time t is denoted by Ei (t), the average payload size by Di , and the power consumption in transmission mode by ξPi + Pc , where Pc is the circuit power consumed by electronic circuits, ξ is the inverse of power amplifier (PA) efficiency, and Pi is the transmit power for reliable data transmission. We define the expected lifetime for node i at the regeneration point as the multiplication of reporting period by the ratio between remaining energy and the average energy consumption per reporting period, as follows: ∆

Li (t) =

Ei (t) Ti , Esi + Edi

(1)

where Edi is the average energy consumption per reporting period for data transmission: Edi = [Pc + ξPi ]Di /Ri , Ri is the average data transmission rate, Ti the expected length of one reporting period, and Esi the average static energy consumption in each reporting period for data gathering, processing, and etc. B. Network Lifetime Definition The network lifetime is the time between the reference time and when a network is considered to be nonfunctional. The instant at which an M2M network is considered to be nonfunctional is application-specific. For example, in safetycritical applications where losing even one node deteriorates the performance or coverage, or in sparse sensor deployments where correlation between gathered data by different nodes is low, the shortest individual lifetime (SIL) may specify the network lifetime. In other cases, e.g. where correlation between gathered data by different nodes is high, the longest individual lifetime (LIL) or the average individual lifetime (AIL) may be defined as the network lifetime. Here, we present our derivations for the case in which, the shortest individual lifetime is considered as the network lifetime, i.e. Lsil net (t) = mini Li (t); however, as we will show in section IV and VII, our proposed lifetime-aware resource allocation framework can be also used with other network lifetime definitions. III. S CHEDULING FOR NARROW-BAND M2M N ETWORKS To facilitate the understanding of the fundamental dependence of network lifetime on the remaining energy, reporting period, channel gain, circuit power, and bandwidth, here we investigate SIL-aware scheduling for a narrow-band M2M system. Examples of such systems are 2G GSM-based M2M networks, LTE networks in which a specific carrier is reserved for an MTC application, and proprietary M2M networks. In these systems, as at most one node occupies the whole bandwidth at each time, the uplink scheduling is equivalent to

finding the transmission time for each node. Denote the length of the resource pool in time domain as τ , the bandwidth as w, and the allocated fraction of time for data transmission of node i as τi . Then, the lifetime expression for node i is found from (1), where Edi = τi [Pc +ξPi ]. Denote the signal-tointerference-plus-noise ratio (SINR) for reliable transmission of Di bits in τi seconds to be ηi = S(Di /τi ). For example, using Shannon capacity formula the data rate function is derived as: ηi ), Ri = w log(1 + Γmcs and correspondingly S(x) is derived as: [ x ] S(x) = 2 w − 1 Γmcs .

(2)

In this expression, Γmcs is the SNR gap between the channel capacity and a practical modulation and coding scheme (MCS), as investigated in [18]. One sees in (2) that S(x) is strictly convex in x and S(0) = 0. Thus, we do not choose a specific MCS, and hence, do not specify the exact form of S(x) in our analysis. Instead, we only assume S(x) to be strictly convex in x and S(0) = 0. Denote the channel gain between node i and the BS as hi . Then, the required transmit power for node i will be Pi = ηi [N0 + I]w/[hi Gtr ], where Gtr is the multiplication of transmit and receive antenna gains, and the power spectral densities (PSDs) of noise and interference at the receiver are denoted by N0 and I, respectively. Then, the scheduling optimization problem that maximizes the network lifetime is formulated as follows: maximizeτi Lsil net (t) ∑ τi ≤ τ, subject to: C.3.1:

(3)

i∈A

C.3.2: τim ≤ τi

∀i ∈ A,

τim

where is the minimum required transmission time, and is found as a function of maximum allowed transmit power Pmax , by solving the following equation: S(

Di Pmax hi Gtr )= . τim [N0 + I]w

One can define Z as an auxiliary variable where Z = maxi∈A Li1(t) , and rewrite (3) as: minimizeτi Z subject to: C.3.1, C.3.2, and

(4) 1 ≤ Z, Li (t)

∀i ∈ A.

Taking the second derivative of the inverse lifetime expression, ξ[N0 + I]w Di2 ¨ ∂ 2 1/Li = S(Di /τi ), ∂τi2 Ei (t)Ti hi Gtr τi3 one sees that it is a strictly convex function of τi because ¨ S(x) > 0, where f˙(x) and f¨(x) show the first and second derivatives of function f (x) respectively. Thus, Z is also a strictly convex function of τi because the point-wise maximum operation preserves convexity [24]. Then, the scheduling problem in (4) is a convex optimization problem, and can be solved

using convex optimization tools, as has been investigated in appendix A. In the special case that S(x) is found from (2) and Γmcs = 1, the real-valued solution of (31) is found from appendix A as: { τi∗ = max τim ,

w + L( 1e

ln(2)Di [ [hi Gtr ][Pc +Ti Ei (t)µ/λi ] ξ(N0 +I)w

} , ] − 1 )w (5)

where e is the Euler’s number, and L(x) is the LambertW function, i.e. inverse of the function f (x) = x exp(x) [25]. Motivated by the facts that: (i) scheduling is done in the time domain; and (ii) nodes which are more critical from network battery lifetime point of view receive a longer transmission time than the minimum required transmission time in order to decrease their transmission powers; the expression in (34) represents priority of nodes in uplink scheduling when network battery lifetime is to be maximized. From (34), one sees that the priority of nodes in lifetime-aware scheduling: 1 1 ∗ • increases with E (t) and T , because τi increases when i i the remaining energy, i.e. Ei (t), decreases or the packet generation rate, i.e. 1/Ti , increases. 1 ∗ • increases with h , because τi increases for devices dei ployed far from the BS or compensate a large pathloss, in order to reduce the transmit power and save energy. ∗ • increases with Di , because τi increases in the size of buffered data to be transmitted. In the case that there is no constraint on the amount of available radio resources, the optimal transmission time is found as: { ln(2)Di τi∗ = max τim , [ Pc [hi Gtr ] ] }. w + L( 1e ξ(N − 1 )w 0 +I)w In this case, when the circuit power consumption increases, the optimal transmission time decreases, and hence, the transmit power increases. Then, for nodes in which the circuit power is so high that is comparable with the transmit power, it is more energy efficient to transmit data with a higher transmit power in order to finish data transmission in a shorter time interval, and hence, reduce the circuit energy consumption. Also, the scheduler must provide time-domain scheduling priority for these devices to decrease their waiting time before receiving services, which decreases their energy consumption in the idle listening to the base station. Based on these preliminary insights to the lifetime-aware scheduling problem, in the next section MTC scheduling in both time and frequency domains for SC-FDMA systems is investigated. IV. MTC S CHEDULING OVER SC-FDMA

[28]. Consider the scheduling problem at time t, where a set of nodes, i.e. A, with cardinally |A| are to be scheduled for uplink transmission. Denote the set and total number of available chunks as C and |C|, where each chunk consists of M adjacent subcarriers with a time duration of τ . Here, we investigate time- and frequency-domain scheduling, i.e. if the number of resource elements is not sufficient to schedule A at once, a subset of A is selected, and then, the available resources are assigned to this subset using a frequency-domain scheduler. The effective SINR for a SC-FDMA symbol is approximated as the average SINR over the set of allocated subcarriers because each data symbol is spread over the whole bandwidth [16]. Then, the achievable data rate for node i is written as: R(Ci , Pi ) = |Ci |M Sv (ηi ),

(6)

where Ci represents the set of allocated chunks to node i, |Ci | Pi Gtr hei , Sv (x) is the inverse of the cardinality of Ci , ηi = |C i |M S(x) which is a strictly convex and increasing function of x, and hence, Sv (x) is a strictly concave function of x [24]. Also, hei is the effective channel gain-to-interference-plus-noise ratio for node i and is defined as [16] [ ∑ [N0 + Ij ]M w ] hei = |Ci |/ , hji j∈Ci

where hji is the channel gain2 of node i over chunk j, Ij is the PSD of interference on chunk j, and w is the bandwidth of each subcarrier. Thus, using (1) the expected lifetime of node i at time t + τ is formulated as a function of Pi and Ci as follows: Li (t + τ ) =

Ei (t + τ ) − [1 − θi ]Edi Ti , Esi + Edi

(7)

where Ei (t + τ ) = Ei (t) − τ Pc [1 − θi ] −

[ ] Di Pc + ξPi θi , R(Ci , Pi ) (8)

θi is 1 if node i is scheduled with |Ci | > 0 and 0 otherwise, and [1 − θi ]Edi is the expected energy consumption from t + τ till the successful transmission, i.e. the regeneration point. Now, one can formulate the scheduling problem as: maximizeCi ,Pi ,θi Lsil net (t + τ ) ∑ subject to: C.9.1: |Ci | ≤ |C|,

(9)

i∈A

C.9.2: Ci : contiguous ∀i ∈ A, C.9.3: Ci ∩ Cj = ∅ ∀i, j ∈ A, i ̸= j, C.9.4: θi Di /R(Ci , Pi ) ≤ τ ∀i ∈ A, C.9.5: Pi ≤ Pmax ∀i ∈ A, C.9.6: [1 − θi ]Qi = 0 ∀i ∈ A,

SC-FDMA is a favorite multiple access scheme for energylimited uplink communications. Using SC-FDMA implies that: (i) only G clusters of adjacent subcarriers can be allocated to each node1 ; (ii) the transmit power over all assigned subcarriers to a node must be the same [16]; and (iii) subcarriers are grouped into chunks, before being assigned to the nodes

where ∅ is the empty set, and Qi ∈ {0, 1} is a binary parameter used for prioritizing traffics from high-priority nodes or traffics with zero remaining delay budgets to be transmitted immediately. The scheduler gives the highest priority to the

1 This is the contiguity constraint [26]. In existing LTE-A networks, G = 1, 2 are used. As PAPR and spectral efficiency increase in G [27], for MTC applications G = 1 is preferred.

2 The channel gain of a user over all subcarriers of one chunk is assumed to be constant in (t, t + τ ) [16].

traffic from node i if Qi = 1. If Qi = 0, node i will be either scheduled by using the remaining radio resources from scheduling high-priority traffic or will be scheduled in later time slots. In (9), C.9.1 is due to the limited set of available uplink radio resources, C.9.2 is due to the contiguity requirement in SC-FDMA, C.9.3 assures each resource element will be at most allocated to one device, C.9.4-5 assure that the Di bits of data can be transmitted over the assigned set of contiguous resources with a transmit power lower than the maximum allowed transmit power, and C.9.6 assures that high priority traffic will be scheduled ahead of low-priority traffic. One sees that the optimization problem in (9) is not a convex optimization problem due to the contiguity constraint, as discussed in [16]. The straightforward solution for this problem consists in reformulating the problem as a pure binary-integer program. From [28], we know that the complexity of search over all feasible resource allocations for SC-FDMA systems is ∑|A| (|A| ) (|C|−1 ) i! i−1 . (10) i i=1

Thus, the straightforward approach is clearly complex, especially for cellular networks with massive machine-type communications. While shifting complexity from device-side to network-side is feasible for enabling massive IoT connectivity in beyond 4G era [29], and quantum-assisted communication for realizing such receivers has been introduced in [30, 31], here we focus on deriving low-complexity scheduling solutions which are applicable even in existing cellular infrastructures. In the following, we present a low-complexity lifetimeaware scheduling solution. A. Low-Complexity Scheduling Solution: The Prerequisites Before presenting our proposed solution, in the sequel a set of prerequisites are investigated. 1) Optimized transmit power for a given scheduling: For a fixed chunk assignment to node i, i.e. Ci , one can use the results presented in [32, section III] to prove that the presented energy consumption expression for time interval [t, t + τ ] in (8) is a strictly quasiconvex function of Pi . Then, one can find the transmit power that minimizes the energy consumption in this interval. Using convex optimization theory, the optimal transmit power is found as the maximum of Pmin and the solution to the following equation: Sv (KPi ) − K[Pc /ξ + Pi ]S¨v (KPi ) = 0, where K = solution to:

Gtr hei |Ci |M ,

and from C.9.4, one can drive Pmin as the

Sv (Pmin K) = Di /[τ |Ci |M ]. As an example, when S(x) is found from (2), the optimal transmit power is: } { { ] 1 [ KPc /ξ − 1 −1 , P(Ci ) = min Pmax , max Pmin , KP /ξ−1 K L( c ) e (11) [ ] Di and Pmin is found as: Pmin = 2 τ M |Ci |w − 1 /K.

2) Scheduling metric: Let us consider the scheduling problem at time t, where |C| − 1 chunks are already allocated to the nodes, Ci shows the set of already assigned resources to node i, and one further chunk is available to be allocated to the already scheduled or non-scheduled nodes. Given Ci and θi , the expected lifetime of node i at t + τ is derived from (7)-(8), as follows: ∆

F(Ei (t), Edi , Esi , Ti , Ci , θi ) = Li (t + τ ) = (12) [ ] Di Pc +ξP(Ci ) i ( ) θi − [1 − θi ]Ed Ei (t) – τ Pc [1 − θi ] − R Ci ,P(Ci )

Ti .

Esi + Edi

When the SIL network lifetime is to be maximized, the index of node with highest priority to be scheduled is found as follows: sil Lsil net (t + τ )|opt. scheduling > Lnet (t + τ )|any scheduling ∗

∗

→ min{Li (t + τ ) + ∆Lii } > min{Lj (t + τ ) + ∆Ljj }, i∈A

j∈A

where i∗ is the index of selected node using optimal scheduler, j ∗ the index of selected nodes using any other scheduler, ∆Lji the change in the lifetime of node i when the extra chunk is allocated to node j, and ∆Lji = 0 for i ̸= j. Then, we have: ∗

min{Li∗ + ∆Lii∗ , Lj ∗ ,

min

i∈A\{i∗ ,j ∗ }

min{Li∗ , Lj ∗ +

∗ ∆Ljj ∗ ,

Li } > min

j∈A\{i∗ ,j ∗ }

Lj },

∗

∗

→ min{Li∗ + ∆Lii∗ , Lj ∗ } > min{Li∗ , Lj ∗ + ∆Ljj ∗ }, (13) where the time index is dropped for the sake of notational continence. One sees that the only choice of i∗ that satisfies (13) for any choice of j ∗ at time t is i∗ (t) = arg min Li (t + τ ) i∈A

= arg min F(Ei (t), Edi , Esi , Ti , Ci , θi ),

(14)

i∈A

i∗ i∗

if it satisfies the lifetime improvement constraint: ∆L > 0, i.e. its battery lifetime can be improved by assigning the new chunk. If the battery lifetime of i∗ (t) cannot be improved, we remove i∗ from A, and repeat the criterion in (14) in order to find node with the shortest lifetime that can improve its lifetime using more chunks. Denote by VX the expected battery lifetime vector, where X is the scheduling criterion, and the ith element of VX shows the expected battery lifetime of node i under criterion X . Definition A feasible scheduling satisfies the max-min fairness criterion if no other scheduling has a lexicographically s s greater sorted lifetime vector, i.e. Vmax-min ≥lex Vany , where the superscript s shows sorting in non-decreasing order [33]. In other words, this definition means that if we schedule machine nodes under criterion X , derive their expected battery lifetimes after scheduling, and sort the battery lifetimes in vector VX ; the resulting lifetime vector from max-min fairness scheduling is lexicographically greater than the resulting lifetime vector from any other scheduling criterion. Then, the smallest resulting battery lifetime from the max-min fairness

criterion will be as large as possible, and the second-smallest resulting battery lifetime will be as large as possible, and so on. Comparing the proposed iterative structure in this subsection for prioritizing nodes in lifetime-aware scheduling with the definition IV-A2 shows that the proposed scheduling procedure achieves the max-min fairness. This is due to the fact that our scheduler first schedules node with the shortest expected lifetime, then schedules node with the second shortest lifetime, and etc. As a result, the increase in battery lifetime of the selected node in each phase will not be at the cost of decrease in the battery lifetime of another node with already shorter battery lifetime. B. Lifetime-Aware Scheduling Solution: The Procedure The basic idea behind our proposed solution is breaking the scheduling problem into two subproblems: (i) satisfying the minimum resource requirement for the set of high-priority nodes which must be scheduled at time t, called Ad ; and (ii) resource allocation for all nodes based on their impacts on the network lifetime. Our proposed solution, presented in Algorithm 1, solves the first and second subproblems in step 1 and 2 respectively. The first subproblem is a frequency-domain scheduling problem, where the scheduler allocates contiguous resource elements to the nodes which must be scheduled immediately to satisfy their minimum resource requirements. The second subproblem is a time/frequency scheduling problem. As the minimum resource requirement of Ad has been already satisfied in step 1, the scheduler in step 2 selects node with the highest impact on the network lifetime among the scheduled and non-scheduled nodes, and allocates contiguous resource elements to it in order to prolong its battery lifetime, and hence, maximizes the network lifetime. In Algorithm 1, we call a resource expansion algorithm named ExpAlg, which is presented in Algorithm 2. Given the set of available resources, i.e. C, the set of already allocated resources to the ith node, i.e. Ci , and the maximum allowed number of allocated resource clusters to a node, i.e. G, this algorithm finds the resource element which satisfies the contiguity constraint, and on which, node i has the best SINR.

Algorithm 1: SIL-aware scheduling for SC-FDMA

1

2

i∈A

and sum of the logarithms of individual lifetimes (SLIL) defined as: ∑ Lslil log(Li (t)). net (t) = i∈A

One sees that AIL-aware scheduling aims at maximizing the average battery lifetime of machine devices without providing fairness among them, while SLIL-aware scheduling

Step 2; - Ad ∪ Acd → H; - while C and H are non-empty do - j ∗ = arg minj∈H F(Ei (t), Edi , Esi , Ti , Ci , θi ); - if θj ∗ ̸= 1 then - 1 → θj ∗ ; ( ) - while Dj ∗ /R Cj ∗ , Pmax > τ do - ExpAlg(C, Cj ∗ , G) → c∗ , c∗ ∪ Cj ∗ → Cj ∗ , C \ c∗ → C; - If c∗ = ∅, then 0 → θi , Cj ∗ ∪ C → C, H \ j ∗ → H, exit the loop; else Di [Pc +ξP (Cj ∗ )] → B; - R(C j ∗ ,P (Cj ∗ )) - ExpAlg(C, Cj ∗ , G) → c∗ ; - if c∗ = ∅ then - H \ j ∗ → H; else - c∗ ∪ Cj ∗ → Cj ∗ , C \ c∗ → C; Di [ξP(Cj ∗ )+Pc ] > B, then - If R(C j ∗ ,P(Cj ∗ )) Cj ∗ \ c∗ → Cj ∗ , C ∪ c∗ → C, H \ j ∗ → H;

3 4

C. Extension to Other Network Lifetime Definitions The proposed framework in subsection IV-B can be also used with other network lifetime definitions. When the longest individual lifetime is considered as the network lifetime, the scheduling solution is found from a modified version of Algorithm 1, in which the argmin operator is replaced with the argmax operator. In the sequel, we consider two other network lifetime definitions including: (i) average individual lifetime defined as: ∑ Lail Li (t), net (t) = [1/|A|]

Initialization; - Define Ad , where i ∈ Ad if Qi = 1; - Define Acd as A \ Ad ; - 0 → θi and ∅ → Ci , ∀i ∈ A ; Step 1; - A → Atd ; - while Atd is non-empty do - arg minj∈Atd Lj (t) → j ∗ ; - Atd \ j ∗ → Atd(, 1 → θi ; ) - while Dj ∗ /R Cj ∗ , Pmax ) > τ do - ExpAlg(C, Cj ∗ , G) → c∗ , c∗ ∪ Cj ∗ → Cj ∗ , C \ c∗ → C; - If c∗ = ∅, then 0 → θi , Ad \ j ∗ → Ad , Cj ∗ ∪ C → C, exit the loop;

P(Ci ) → Pi , ∀i ∈ A; return Pi , θi , Ci , ∀i ∈ A

Algorithm 2: Resource expansion algorithm (ExpAlg) - Inputs: C, Ci , G; - Number of existing resource clusters in Ci → n; - if n < G then mi - c∗ = arg maxm∈C Nh0 +I ; m else - Adjacent resource elements to Ci → C; mi - c∗ = arg maxm∈C Nh0 +I ; m return c∗

aims at maximizing the average battery lifetime of machine devices with providing proportional fairness among them. [34, chapter 4]. In order to modify Algorithm 1 for AIL- and SLIL-aware scheduling, one needs to derive the respective

scheduling metrics, which are investigated in the following. 1) Scheduling metric for AIL: Following the discussion in subsection IV-A2, when AIL network lifetime is to be maximized, index of the node with highest priority in scheduling is found as: ail Lail net (t + τ )|opt. scheduling > Lnet (t + τ )|any scheduling ∑ ∑ ∗ i∗ → Li + ∆Li > Lj + ∆Ljj , i∈A

j∈A ∗

→ Li∗ + ∆Lii∗ + Lj ∗ +

∑

Li > Li∗ + Lj ∗

i∈A\i∗ ,j ∗

∑

∗

+ ∆Ljj ∗ +

Lj ,

j∈A\i∗ ,j ∗ ∗

∗

→ ∆Lii∗ > ∆Ljj ∗ ,

(15)

where the time index is dropped for the sake of notational convenience. One sees that the only choice of i∗ that satisfies (15) for any choice of j ∗ at time t is i∗ (t) = arg max ∆Lii ,

(16)

i∈A

= arg max F(Ei (t), Edi , Esi , Ti , Ci+ , 1)− i∈A

F(Ei (t), Edi , Esi , Ti , Ci , θi ), where Ci+ shows the updated set of assigned chunks to node i, i.e. the set of already assigned chunks plus the new chunk. Then, for each available chunk we need to find the node that its lifetime improvement with the extra chunk is higher than the others. 2) Scheduling metric for SLIL: If SLIL network lifetime is the case, index of the node with highest priority to be scheduled is found as follows: slil Lslil net (t + τ )|opt. scheduling > Lnet (t + τ )|any scheduling ∑ ∑ ∗ ∗ → log(Li (t + τ )+∆Lii ) > log(Lj (t + τ ) + ∆Ljj ), i∈A

→

∏

j∈A

∗

Li (t + τ ) + ∆Lii >

i∈A

∏

∗

Lj (t + τ ) + ∆Ljj ,

j∈A ∗

∗

Lj ∗ (t + τ ) + ∆Ljj ∗ Li∗ (t + τ ) + ∆Lii∗ → > . Li∗ (t + τ ) Lj ∗ (t + τ )

(17)

One sees that the only choice of i∗ that satisfies (17) for any choice of j ∗ at time t is ∗ ∆Lii∗

Li∗ (t + τ ) + Li∗ (t + τ ) F(Ei (t), Edi , Esi , Ti , Ci+ , 1) = arg max . i∈A F(Ei (t), E i , Esi , Ti , Ci , θi ) d

i∗ (t) = arg max i∈A

(18)

Comparing (14), (16), and (18), one sees how SIL- and SLILaware scheduling provide max-min and proportional fairness among machine nodes, respectively. D. Performance Analysis In the outer loop of the first step of Algorithm 1, the scheduler iterates over the set of prioritized nodes. In each iteration, it satisfies the minimum resource requirement of a selected node, and hence, the maximum number of iterations

in step 1 will be |Cd | ≤ |C|, where Cd is the set of allocated resources to Ad . In the second step, the scheduler iterates over the set of remaining resources, and in each iteration it assigns one resource element to a selected node. Then, the maximum number of iterations in the second step is |C \ Cd |, and hence, the complexity order of Algorithm 1 is O(|C|), which is significantly lower than complexity of optimization problem (9), derived in (10). This complexity reduction comes at the cost of performance degradation in comparison with the non-relaxed problem in (9). The gap in performance is due to the fact that the proposed algorithm 1 works in a sequential manner, i.e. it selects the most energy-critical node, allocates the best resource chunk to it, and continues by allocating neighbor resource chunks to it3 until it becomes satisfied. One can see that while this sequential structure simplifies the solution, it degrades the performance by not selecting a bunch of neighbor chunks at once. In other words, a chunk on which a tagged node has the best SINR may lead to selection of a suboptimal set of neighbor chunks due to the contiguity constraint in SC-FDMA. However, regarding the fact that in recent releases of LTE like LTE-A clusters of chunks can be allocated to the users, i.e. G > 1 is available, the proposed scheduler will not suffer much from the contiguity constraint. On the other hand, the main drawback of this algorithm is seemed to be the need for channel state information (CSI), as the level of consumed energy in CSI exchange with the BS is comparable with, or even higher than, the consumed energy in actual data transmission. We discuss this problem in section V, and tackle it by presenting limited-feedback requiring variants of Algorithm 1. One must note that in algorithm 1 we have used a binary variable for prioritizing different M2M traffic types. In practice, regarding the diverse set of QoS requirements of different M2M applications, scheduler must be able to handle traffic streams with multi-level priorities, e.g. alarms, surveillance cameras, temperature sensors, and etc. Multilevel prioritized scheduling has been recently investigated in [12]. Then, design of a hybrid scheduler combining energy preserving features presented in Algorithm 1 with multi-level prioritized scheduling features presented in [12] makes an interesting research direction for our future research. V. L IFETIME -AWARE MTC S CHEDULING WITH L IMITED F EEDBACK In the previous section, we have presented a lifetime-aware scheduling solution which aims at maximizing the network lifetime. This scheme requires the CSI of machine nodes as well as other communications characteristics to be available at the BS. However optimal scheduling based on these information sets can improve the network lifetime, it requires machine nodes to send several status packets while the actual amount of useful data to be transmitted in many MTC applications is very limited [15]. Thus, in this section we present a low-complexity low-feedback frequency-domain scheduling solution to be used with other time-domain schedulers. This scheduler aims at maximizing the network lifetime while the remaining energy 3 due

to the contiguity constraint

level, reporting period, average static energy consumption, and average pathloss for each device are required at the BS. For M2M applications with limited device mobility, the average pathloss, reporting period, and static energy consumption are semi-constant during the lifetime of a device, and hence, the BS can save them for future use. Also, the energy consumption of each machine device is expected to be very low, then change in the remaining energy will happen in long time intervals and the BS needs to update it in long time intervals, from days to months. Potential applications of this scheduling solution will be presented in subsection V-A. Denote the subset of devices which are to be scheduled at time t as A. From (1), the expected lifetime of node i at time t, which is scheduled with |Ci | > 0 chunks, is formulated as follows: Li (t) =

Ei (t)Ti , Esi + Di [ξPi + Pc ]/R(|Ci |, Pi )

Algorithm 3: SIL-aware scheduling with limited feedback 1

2

- If At is empty, then 0 → Ctn ;

(19) 3

where

4

R(|Ci |, Pi ) = |Ci |M Sv (

Pi Gtr ), |Ci |M γi [N0 + I]w

maximize|Ci |,Pi subject to: C.21.1:

∑

i∈A

(21)

|Ci | ≤ |C|,

Di ≤ τ ∀i ∈ A, R(|Ci |, Pi ) C.21.3: Pi ≤ Pmax ∀i ∈ A, C.21.4: |Ci | ∈ N0 , C.21.2:

where N0 is the set of non-negative integers. Using C.21.2, the minimum resource requirement of node i, i.e. |Ci |min , is found by solving the following equation: |Ci |min M Sv (

Pmax Gtr ) = Di /τ. |Ci |min M γi [N0 + I]w

(22)

One sees that the optimization problem in (21) is not a convex optimization problem because of the Pi /|Ci | term in (20). To find an efficient solution for this problem, one can pursue a similar approach as in Algorithm 1. The overall solution procedure is presented in Algorithm 3. This algorithm first satisfies the minimum resource requirements of all nodes. Then, if the set of remaining chunks, i.e. Ctn , is non-empty, it finds node with the shortest individual lifetime that its lifetime can be improved by assigning more chunks and assigns it one more chunk. In this algorithm, F(Ei (t), Esi , Ti , |Ci |) =

Ei (t)Ti Esi +

Di [ξP(|Ci |)+Pc ] R(|Ci |,P(|Ci |))

P(y(i)) → p(i), ∀i ∈ A; return y and p;

(20)

and γi is the distance-dependent path-loss between node i and the BS. Then, the problem in (9) is rewritten as: Lsil net (t)

Initialization; - Derive |Ci |min , ∀i ∈ A, from (22); - |Ci |min → y(i), ∀i ∈ A; - F(Ei (t), Esi , Ti , y(i)) → f (i), ∀i ∈ A; - A → At ; while Ctn do - arg mini∈A f (i) → m; - y(m) + 1 → x; - if F(Em (t), Esm , Tm , x) > f (m) then - Ctn − 1 → Ctn , x → y(m); - F(Em (t), Esm , Tm , y(m)) → f (m); else - At \ m → At , and ∞ → f (m);

,

where the optimized transmit power for a given number of chunks, i.e. P(|Ci |), is found from (11). The outputs of this algorithm are y and p vectors, where the ith entries of them show the number of allocated chunks and the transmit power for the ith node, respectively.

A. Complexity Analysis and Potential Applications Algorithm 3 works over the set of all nodes in the first step, and over the set of remaining chunks in the later steps. Thus, its complexity order is O(|C|). Algorithm 3 can be used along with the specified time-domain schedulers in [9] in order to allocate frequency domain resources to energy-limited nodes and prolong the network lifetime. Another important application of this low complexity scheduler consists in uplink scheduling for time-controlled M2M communications. The 3GPP and IEEE have defined specific service requirements and features for M2M communications where one of the most important ones is the time-controlled feature [2, 35]. Based on this feature, the BS can assign a set of resources, which are repeated in time in regular time intervals, to a node based on its QoS requirements [36]. For M2M applications that support the time-controlled feature, the derived scheduling solutions using Algorithm 3 are valid for a long time interval because the communications characteristics of machine nodes, e.g. reporting periods, are semi-constant during their lifetimes. Then, for a group of machine devices with similar delay requirements, one can use the lifetime-aware scheduler in Algorithm 3, and assign them a set of persistent uplink transmission grants. The interested reader may refer to [36], where persistent resource provisioning for M2M communications has been introduced. VI. MTC S CHEDULING OVER LTE

NETWORKS

Here, we focus on the air interface of 3GPP LTE Release 13 [26]. In this standard, radio resources for uplink and downlink transmissions are distributed in both time and frequency domains. In the time domain, data transmissions are structured in frames where each frame consists of 10 subframes each with 1 ms length, while in the frequency domain, the available bandwidth is divided into a number of subcarriers each with 15 KHz bandwidth. The minimum allocatable resource element in a frame is a physical resource block pair (PRBP) which consists of 12 subcarriers spanning over one transmission time interval (TTI) [26]. Each TTI consists of two slots and includes

12 (or 14) OFDM symbols if long (or short) cyclic prefix is utilized. Based on the LTE open-loop power control [26], each node determines its uplink transmit power using downlink pathloss estimation as: P owC(|Ci |, δi ) = |Ci |P0 βi γi [2

ks TBS (|Ci |, δi ) |Ci |Ns Nsc

− 1].

(23)

In this expression, the number of assigned PRBPs to node i is denoted by |Ci |, the estimated downlink pathloss by γi , the compensation factor by βi , the number of symbols in a PRBP by Ns , and the number of subcarriers in a PRBP by Nsc . Also, ks is usually set to 1.25 and the transport block size (TBS) can be found in Table 7.1.7.2.1-1 of [26] as a function of |Ci | and TBS index. The TBS index, δi ∈ {0, · · · , 33}, is a function of modulation and coding scheme as in Table 8.6.1-1 of [26]. Based on the LTE specification in [26], P0 is set based on the required SNR level at the receiver as: P0 = βi [SNRtarget + Pn ] + [1 − βi ]Pmax , where Pn = −209.26 dB is the noise power in each resource block. Based on these specifications, one can rewrite the presented scheduling problems in sections IV and V in the context of LTE. Also, one sees that by tuning G in Algorithm 2, our proposed scheduling solutions can be used for both LTE and LTE-A networks which utilize SC-FDMA and clustered SCFDMA for uplink transmissions, respectively. Let us consider the scheduling problem in section V in the context of LTE. For node i, the expected battery lifetime is found from (1) as: Li (t) =

Ei (t)Ti . Esi + T T I[Pc + ξP owC(|Ci |, δi )]

(24)

Also, the resource allocation problem in (21) reduces to finding the optimal |Ci | and δi values, as follows:

δi∗ can be found by referring to the (|Ci |)-th column of the TBS table in [26], and finding the minimum TBS index for which, the constraint in (27) is satisfied. A. Low-Complexity Solution We can also use linear relaxation in order to transform the discrete optimization problem in (25) to a continuous optimization problem. Let us introduce an auxiliary variable Z and rewrite the optimization problems in (25) as follows: minimize|Ci | Z ∑ subject to: C.28.1:

i∈A

(28) |Ci | ≤ |C|,

C.28.2: |Ci |min ≤ |Ci |, ∀i ∈ A, C.28.3: Z ≤ Li (t), ∀i ∈ A, where Z = max Li1(t) . From the TBS table in [26], one sees that the maximum TBS for one PRBP is 968, then ¯ i /968}, |Ci |min = max{|Ci |m , D in which |Ci |m is found by satisfying C.25.2 and C.25.3 with equality, as follows: ¯ ks D i

|Ci |m P0 βi γi [2 |Ci |m Ns Nsc − 1] = Pmax . The scheduling problem in (28) is a convex optimization problem because the objective function is concave and the constraints make a convex set. Then, one can use the dual Lagrangian scheme and find the desired solution as: ¯ i /[Ns Nsc ] } { ks ln(2)D |Ci |∗ = max |Ci |min , ( ) , (29) i (t)Ti µ 1 + L eP0Eβγ − 1e i λi ξT T I

where µ and λi :s are Lagrange multipliers. The derived |Ci |∗ values from (29) are fractional solutions to the relaxed problem maximize|Ci |,δi (25) ∑ in (28). Then, we can use randomized rounding to find the s.t.: C.25.1: |Ci | ≤ |C|, number of assigned PRBPs to each node [37]. Given the i∈A ¯ assigned number of PRBPs to node i as |Ci |∗ and the queued C.25.2: Di ≤ TBS(|Ci |, δi ), ∀i ∈ A, data length as Di , the optimal TBS index is found from (27) C.25.3: P owC(|Ci |, δi ) ≤ Pmax , ∀i ∈ A, as δi∗ = FunD(|Ci |∗ , Di ). Then, the corresponding transmit C.25.4: δi ∈ {0, · · · , 33}; |Ci | ∈ {1, · · · , |C|}, ∀i ∈ A, power for node i is computed by inserting the derived |Ci |∗ , ¯ i = Di + δi∗ , and TBS(|Ci |∗ , δi∗ ) in (23). where |C| is the total number of available PRBPs, D Doh , and Doh is the size of overhead information for User VII. P ERFORMANCE E VALUATION Datagram Protocol (UDP), Internet Protocol (IP), and etc. In order to solve this problem, we can use a modified version of In this section, we apply our proposed scheduling algorithms Algorithm 3. The solution procedure is presented in Algorithm to a 3GPP LTE-A system and provide simulation results to 4. In this algorithm, F(x, y) = Li (t) |Ci |=x,δi =y , and |Ci |min demonstrate lifetime improvements. The testbed for simuis the minimum PRBP requirement for node i found as: lations is based on the uplink of a single cell multi-user 3GPP LTE-A network with 1.4 MHz bandwidth [26]. The min |Ci | = minimizeδi |Ci |, (26) deployment of machine devices and their traffic model follow ¯ subject to: TBS(|Ci |, δi ) ≥ Di ; P owC(|Ci |, δi ) ≤ Pmax . the proposed models in [27, annex A] for smart metering Also, given the assigned number of PRBPs to node i, i.e. |Ci |, applications, and are reflected in Table I. Upon having data to and the queued data length as Di , Algorithm 4 calls function transmit, machine nodes send scheduling request on PUCCH FunD(|Ci |, Di ) in order to derive the corresponding TBS index to the BS. As per [9, 36], we consider that part of PUSCH radio resources are assigned to M2M communications. Here, δi∗ that minimizes the transmit power as: we assume the first two radio frames in each second, i.e. 20 ∆ δi∗ =FunD(|Ci |, Di ) = minimize δi , (27) subframes each containing 6 PRBPs, have been reserved for uplink transmissions of machine devices. The BS schedules subject to: TBS(|Ci |, δi ) ≥ Di + Doh . machine devices and sends back the scheduling grants on Lsil net

2

- If At is empty, then 0 → Ctn ; 3

return y and p; TABLE I: Simulation parameters

Parameter

Value

Cell radius Path loss model PSD of noise System bandwidth Transmission time interval, TTI Number of PRBPs in each TTI ks , Ns , Nsc TBS index, δi Transport block size Pathloss compensation factor, βi Number of nodes Data generation at each node Duty cycle, Ti ¯i Payload+overhead size, D Circuit power, Pc SNRtarget Maximum transmit power, Pmax Static energy consumption, Esi

500 m r 128 + 38 log10 ( 1000 ) -174 dBm/Hz 1.4 MHz 1 ms 6 1.25, 12, 12 {0, · · · , 26} Tab. 7.1.7.2.1-1 [26] 0.92 18000 Poisson, rate 1/300 300 sec, ∀i ∈ A 600 Bits 7 dBm 1 dB 24 dBm 10 µJ

PDCCH to let them transmit their packets in the upcoming reserved resources. Six different MTC scheduling schemes that have been implemented in our simulations are as follows: • •

Scheme 1: This scheme is based on Algorithm 1, and aims at maximizing the SIL network lifetime. Scheme 2: This scheme is based on Algorithm 3 and 4, and provides a low-complexity solution with limited feedback requirement. In this scheme, a round robin (RR) scheduler is used for time-domain scheduling, and Algorithm 4 is used for frequency-domain scheduling.

PDF PDF PDF

Initialization; - Derive |Ci |min , ∀i ∈ A, from (26); - |Ci |min → y(i), ∀i ∈ A; - FunD(y(i), Di ) → δi∗ , ∀i ∈ A; - P owC(y(i), δi∗ ) → p(i), ∀i ∈ A; - F(y(i), δi∗ ) → f (i), ∀i ∈ A; - A → At ; while Ctn do - arg mini∈A f (i) → m; - y(m) + 1 → x; ∗ - FunD(x, Dm ) → δm ; ∗ - P owC(x, δm ) → P ; ∗ - if P ≤ Pmax , and F(x, δm ) > f (m) then - Ctn − 1 → Ctn , x → y(m), P → p(m); ∗ - F(x, δm ) → f (m); else - At \ m → At , and ∞ → f (m);

PDF PDF

1

PDF

Algorithm 4: Scheduling with limited feedback for LTE

0.11 0 0.11

scheme 1

scheme 2

0 0.11 0 0.11

scheme 3 scheme 4

0 0.11 0 0.11

scheme 5 scheme 6

0 103

104

105

106

107

108

Time (× Ti)

Fig. 1: Empirical PDF of individual lifetimes using different scheduling schemes

Scheme 3: This scheme is based on Algorithm 1, and aims at maximizing the LIL network lifetime. • Scheme 4: This scheme consists of two RR schedulers for time- and frequency-domain scheduling, and represents the delay/priority-aware scheduling schemes in literature [11, 16] when the MTC traffic has no strict delay/priority requirement. • Scheme 5: This scheme consists of a channel-aware scheduler for time-domain scheduling, a RR scheduler for frequency domain scheduling, and represents the proposed channel-aware scheduling schemes in [10, 11]. • Scheme 6: This scheme represents the proposed energy efficient MTC scheduling algorithm in [14], where the ratio between the sum data rates and the transmit power consumptions of all devices is maximized. One must note that a fair comparison requires us to compare scheme 1, as a lifetime-aware time/frequency-domain scheduling solution against schemes {3, 4, 6} which either benefit from RR scheduling or lifetime-aware scheduling with full CSI. On the other hand, scheme 2 which benefits from the low-complexity lifetime-aware solution with limited-CSI, can be compared against scheme 4 and 5, which benefit from RR scheduling and channel-aware scheduling with limited CSI, respectively. •

A. Performance Evaluation of the Proposed Schedulers Fig. 1 represents the probability density function (PDF) of battery lifetimes of machine nodes using different scheduling schemes. The x-axis has been depicted in log-scale to highlight the differences in PDFs when the initial energy drains happen. One sees that the first-energy-drain using scheme 1, which aims at maximizing the SIL network lifetime, happens much later than the first energy drain using the benchmarks, i.e. scheme 4, 5, and 6. Also, one sees that the last energy drain using scheme 3, which aims at maximizing the LIL network lifetime, happens much later than the benchmarks. Furthermore, we see that the PDF of scheme 1 has a compact shape, which shows that the individual lifetimes of machine devices are distributed in a limited time interval. The detailed

13

6

3

LIL network lifetime SIL network lifetime

5

2.5

4

2

3

1.5

2

1

1

0.5

0

0 Sch. 1

Sch. 2

Sch. 3

Sch. 4

Sch. 5

Sch. 6

Scheduling Schemes

x 10 10

1

Jain index of ind. lifetimes

3.5

Jain index Variance

0.8

8

0.6

6

0.4

4

0.2

2

0

Sch. 1

Sch. 2

Sch. 3

Sch. 4

Sch. 5

Sch. 6

Variance of ind. lifetimes

104

107

SIL network lifetime ( Ti)

LIL network lifetime ( Ti)

7

0

Scheduling schemes

Fig. 2: Network lifetime comparison SIL network lifetime comparison of the proposed scheduling schemes is presented in Fig. 2. In this figure, it is evident that the achieved SIL network lifetime from scheme 1 is 2.4 times higher than scheme 4, three times higher than scheme 5, and 13.4 times higher than the scheme 6. Also, one sees that scheme 2, which aims at maximizing the SIL network lifetime with limited feedback requirement, outperforms the baseline schemes 4 and 5. The detailed LIL network lifetime comparison is presented in Fig. 2. In this figure, it is evident that the achieved LIL network lifetime from scheme 3 is 1.55 times higher than scheme 4, 1.15 times higher than scheme 5, and 0.02 times higher than scheme 6. Fairness of the proposed scheduling schemes is investigated in Fig. 3. The right axis of Fig. 3 shows the variance of individual lifetimes, while the right axis represents the modified Jain’s fairness index [38], calculated as: ∑ ( i∈A Li )2 ∑ J = . |A| i∈A L2i One sees that scheme 1 achieves the highest fairness index. Recall from Fig. 2, where it was shown that SIL-aware scheduling prolongs the shortest individual lifetime in the network. Comparing Fig. 2 and Fig. 3 indicates that using SIL-aware scheduling, machine nodes will last all together for a long period of time, and will die almost at the same time. Fig. 4 indicates the impact of link budget on the network battery lifetime. Recall the transmit power expression in (23). From this expression, one sees that transmit power is an increasing function of SNRtarget . In Fig. 4, one sees that the battery lifetime significantly decreases in SNRtarget . Also, one sees that the achieved battery lifetime from scheme 2 is approximately 2 times higher than the baseline scheme for different SNRtarget values. Similar results can be seen in Fig. 5 ¯ i on the network battery lifetime. One sees for the impact of D when the packet size increases the network lifetime decreases. Also, one sees that the achieved network lifetime from scheme 2 is approximately 2 times higher than the baseline scheme ¯ i values. for different D B. Comparison of Network Lifetime Definitions In Fig. 3, one sees that by SIL-aware scheduling, all nodes are expected to have their batteries drained approximately at the same time. This is due to the fact that SIL scheduler tries to prolong battery lifetimes of low-battery nodes at the cost

SIL network lifetime (× Ti )

Fig. 3: Fairness of the proposed schemes

4

×10 4

3

Lifetime-aware sched. (scheme 2) RR scheduling (scheme 4)

2 1 0 -2

1

4

7

SNRtaget (dB)

Fig. 4: Impact of the link budget on lifetime

of sacrificing high-battery nodes’ lifetimes. On the other hand, by LIL-aware scheduling, battery lifetimes of all other nodes are sacrificed to prolong battery lifetime of the node with the longest battery lifetime. Then, it is clear that depending on the M2M application, the choice of network lifetime to be used for scheduling and consequently the scheduler design, can be different from one network to the other. For example, for a network in which loosing even a small number of nodes deteriorates the performance like sensors installed in urban trash bins, and hence batteries must be replaced when drained, SIL-aware scheduling may significantly reduce the network maintenance costs by reducing the efforts to monitor the network continuously and replace battery-drained machine nodes one by one. On the other hand, for networks in which correlation between gathered data from different nodes is high like sensors installed in an area for temperature monitoring, LIL-aware scheduling may minimize the maintenance costs by prolonging battery lifetimes of a subset of nodes. C. Lifetime-Aware Resource Provisioning for MTC As discussed above, provisioning uplink resources for MTC can impact the network lifetime. If the amount of reserved resources is larger than required, some resources are wasted, spectral efficiency of the network decreases, and the QoS for other services, e.g. web surfing, may be decreased. If the amount of reserved resources is smaller than required, machine nodes must wait for a longer period of time to get access to the reserved resources, and send data with a higher transmit

SIL network lifetime (×Ti )

8

energy efficiency, i.e. it increases in sending more data in each resource block. Fig. 7 compares the SIL network lifetime. One sees that the network lifetime follows a similar trend to the energy efficiency, i.e. it decreases as the amount of data to be transmitted over a resource block increases.

×10 4 Lifetime-aware sched. (scheme 2) RR scheduling (scheme 4)

6 4 2 0 150

VIII. C ONCLUSIONS 300

450

600

700

Packet size (bits)

¯ i on lifetime Fig. 5: Impact of D ×10 -7

3 EE of scheme 1 with 1 reserved frame EE of scheme 1 with 2 reserved frames SE of scheme 1 with 1 reserved frame SE of scheme 1 with 2 reserved frames

4 3.5

2.5 2

3

1.5

2.5

1

2

0.5

1.5 150

225

300

375

Spectral efficiency (bit/sec/Hz)

Energy efficiency (bit/joule)

4.5

0 450

Packet size (bits)

SIL network lifetime (× Ti)

Fig. 6: Energy/spectral efficiency tradeoff in MTC resource provisioning 8

In this paper, a lifetime-aware resource allocation framework for cellular-based M2M networks is introduced. Theoretical analyses on the impact of scheduling and power control on the energy consumptions of machine nodes, and hence, network lifetime are presented. Based on these analyses, battery lifetime aware scheduling algorithms are derived. The obtained results show that the optimal scheduling decision depends on the priority class of traffic, remaining battery lifetime of the devices, and the transmission-dependent and -independent sources of energy consumptions. Comparing the lifetimeaware schedulers with the existing scheduling solutions in literature shows that modeling the energy consumption of MTC, and designing respective scheduling schemes can significantly prolong the network lifetime. Furthermore, the energy efficiency, spectral efficiency, and network lifetime tradeoffs in uplink MTC scheduling are investigated. It is also shown that uplink scheduling based on the max-min fairness enables machine nodes to last for a long time and die approximately at the same time, which contributes significantly in network’s maintenance costs reduction. The results of this article can be used to analyze and optimize the lifetime performance of deployed machine-type devices over cellular networks.

×10 4 scheme 1 with 1 reserved frame scheme 1 with 2 reserved frames

6

A PPENDIX A

4 2 0 150

225

300

375

450

Packet size (bits)

Fig. 7: Lifetime tradeoffs in MTC resource provisioning

power which in turn reduces their battery lifetimes. Then, one sees that there is a tradeoff between energy efficiency and spectral efficiency in uplink transmission. This tradeoff is presented in Fig. 6. In this figure, the solid curves illustrate the energy efficiency of uplink transmissions in Bit-per-Joule for two resource provisioning approaches: (i) when 1 radio frame consisting of 10 subframes is allocated to MTC in each second; and (ii) when 2 radio frames are allocated to MTC in each second. One sees that the energy efficiency decreases when either the amount of reserved resources decreases or the amount of data to be transmitted over a given set of resources increases. The dashed curves illustrate the spectral efficiency of uplink transmissions in Bit/Sec/Hz. One sees that spectral efficiency presents a reverse trend when compared with the

To solve the optimization problem in (4), we first relax C.3.2, solve the relaxed problem, and then apply C.3.2. The Lagrangian function for the relaxed problem is written as follows: ∑ ∑ 1 − Z], (30) F = Z + µ[ τi − τ ] + λi [ i∈A i∈A Li (t) where µ and λi :s are Lagrange multipliers. Using convex optimization theory [24], the solution for relaxed problem, i.e. τi∗ , is found by solving: ∂L−1 (t) → µ + λi i = 0, ∂τi [ λi w Di Pc + ξ[N0 + I] S( ), →µ+ Ei (t)Ti hi Gtr τ ]i Di ˙ Di w − ξ S( )[N0 + I] = 0. τi τi hi Gtr

∂F = 0, ∂τi

(31)

Also, µ and λi :s, i.e. the Lagrange multipliers, are found due to the following Karush Kuhn Tucker (KKT) conditions [24]: (∑ ) µ ≥ 0; τi − τ µ = 0; (32) i∈A ( ) 1/Li (t) − Z λi = 0; λi ≥ 0; ∀i ∈ A. (33)

For example, in the special case that S(x) is found from (2) and Γmcs = 1, the real-valued solution of (31) is found as: τi∗ =

1 + L( 1e

ln(2)Di /w [ [hi Gtr ][Pc +Ti Ei (t)µ/λi ] ξ(N0 +I)w

] , −1 )

(34)

where e is the Euler’s number, and L(x) is the LambertW function, i.e. inverse of the function f (x) = x exp(x) [25]. Now, by applying C.3.2 the optimal transmission time is found as the maximum of τim and τi∗ , where τi∗ has been introduced in (3). R EFERENCES [1] Ericsson, Huawei, NSN, and et al., “A choice of future M2M access technologies for mobile network operators,” Tech. Rep., Mar. 2014. [2] 3GPP TS 22.368 V13.1.0, “Service requirements for machinetype communications,” Tech. Rep., 2014. [Online]. Available: http://www.3gpp.org [3] Nokia Networks, “Looking ahead to 5G: Building a virtual zero latency gigabit experience,” Tech. Rep., 2014. [4] A. Laya, L. Alonso, and J. Alonso-Zarate, “Is the random access channel of LTE and LTE-A suitable for M2M communications? A survey of alternatives,” IEEE Commun. Surveys Tuts., vol. 16, no. 1, pp. 4–16, Dec. 2013. [5] G. Miao, A. Azari, and T. Hwang, “E 2 -MAC: Energy efficient medium access for massive M2M communications,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4720 – 4735, Nov. 2016. [6] Z. Alavikia and A. Ghasemi, “A multiple power level random access method for M2M communications in LTE-A network,” Transactions on Emerging Telecommunications Technologies, vol. 28, no. 6, June 2017. [7] Z. Zhou, J. Feng, Y. Jia, S. Mumtaz, K. M. S. Huq, J. Rodriguez, and D. Zhang, “Energy-efficient game-theoretical random access for M2M communications in overlapped cellular networks,” Computer Networks, July 2017. [8] Nokia Networks, “LTE-M – optimizing LTE for the Internet of things,” Tech. Rep., 2015. [9] M. Mehaseb, Y. Gadallah, A. Elhamy, and H. El-Hennawy, “Classification of LTE uplink scheduling techniques: An M2M perspective,” IEEE Commun. Surveys Tuts., no. 99, Nov. 2015. [10] S. Zhenqi, Y. Haifeng, C. Xuefen, and L. Hongxia, “Research on uplink scheduling algorithm of massive M2M and H2H services in LTE,” in IET International Conference on Information and Communications Technologies, Apr. 2013, pp. 365–369. [11] A. S. Lioumpas and A. Alexiou, “Uplink scheduling for machineto-machine communications in LTE-based cellular systems,” in IEEE GLOBECOM Workshops, 2011, pp. 353–357. [12] A. E. Mostafa and Y. Gadallah, “A statistical priority-based scheduling metric for M2M Communications in LTE Networks,” IEEE Access, vol. 5, pp. 8106–8117, May 2017. [13] S. A. Mahmud et al., “Fairness evaluation of scheduling algorithms for dense M2M implementations,” in IEEE Wireless Communications and Networking Conference Workshops, Apr. 2014, pp. 134–139. [14] A. Aijaz et al., “Energy-efficient uplink resource allocation in LTE networks with M2M/H2H co-existence under statistical QoS guarantees,” IEEE Trans. Commun., vol. 62, no. 7, pp. 2353–2365, July 2014. [15] K. Wang, J. Alonso-Zarate, and M. Dohler, “Energy-efficiency of LTE for small data machine-to-machine communications,” in IEEE International Conference on Communications (ICC), June 2013, pp. 4120–4124. [16] M. Kalil et al., “Low-complexity power-efficient schedulers for LTE uplink with delay-sensitive traffic,” IEEE Trans. Veh. Technol., vol. 64, no. 10, pp. 4551–4564, Nov. 2015. [17] H. S. Dhillon et al., “Power-efficient system design for cellular-based machine-to-machine communications,” IEEE Trans. Wireless Commun., vol. 12, no. 11, pp. 5740–5753, Nov. 2013. [18] G. Miao, N. Himayat, G. Li, and S. Talwar, “Low-complexity energyefficient scheduling for uplink OFDMA,” IEEE Trans. Commun., vol. 60, no. 1, pp. 112–120, Jan. 2012. [19] Y. Liu et al., “Design of a scalable hybrid MAC protocol for heterogeneous M2M networks,” IEEE Internet Things J., vol. 1, no. 1, pp. 99–111, Feb. 2014. [20] A. Azari, “Energy-efficient scheduling and grouping for machine-type communications over cellular networks,” Ad Hoc Networks, vol. 43, pp. 16–29, june 2016.

[21] A. Azari and G. Miao, “Lifetime-aware scheduling and power control for cellular-based M2M communications,” in IEEE Wireless Communications and Networking Conference (WCNC), 2015. [22] ——, “Lifetime-aware scheduling and power control for M2M communications over LTE networks,” in IEEE Vehicular Technology Conference (VTC), 2015. [23] 3GPP, “USF capacity evaluation for MTC,” Tech. Rep., 2010, TSG GERAN 46 GP-100894. [24] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. [25] R. Corless, G. Gonnet, D. Hare, D. Jeffrey, and D. Knuth, “On the LambertW function,” vol. 5, no. 1, pp. 329–359, 1996. [26] 3GPP TS 36.213, “Evolved universal terrestrial radio access, physical layer procedures,” Tech. Rep., May 2016, (Rel. 13). [27] 3GPP, “Study on provision of low-cost machine-type communications (MTC) user equipments (UEs),” 3GPP TR 36.888 V12.0.0, 2013. [28] I. C. Wong, O. Oteri, and W. McCoy, “Optimal resource allocation in uplink SC-FDMA systems,” IEEE Trans. Wireless Commun., vol. 8, no. 5, pp. 2161–2165, May 2009. [29] M. R. Palattella, M. Dohler, A. Grieco, G. Rizzo, J. Torsner, T. Engel, and L. Ladid, “Internet of things in the 5G era: Enablers, architecture, and business models,” IEEE J. Sel. Areas Commun, vol. 34, no. 3, pp. 510–527, Mar. 2016. [30] L. Hanzo, H. Haas, S. Imre, D. O’Brien, M. Rupp, and L. Gyongyosi, “Wireless myths, realities, and futures: From 3G/4G to optical and quantum wireless,” Proceedings of the IEEE, vol. 100, no. Special Centennial Issue, pp. 1853–1888, May 2012. [31] S. Imre and L. Gyongyosi, Advanced quantum communications: an engineering approach. John Wiley & Sons, 2012. [32] G. Miao, N. Himayat, and G. Y. Li, “Energy-efficient link adaptation in frequency-selective channels,” IEEE Trans. Commun., vol. 58, no. 2, pp. 545–554, Feb. 2010. [33] H. Luss, Equitable Resource Allocation: Models, algorithms and applications. John Wiley & Sons, 2012, vol. 101. [34] G. Miao, J. Zander, K. W. Sung, and S. B. Slimane, Fundamentals of Mobile Data Networks. Cambridge University Press, 2016. [35] H. Cho, “Machine to machine (M2M) communications technical report,” 2011, IEEE 802.16 Broadband Wireless Access Working Group. [36] S. Y. Lien and K. C. Chen, “Massive access management for QoS guarantees in 3GPP machine-to-machine communications,” IEEE Commun. Lett., vol. 15, no. 3, pp. 311–313, Mar. 2011. [37] P. Raghavan and C. D. Tompson, “Randomized rounding: A technique for provably good algorithms and algorithmic proofs,” Combinatorica, vol. 7, no. 4, pp. 365–374, 1987. [38] R. Jain, D. M. Chiu, and W. Hawe, “A quantitative measure of fairness and discrimination for resource allocation in shared computer systems,” Digital Equipment Corporation Hudson, Sept. 1984.

Paper B: Latency-Energy Tradeoff based on Channel Scheduling and Repetitions in NB-IoT Systems Amin Azari, Guowang Miao, Cedomir Stefanovic, and Petar Popovski Accepted in IEEE Global Communication Conference (Globecom), 2018. (©2018 IEEE. Reprint with permission.)

117

Latency-Energy Tradeoff Based on Channel Scheduling and Repetitions in NB-IoT Systems ˇ Amin Azari∗ , Guowang Miao∗ , Cedomir Stefanović+ , and Petar Popovski+ ∗ KTH Royal Institute of Technology, + Aalborg University Email: {aazari,guowang}@kth.se, {cs,petarp}@es.aau.dk Abstract—Narrowband Internet of Things (NB-IoT) is the latest IoT connectivity solution presented by the 3rd generation partnership project (3GPP). NB-IoT introduces coverage classes and offers a significant link budget improvement by allowing repeated transmissions by nodes that experience high path loss. However, those repetitions necessarily increase the energy consumption and the latency in the whole NB-IoT system. The extent to which the whole system is affected depends on the scheduling of the uplink and downlink channels. We address this question, not treated previously, by developing a tractable model of NB-IoT access protocol operation, comprising message exchanges in random-access, control, and data channels, both in the uplink and downlink. The model is then used to analyze the impact of channel scheduling as well as the interaction of coexisting coverage classes, through derivation of the expected latency and battery lifetime for each coverage class. These results are subsequently employed in investigation of latency-energy tradeoff in NB-IoT channel scheduling as well as determining the optimized operation points. Simulations results show validity of the analysis and confirm that channel scheduling and coexistence of coverage classes significantly affect latency and battery lifetime performance of NB-IoT devices.

I. I NTRODUCTION Internet of Things (IoT) is behind 2 out of 3 major drivers of next generation wireless networks, which are massive machine type communications (mMTC), ultra reliable low latency communications (uRLLC), and enhanced mobile broadband (eMBB) [1]. Due to the fundamental differences in characteristics and service requirements between IoT and legacy traffic in cellular networks, which are seen in massive number of connected devices, short packet sizes, and long battery lifetimes, revolutionary connectivity solutions have been proposed and implemented by industry [2, 3]. The most prominent examples of such solutions are SigFox, introduced in 2009, and LoRa, introduced in 2015, both implemented in the unlicensed band, i.e., 868 MHz in Europe [3, 4]. On the other hand, the accommodation of IoT traffic over cellular networks has been investigated by the 3rd generation partnership project (3GPP), proposing evolutionary solutions like LTE Cat-1 and LTE Cat-M [5, 6]. Recently, these efforts have been also complemented by introduction of revolutionary solutions like NB-IoT [7]. NB-IoT represents a big step towards realization of massive IoT connectivity over cellular networks. Communication in NB-IoT systems takes place in a narrow, 200 KHz bandwidth, resulting in more than 20 dB link budget improvement over the legacy LTE [9]. Furthermore, NB-IoT introduces a set

Downlink frame (10 subframes) 1

2

NPBCH

3

4

5

6

NPSS

NPDCCH or NPDSCH

7

8

9

NPDCCH or NPDSCH

10 NSSS/NPDCCH or NPDSCH

Uplink frame (10 subframes) 1

2

3

4

5

6

7

8

9

10

NPUCSH or NPRACH

Fig. 1: NB-IoT features frequency-division duplex for uplink and downlink [8]. Downlink/uplink NP channels and signals are time multiplexed, as depicted in the figure.

of coverage classes, each associated with a number of signal repetitions, which are assigned to users based on their experiencing pathloss in communications with the base station (BS). The narrow communication bandwidth and signal repetitions enable reliable communication of smart devices deployed in remote or isolated areas, e.g., basements, with the BS. As the legacy signaling and communication protocols were designed for large bandwidths, NB-IoT introduces a solution with five new narrowband physical (NP) channels [8, 10], see Fig. 1: random access channel (NPRACH), uplink shared channel (NPUSCH), downlink shared channel (NPDSCH), downlink control channel (NPDCCH), and broadcast channel (NPBCH). NB-IoT also introduces four new physical signals: demodulation reference signal (DMRS) that is sent with user data on NPUSCH, narrowband reference signal (NRS), narrowband primary synchronization signal (NPSS), and narrowband secondary synchronization signal (NSSS). Prior works on NBIoT investigated preamble design for access reservation of devices over NPRACH [11, 12], uplink resource allocation to the connected devices [13], coverage and capacity analysis of NB-IoT systems in rural areas [14], coverage of NB-IoT with consideration of external interference due to deployment in guard band [15], and impact of channel coherence time on coverage of NB-IoT systems in [16]. Further, in [17], energy consumption of IoT devices in data transmission over NB-IoT systems in normal, robust, and extreme coverage scenarios has been investigated. The results obtained in [17] illustrate that NB-IoT significantly reduces the energy consumption compared to the legacy LTE, due to the existence of the deep sleep mode for the devices that are registered to the BS. In this paper, we study an important and so far untreated problem: when and how much resources to allocate to NPRACH, NPUSCH, NPDCCH, and NPDSCH in coexistence scenarios, where BS is serving NB-IoT devices with random activations that belong to different coverage classes. The

In this paper, we extend the state-of-the-art latency/energy models in [17, 20, 23, 24], incorporate the NB-IoT channel multiplexing problem, and consider coexistence of devices from a diverse set of coverage classes in the same cell. Specifically, the main contributions of this work are: •

•

• •

Derivation of a tractable analytical model of channel scheduling problem in NB-IoT systems that considers message exchanges on both downlink/uplink channels, from synchronization to service completion. Derivation of closed-form analytical expressions for service latency and energy consumption, and derivation of the expected battery lifetime model for devices connected to the network. Investigation of a latency-energy tradeoff in channel scheduling for NB-IoT systems. Investigation of the interaction among the coverage classes coexisting in the system: performance loss in one coverage class due to an increase in number of connected devices from another coverage class.

The remainder of the paper is structured as follows. The next section is devoted to the system model. Section III presents the analysis. Investigation of the operational tradeoffs and performance evaluation are presented in Section IV. Concluding remarks are given in Section V.

RA

Uplink

Data RAR

Power

Downlink

Transmit data PC

NRS, NPSS, NSSS, master information block (MIB)

Time $%$

%$!"#

&'

Deep sleep

a) Device demanding uplink service Uplink

RA RAR

Downlink Power

solution to this problem has a significant impact on the service execution and devices’ performance, as the resource allocation to different channels faces inherent tradeoffs. The essence of the tradeoff can be explained as follows. If random access opportunities (NPRACH) occur frequently, less uplink radio resources remain for uplink data channel (NPUSCH), which increases the latency in data transmissions. On the other hand, if NPRACH is scheduled infrequently, latency and energy consumption in access reservation increase due to the extended idle-listening time and increased collision probability. Further, as device scheduling for uplink/downlink channels is performed over NPDCCH, infrequent scheduling of this channel may lead to wasted uplink resources in NPUSCH and increased latency in data transmissions. Conversely, if NPDCCH occurs frequently, the latency and energy consumption of transmissions over NPUSCH will increase. While the channel scheduling itself is a complicated problem, introduction of coexisting coverage classes, and adapting channel scheduling to their diverse quality of service requirements pose further challenges to the problem. The literature on latency and energy analysis and optimization for LTE networks is mature, e.g. refer to [20, 23, 24]. However, while the NB-IoT has been significantly inspired by the LTE, the revolutionary features of NB-IoT, including coexistence of coverage classes and timemultiplexing of physical channels, mandate a detailed analysis of user’s experienced latency and consumed energy in NB-IoT connectivity. This analysis could be subsequently employed in optimized operation control of NB-IoT networks.

Data $'

Time Synch. PC

idle PC

RACH PC

Waiting for RAR PC

b) Device demanding downlink service

NPBCH

NPSS

NSSS

Receive data PC

Optional check for data from BS

NPRACH NPDCCH NPUSCH NPDSCH

Fig. 2: Communications exchanges and power consumption in NB-IoT access networking. Note: Reference signals, including NRS, NPSS, NSSS, and master information block (MIB), are broadcasted regularly; here we show only a single realization.

II. S YSTEM M ODEL A. NB-IoT Access Networking Assume a NB-IoT cell with a BS located in its center, and N devices uniformly distributed in it. In general, there are C coverage classes defined in an NB-IoT cell, where the BS assigns a device to a class based on the estimated path loss between them and informs the device of its assignment. Class j, ∀j, is characterized by the number of replicas cj that must be transmitted per original data/control packet. For example, based on the specifications in [8], each device belonging to group j shall repeat the preamble transmitted over NPRACH cj ∈ {1, 2, 4, 8, 16, 32, 64, 128} times. Further, let us denote by fj the fraction of devices belonging to class j, by S the number of communication sessions that a typical IoT device performs daily and by p the probability that a device requests uplink service. The arrival rates of uplink/downlink service requests, per second (sec), to the system are, respectively: NSp N S (1 − p) sec−1 , Gd = sec−1 . (1) 24 · 3600 24 · 3600 Initially, when a NB-IoT device requires an uplink/downlink service, it first listens for the cell information, i.e., NPSS and NSSS, through which it synchronizes with the BS. Then, the device performs access reservation, by sending a random access (RA) request to the BS over NPRACH. The BS answers to a successfully received RA by sending the random access response (RAR) message over NPDCCH, indicating the resources reserved for serving the device. Finally, the device sends/receives data to/from the BS over NPUSCH/NPDSCH channels, which, depending on the application, may be followed by an acknowledgment (ACK) [8]. In contrast to LTE, a device that is connected to the BS can go to the deep sleep state [7, Section 7.3], which does not exist in LTE and from which the device Gu =

#$% &/ ()*+", -% .

NPUSCH

NPRACH

B. Problem Formulation Based on the model presented in Fig. 2, the expected latencies in uplink/downlink communication in class j are, respectively: Duj =Dsyj +Drrj +Dtxj Ddj =Dsyj +Drrj +Drxj

(2)

where Dsyj , Drrj , Dtxj , Drxj are the expected time spent in synchronization, resource reservation, data transmission in uplink service, and data reception in downlink service, respectively. Similarly, the models of expected energy consumption of an uplink/downlink communication in class j are: Euj = Esyj + Errj + Etxj + Es Edj = Esyj + Errj + Erxj + Es

(3)

where Esyj , Errj , Etxj , Erxj , and Es are the expected device energy consumption in synchronization, resource reservation, data transmission in uplink service, data reception in downlink service, and optional communications like acknowledgment, respectively. Since the energy consumption of a typical IoT device involved reporting application can be modeled as a semi-regenerative Poisson process with regeneration point at the end of each reporting period [25], one may define the expected battery lifetime as the ratio between stored energy and energy consumption per reporting period. In this case, the expected battery lifetime can be derived as: Lj =

E0 [day] SpEuj + S(1 − p)Edj

(4)

where E0 is the energy storage at the device battery. In order to derive closed-form latency and energy consumption expressions, e.g., model Errj and Drrj , in the sequel we analytically investigate the performance impacts of channel scheduling, arrival traffic, and coexisting coverage classes on the performance indicators of interest. III. A NALYSIS As mentioned in Section II, in NB-IoT systems the control, data, random access, and broadcast channels are multiplexed on the same set of radio resources. Thus, their mutual impact in both uplink and downlink directions are significant, which 1 For the sake of completeness, we also mention another novel reconnection scheme designed for NB-IoT, in which a device can request to resume its previous connection after receiving the random access response (RAR) [10, Section III]. Towards this end, it needs to respond to the RAR message by transmission of its previous connection ID as well as the cause for resuming the connection.

!"

BS initiated control signals

NPDCCH

#$% &' ()*+", -% .

NPDSCH

can become reconnected just by transmitting a RA request accompanied by a random number [7, Fig. 7.3.4.5-1]. This new functionality aims to address the inefficient handling of IoT communications by LTE , as it significantly saves energy due to the fact that IoT devices do not need to restart all steps of the connection establishment procedure [17, 22]. Fig. 2 represents the access protocol exchanges for NB-IoT, as described in [7, Section 7.3].1

Cell refere-nce signals

IoT devices

Fig. 3: Queuing model of the NB-IoT access networking. The yellow and gray circles represent servers for downlink and uplink channels, respectively.

is not the case in legacy LTE due to wide set of available radio resources. In the following, we propose a queuing model of NB-IoT access networking, which captures these interactions. A. Queuing Model of NB-IoT Access Protocol Fig. 3 depicts the queuing model of NB-IoT access networking, comprising operation of NP random access, control, and data channels. The gray circle represents the uplink server serving two channel queues, NPRACH and NPUSCH, while the yellow circle represents the downlink channel serving three channel queues, NPDCCH, NPDSCH, as well as the reference signals, such as NPSS. Let tj be the average time interval between two consecutive scheduling of NPRACH of class j and Mj the number of orthogonal random access preambles available in it. The duration of scheduled NPRACH of class j is cj τ , where τ is the unit length, equal to the NPRACH period for the coverage class with cj = 1. The interarrival times between two NPRACH periods in NB-IoT can vary from 40 ms to 2.56 s [8]. Further, b denotes the fraction of time in which reference signals are scheduled in a downlink radio frame, e.g., NPBCH, NPSS, and NSSS. Five subframes in every two consecutive downlink frames are allocated to reference signals [8], implying b = 0.2. Finally, a semi-regular scheduling of NPDCCH has been proposed by 3GPP in order to prevent waste of resources in the uplink channel when BS serves another device with poor coverage in the downlink [26]; we denote by d the average time interval between two consecutive NPDCCH instances. In the next section, we derive closed-form expressions for components of latency and battery lifetime models, given in (2)-(3). B. Derivations Dsyj in (2) is a function of the coverage class j. Its average value has been reported in [7, Sec. 7.3]. Drrj is given by: ∑Nrmax Drrj = (1 − Pj )ℓ−1 Pj ℓ(Draj + Drarj ) (5) ℓ=1

in which Nrmax denotes the maximum allowed number of attempts, Pj denotes the probability of successful resource

reservation in an attempt that depends on the number of devices in the class attempting the access, Draj denotes the expected latency in sending a RA message, and Drarj denotes the expected latency in receiving the RAR message. Draj is a function of time interval between consecutive scheduling of NPRACHs and is equal to 0.5 tj + cj τ , while Drarj depends on the operation of NPDCCH. NPDCCH can be seen as a queuing system in which the downlink server (see Fig. 3) visits the queue every d seconds and serves the existing requests. Thus, Drarj consists of i) waiting for NPDCCH to occur, which happens on average d/2 seconds, ii) time interval spent waiting to be served when NPDCCH occurs, denoted by Dw , and iii) transmission time, denoted by Dtj . We first characterize Dw . When the server visits the NPDCCH queue, on average there are: ∑C Q= fj (Gu + Gd ) max {d, tj } + λb d (6) j=1

requests waiting to be served, where the first term in Q corresponds to NPRACH-initiated random access requests, see (1), and λb d models the the arrival of BS-initiated control signals, see Fig. 3. Thus, the average waiting time before the service of a newly arrived RA message starts is Dw = 0.5 Q Dt , where Dt is the average service time in NPDCCH. Using u as the average control packet transmission time, the average transmission time for class j is Dtj = cj u. Thus: ∑C ∑C Dt = f j D tj = fj c j u (7) j=1

j=1

and Drarj becomes: Drarj = 0.5 d + 0.5 Q Dt + cj u.

(8)

Resource reservation of a device over NPRACH is successful if its transmitted preamble does not collide with other nodes’ preambles, which happens with probability PjRACH , and the RA response is received within period Tth , which happens with probability PjRAR . Thus, the probability of successful resource reservation can be approximated as Pj = PjRACH PjRAR . For a device belonging to class j, there are Mj orthogonal preambles available every t seconds, during which it contends on average with Nj = fj (Gu + Gd )tj devices. Then, PjRACH is derived as: ∑N (Nj )k e−Nj ( Mj − 1 )k−1 PjRACH = . (9) k=2 k! Mj The cumulative distribution function of service time for a device and sum of service times for n > 1 devices are: ∑C F1 (x) = fj H(x − cj u), (10) j=1 ∑C Fn (x) = fj Fn−1 (x − cj u) j=1

respectively, where H(x) is the unit step function. Then, PjRAR , which is the probability that RAR is received within

Tth , is: PjRAR =1−

(11)

∞ K−1 ∑ ∑ k QK e−Q (1 − FK−k (Tth )) FK−k−1 (Tth ). K K!

K=2 k=1

Dtxj is a function of scheduling of NPUSCH. Operation of NPUSCH can be seen as a queuing system in which server handles requests in a fraction of each uplink ∑ frame that is C allocated to NPUSCH; this fraction is w = 1 − j=1 cj τ /tj . Arrival of service requests to the NPUSCH can be modeled as a batch Poisson process (BPP), as resource reservation happens only ∑Cin NPRACH periods. The mean batch-size is G = C1 j=1 fj Gu tj , and the rate of batch arrivals is ∑C j=1 1/tj . The uplink transmission time is determined by the packet size and coverage class j. We assume that the packet length follows a general distribution with the first two moments equal to l1 and l2 . Then, the transmission (i.e., service) time for the uplink packet follows a general distribution with the first two moments: ∑C fj cj l1 ∑C fj c2j l2 s1 = and s2 = j=1 Rj w j=1 R2 w 2 j where Rj is the average uplink transmission rate for class j. This queuing system is a BPP/G/1 system, hence, using the results from [27], one can derive the latency in data transmission for class j as: Dtxj =

ρs2 Gs1 cj l1 + + 2s1 (1 − ρ) 2(1 − ρ) Rj w

(12)

∑C where ρ = j=1 Gs1 /tj . Similarly, performance of NPDSCH can be seen as a queuing system in which server visits the queue in a fraction of frame time and serves the requests. This fraction comprises to subframes in which NPDCCH, NPBCH, NPSS, and NSSS are not scheduled, and can be derived similarly to (8) as: Q ∑C y =1−b− fj cj u. (13) j=1 d The arrival of downlink service requests to the NPDSCH queue can be also seen as a BPP, as they arrive only after∑NPRACH has occurred. The mean batch-size is G = ∑C C 1 j=1 fj Gd tj , and the arrival rate is j=1 1/tj . The C downlink transmission time is determined by the packet size and coverage class j. Assuming that packet length follows a general distribution with moments m1 and m2 , then first two moments of the distribution of the packet transmission time are: ∑C fj cj m1 ∑C fj c2j m22 h1 = and h2 = j=1 j=1 R2 y 2 Rj y j where Rj is the average downlink data rate for coverage class ∑C 1 j. Defining ν = j=1 Gh tj , the latency in data reception Drxj becomes: 0.5νh2 Gh1 cj m1 + + . (14) Drxj = h1 (1 − ν) 2(1 − ν) Rj y

TABLE I: Parameters for performance analysis.

category parameters Traffic Traffic Traffic Traffic Power Coverage Coverage Other Other

N , S, p l1 , m1 , Tth u, τ , λb , b f1 , f2 Pt , Pc , PI , Pl c1 , c2 , M1 , M2 R1 , R2 , R1 , R2 E0 , Dsy1 , Dsy2 Commun. frame (CF)

values 20000, 0.5 h−1 , 0.8 500, 5 Kbit, 2 s 2 ms, 10 ms, 1/CF, 0.2 0.5, 0.5 0.2, 0.01, 0.01, 0.1 W 1, 2, 16, 16 5, 5, 15, 15 Kbit/s 1 KJ, 0.33 s, 0.66 s 10 ms

Finally, we derive the average energy consumption of an uplink/downlink service. Denote by ξ, PI , Pc , Pl , and Ptj the power amplifier efficiency, idle power consumption, circuit power consumption of transmission, listening power consumption, and transmit power consumption for class j. Then, Esyj = Pl Dsyj

(15)

Erarj = Pl Drarj ∑Nrmax Err = (1 − Pj )l−1 Pj (Eraj + Erarj )

(16)

Eraj = (Dra − cj τ )PI + cj τ (Pc + ξPtj ) cj l1 cj l1 )PI + (Pc + ξPtj ) Etxj = (Dtxj − Rj w Rj w cj m1 cj m1 Erxj = (Drxj − )PI + Pl Rj y Rj y

(18)

l=1

(17)

(19) (20)

from which the battery lifetime model (4) is derived as: ( Lj = E0 Sp[Esyj + Errj + Etxj + Es ] + )−1 S(1 − p)[Esyj + Errj + Erxj + Es ] . (21) IV. P ERFORMANCE E VALUATION In this section, we validate the derived expressions, highlight performance tradeoffs in channel scheduling, find optimized system operation points, and identify the mutual impact among the coexisting coverage classes. System parameters are presented in Table I. Fig. 4 compares the analytical lifetime and latency expressions derived in Section III-B (dashed curves) against the simulation results (solid curves) for class 1 of devices. The x-axis represents t, the average time between two scheduling of random access resources. It is obvious that the simulations results, including battery lifetime and service latency in uplink and downlink, match well with the respective analytical results. One further observe that by decrease in t, i.e. increase in random access opportunities, the amount of resources for uplink data transmission decreases. This in turn, results in significant increase in latency in data transmission over the

uplink channel, and hence, decreases the battery lifetime accordingly. Fig. 5 shows the mutual impact of two coexisting coverage classes in a cell, i.e., class 1 and class 2. The y-axis represents the expected battery lifetime for both classes, while the xaxis represents the the number of repetitions for class 2, i.e., c2 . Increase in c2 increases the amount of radio resources which are used for signal repetitions (i.e., coverage extension) of devices in class 2. This results in an increased latency both for class 1 and class 2 devices, and hence, increases the energy consumptions per reporting period and decreases the battery lifetime. Also, it can be seen that an increase in the fraction of nodes belonging to class 2, adversely impacts the battery lifetime performance for class 1 devices. For instance, increasing c2 from 11 to 13 decreases the average battery lifetime of class 1 nodes for 6 % when f1 = 0.95 (i.e., f2 = 0.05) and for 28 % when f1 = 0.90 (i.e., f2 = 0.1). Nevertheless, the extended coverage enables devices in class 2 to become connected to the BS, i.e., provides a deeper coverage to indoor areas. Fig. 6a shows the expected battery lifetime changing with t and d, i.e., the time intervals between two consecutive scheduling of NPRACH and NPDCCH, respectively, for the same coexistence scenario. Increasing t at first increases the lifetime of devices in both classes, as it provides more resources for NPUSCH scheduling and decreases time spent in data transmission, i.e., Dtx . After a certain point, increasing t reduces the lifetime due to the increase of the expected time in resource reservation. Similarly, increasing d at first increases the lifetime by providing more resources for NPDSCH, decreasing the time spent in data reception, Drx while after a certain point it decreases the lifetime by increasing the expected time in resource reservation. The impact of t and d on latency in uplink/downlink services is shown in Fig. 6b/Fig. 6c. If the uplink/downlink latency, or the battery consumption represents the only optimization objective, it is straightforward to derive the optimized operation points. However, Figs. 6a-6c show that overall optimization of the objectives is coupled in conflicting ways. Furthermore, Figs. 6a-6c show that the latency- and lifetimeoptimized resource allocation strategy differ on class basis; thus, selecting the optimized values of t and d depends on required quality of service (lifetime and/or latency) for each class. In Fig. 6c, we observe that the optimized t and d values for minimizing downlink latency of users belonging to coverage class 1 and 2 are the same. The reason for t consists in weak dependence of the downlink delay to t. For d, one must note that while the extra communications need of coverage class 2’s users may call for extra radio resources, and hence a higher d value, but increasing d increases the latency for downlink communications. Then, the optimized d value for both coverage classes are the same. Fig. 7 illustrates normalized lifetime and latency for class 1 when d is fixed. For instance, when d = 2 ms, the downlink and uplink latency are minimized for t = 25 ms and t = 200 ms, and lifetime is maximized for t = 65 ms. In the same way, Fig. 8 illustrates

L: Lifetime (Analytical) L: Lifetime (Simulation) Du : Uplink latency (Analytical) Du : Uplink latency (Simulation) Dd : Downlink latency (Analytical) Dd : Downlink latency (Simulation)

950

900

3

2

850

1

800

0 0.1

0.04

0.06

0.08


1500


4

Du , Dd : Up/downlink latencies (sec)


1000

1000

500

0 10 0

Du : Uplink latency (sec)


Class 2, f 1 =0.9 Class 1, f 1 =0.95 Class 2, f 1 =0.95

500

Class 1, f 1 =0.98

2.5

Class #2 Min latency @ {d=0.0057,t=0.1466}

2 1.5 1 0.5

Class 2, f 1 =0.98

d (sec)

(a) Battery lifetime L versus t and d. 3

Class 1, f 1 =0.9

Class #1 Min latency @ {d=0.0052,t=0.132}

Class 1, f 1 =0.995 Class 2, f 1 =0.995

d (sec)

0 0

5

10

15

20

V. C ONCLUSION NB-IoT access protocol scheduling has been investigated, and a tractable queuing model has been proposed to investigate impact of scheduling on service latency and battery lifetime. Using derived closed-form expressions, it has been shown that scheduling of random access, control, and data channels cannot be treated separately, as the expected latencies and energy consumptions in different channels are coupled in conflicting ways. Furthermore, the derived analytical model has been leveraged to investigate the performance impact of serving devices experiencing high pathloss, and thus needing of more signal repetitions, on latency and battery lifetime performance of other nodes. Finally, given the set of provisioned radio resources for NB-IoT and arrival traffic, optimized scheduling policies minimizing the experienced

Dd : Downlink latency (sec)

normalized lifetime and latency for class 1 when t is fixed. One observes, when t = 200 ms, the downlink and uplink latencies are minimized for d = 200 ms and d = 2 ms, and lifetime is maximized for d = 10 ms.

10 -2

10 0

10 -1

t (sec)

(b) Uplink latency Du versus t and d.

c2 : Number of repetitions for class 2

Fig. 5: Mutual impact among two coexisting classes in a cell versus number of repetitions for the second class (C = 2, c1 = 1, f2 = 1 − f1 , τ = 2 ms, d = 10 ms, t = 65 ms).

10 -2

t (sec)

Fig. 4: Comparison of analytical and simulation results versus t for class 1. d = 10 ms, S = 0.25h−1 , l1 = 200 bits, and m1 = 5Kbits.

1000

10 -1

10 -1

t (sec)

1500


10

5 Class2: Min latency@ {d=0.0469,t=0.0473}

4 3 2 0

Class1: Min latency@ {d=0.0469,t=0.0473}

10 -1

t (sec)

10 -1

10 -2

d (sec)

(c) Downlink latency Dd versus t and d.

Fig. 6: Performance as function of t and d, which are time intervals between two scheduling of NPRACH and NPDCCH, respectively. latency and maximizing the expected battery lifetime have been investigated. ACKNOWLEDGMENT The research presented in this paper was supported in part by the EIT Digital Project ACTIVE (Advanced Connectivity Platform for Vertical Segments), in part by the Celtic Plus


1 Max lifetime t=0.07

0.8

Battery lifetime Downlink latency Uplink latency

0.6

Min downlink latency t=0.11

0.4 0.2

Min downlink latency t=0.05

0 0.05

0.06

0.07

0.08

0.09

0.1

0.11

t (sec)


Fig. 7: Overall performance analysis for class 1 vs. t; d = 0.0044. 1 0.8 Max lifetime d=0.08

0.6

Min uplink latency d=0.04

0.4

Battery lifetime Downlink latency Uplink latency Min downlink latency d=0.35

0.2 0 0.05

0.09

0.13

0.23

0.33

0.5

d (sec)

Fig. 8: Overall performance analysis for class 1 vs. d; t = 1.

Project SooGreen (Service Oriented Optimization of Green Mobile Networks), and in part by the European Research Council (ERC Consolidator Grant Nr. 648382 WILLOW) within the Horizon 2020 Program. R EFERENCES [1] C. Mavromoustakis, G. Mastorakis, and J. M. Batalla, Internet of Things (IoT) in 5G mobile technologies. Springer, 2016, vol. 8. ´ Morin, M. Maman, R. Guizzetti, and A. Duda, “Comparison of the [2] E. device lifetime in wireless networks for the internet of things,” IEEE Access, vol. 5, pp. 7097–7114, 2017. [3] W. Yang et al., “Narrowband wireless access for low-power massive internet of things: A bandwidth perspective,” IEEE Wireless Communications, vol. 24, no. 3, pp. 138–145, 2017. [4] A. Azari and C. Cavdar, “Self-organized low-power IoT Networks: A distributed learning approach,” arXiv preprint arXiv:1807.09333, 2018. [5] M. E. Soussi, P. Zand, F. Pasveer, and G. Dolmans, “Evaluating the Performance of eMTC and NB-IoT for Smart City Applications,” arXiv preprint arXiv:1711.07268, 2017. [6] R. Ratasuk, N. Mangalvedhe, A. Ghosh, and B. Vejlgaard, “Narrowband LTE-M System for M2M Communication,” in IEEE VTC-Fall), Sept 2014, pp. 1–5. [7] 3GPP TR 45.820, “Technical Specification Group GSM/EDGE Radio Access Network; Cellular system support for ultra-low complexity and low throughput Internet of Things (CIoT),” 2015. [8] J. Schlienz and D. Raddino, “Narrowband internet of things,” Rohde and Schwarz, Tech. Rep., 08 2016. [9] R. Ratasuk, B. Vejlgaard, N. Mangalvedhe, and A. Ghosh, “NB-IoT system for M2M communication,” in IEEE WCNC, 2016, pp. 1–5. [10] Y. P. E. Wang et al., “A primer on 3GPP narrowband internet of things,” IEEE Communications Mag., vol. 55, no. 3, pp. 117–123, March 2017.

[11] X. Lin, A. Adhikary, and Y. P. E. Wang, “Random access preamble design and detection for 3GPP Narrowband IoT systems,” IEEE Wireless Communications Letters, vol. 5, no. 6, pp. 640–643, Dec 2016. [12] T. Kim, D. M. Kim, N. Pratas, P. Popovski, and D. K. Sung, “An enhanced access reservation protocol with a partial preamble transmission mechanism in NB-IoT systems,” IEEE Communications Letters, June 2017. [13] C. Yu et al., “Uplink scheduling and link adaptation for narrowband internet of things systems,” IEEE Access, vol. 5, pp. 1724–1734, 2017. [14] M. Lauridsen et al., “Coverage and capacity analysis of LTE-M and NB-IoT in a rural area,” in IEEE VTC Fall, 2016, pp. 1–5. [15] A. Adhikary, X. Lin, and Y. P. E. Wang, “Performance Evaluation of NB-IoT Coverage,” in IEEE VTC-Fall, Sept 2016, pp. 1–5. [16] Y. D. Beyene, R. Jantti, K. Ruttik, and S. Iraji, “On the Performance of Narrow-Band Internet of Things (NB-IoT),” in 2017 IEEE WCNC, March 2017. [17] P. A. Maldonado et al., “Narrowband IoT data transmission procedures for massive machine-type communications,” IEEE Network, vol. 31, no. 6, pp. 8–15, November 2017. [18] Y. Chen, S. Zhang, S. Xu, and G. Y. Li, “Fundamental trade-offs on green wireless networks,” IEEE Communications Magazine, vol. 49, no. 6, pp. 30–37, June 2011. [19] S. H. Alonso, M. R. Prez, M. F. Veiga, and C. L. Garca, “Adaptive DRX Scheme to Improve Energy Efficiency in LTE Networks With Bounded Delay,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 12, pp. 2963–2973, Dec 2015. [20] A. Azari and G. Miao, “Network lifetime maximization for cellularbased M2M networks,” IEEE Access, vol. 5, pp. 18 927–18 940, 2017. [21] A. Aijaz, M. Tshangini, M. R. Nakhai, X. Chu, and A. H. Aghvami, “Energy-Efficient Uplink Resource Allocation in LTE Networks With M2M/H2H Co-Existence Under Statistical QoS Guarantees,” IEEE Transactions on Communications, vol. 62, no. 7, pp. 2353–2365, July 2014. [22] K. Wang, J. A. Zarate, and M. Dohler, “Energy-efficiency of LTE for small data machine-to-machine communications,” in 2013 IEEE ICC, June 2013, pp. 4120–4124. [23] G. C. Madueno, J. J. Nielsen, D. M. Kim, N. K. Pratas, C. Stefanovic, and P. Popovski, “Assessment of LTE Wireless Access for Monitoring of Energy Distribution in the Smart Grid,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 3, pp. 675–688, March 2016. [24] P. A. Maldonado et al., “Optimized LTE data transmission procedures for IoT: Device side energy consumption analysis,” in 2017 ICC Workshops, May 2017, pp. 540–545. [25] A. Azari and G. Miao, “Network lifetime maximization for cellularbased M2M networks,” IEEE Access, vol. 5, pp. 18 927–18 940, 2017. [26] 3GPP TSG- RAN1 AdHoc NB-IoT, “NPDSCH resource allocation,” Tech. Rep., March 2016. [27] H. Akimaru and K. Kawashima, Teletraffic: theory and applications. Springer Science and Business Media, 2012.

Paper C: Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks Amin Azari, Meysam Masoudi, and Cicek Cavdar Submitted to IEEE Transactions on Communications, 2018. The conference version of this manuscript has been accepted in IEEE Globecom 2018 [96].

127

Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks Amin Azari, Meysam Masoudi and Cicek Cavdar KTH Royal Institute of Technology, Email: {aazari,masoudi,cavdar}@kth.se Abstract—Grant-free radio access is a promising solution for reducing energy consumption and access delay in lowpower wide-area (LPWA) Internet of Things (IoT) networks. This work is devoted to reliability modeling, battery-lifetime analysis, resource provisioning, and operation control for grantfree IoT networks. Our modeling captures correlation in devices’ locations, benefits from 3D (time/frequency/code) interference analysis, and enables coexistence analysis of multi-type IoT technologies. We derive the interplay amongst density of the access points, communication bandwidth, traffic volume, and quality of service (QoS) of communications. Deriving the interplay enables scalability analysis, i.e. it figures out the required increase in device’s energy consumption (or access network’s resources) for compensating the increase in traffic volume or QoS demand. Our major contribution consists in deriving traffic loads and respective exchange rates in which, energy and cost resources of devices and the access network, respectively, could be traded to achieve a given level of QoS. We further indicate operation regions in which scaling a parameter turns from being a friend into a foe. Finally, we present energy- and cost-optimized operation control and resource provisioning strategies, respectively. The simulation results confirm tightness of the analytical expressions, and indicate usefulness of them in planning and operation control of IoT networks. Index Terms—5G, Battery lifetime, Coexistence, Grant-free, Reliability, LPWA IoT.

I. I NTRODUCTION Providing connectivity for massive Internet-of-Things (IoT) devices is a key driver of 5G [1]. Until now, several solutions have been proposed for enabling large-scale IoT connectivity, including evolutionary and revolutionary solutions [2]. Evolutionary solutions aim at enhancing connectivity procedure of existing LTE networks, e.g. access reservation and scheduling improvement [3, 4]. On the other hand, revolutionary solutions aim at providing scalable low-power IoT connectivity by redesigning the access network. In 3GPP LTE Rel. 13, narrowband IoT (NB-IoT) has been announced as a revolutionary solution which handles communications over a 200 KHz bandwidth [5]. This narrow bandwidth brings high link budget, and offers extended coverage [5]. In order to provide low-latency access to radio resources, grant-free radio access is a study item in 3GPP IoT working groups, and it is expected to be included in 5g new radio (NR) [6]. Thanks to the simplified connectivity procedure, and removing the need for pairing and fine synchronization, grant-free radio access has attracted lots of interests in recent years for providing low-power wide-area (LPWA) IoT connectivity, especially when more than 10 years lifetime is required. SigFox and LoRaWAN are two dominant grant-free radio access solutions

over the unlicensed spectrum [2]. While energy consumptions of LoRaWAN and SigFox solutions are extremely low, and their provided link budget is enough to penetrate to most indoor areas, e.g. LoRaWAN signal can be decoded when it is 20 dB less than the noise level, reliability of their communications in coexistence scenarios is questionable [7, 8]. Experimental measurements in several coexistence scenarios, where multiple IoT technologies share a set of radio resources, have been carried out in [7]. The results in [7] confirm significant impact of inter-technology interference on reliability of IoT communications. Regarding the growing interest in grant-free radio access for IoT communications in the 3GPP and proprietary cellular networks [2, 6], here we investigate the reliability, energy efficiency, and scalability in grant-free radio access. For grant-free enabled 3GPP networks, in which a limited set of authenticated devices get grant-free access to a pool of radio resources [6], our study offers an analytical scalability analysis tool, which indicates how the provisioned pool of radio resources (or device-side parameters) should be scaled with the number of authenticated devices. For LPWA IoT networks over the unlicensed spectrum, e.g. LoRaWAN, our study offers an analytical scalability analysis tool, which indicates how the number of access points (or device-side parameters) should be scaled with the traffic/interference load. A. Literature Study Non-orthogonal radio access, in which simultaneous transmissions have time/frequency overlapping, has attracted lots of attentions in recent years as a complementary radio access scheme for future generations of wireless networks [9, 10]. In literature, non-orthogonal access has been mainly employed in order to increase the network throughput [11], reliability [12], battery lifetime [13], and reduce the access delay [11]. In [14], grant-free access to uplink radio resources of cellular networks has been analyzed for intra-group communications of IoT devices. In [13], a novel receiver for grant-free radio access IoT networks has been designed, which benefits from oscillator imperfection of cheap IoT devices for contention resolution. In [15], a grant-free rateless multiple access scheme has been introduced. This scheme leverages the unique pattern of each device in uplink transmission for joint user detection and data decoding through an iterative algorithm. Application of compressed sensing techniques in grant-free multiple access has been investigated in [16]. In [17], authors show that massive multiple input multiple output (MIMO) is well-suited for massive grant-free IoT networks because the device detection

error tends to zero asymptotically as the number of antennas at the receiver goes to infinity by using the multiple measurement vector (MMV) compressed sensing techniques. In [18], a novel grant-free radio access scheme based on direct sequence spread spectrum (DSSS) technique has been proposed in order to transmit emergency data without access reservation over other user’s uplink traffic. In [19], outage probability in grantfree access has been studied by assuming a constant received power from all contending devices, which is not the case in practice regarding the limited transmit-power of IoT devices, as well as lack of channel state information at the device-side for power control. The success probability in grant-free radio access has been also analyzed in [8, 20] by assuming a Poisson point process (PPP) distribution of IoT devices. Performance of successive joint decoding (SJD) and successive interference cancellation (SIC) in detection/decoding of collided packets in a grant-free IoT network has been investigated in [21]. One sees the research on grant-free radio access has been mainly focused on success probability analysis in homogeneous scenarios, and there is lack of research on performance analysis of large-scale IoT networks with multi-type IoT devices with heterogeneous communications characteristics. Furthermore, when it comes to the distribution of devices in wide-area IoT networks, PPP has been mainly used. However, this assumption may lead to inaccurate results [22, 23] due to the cell ranges that can go up to tens of kilometers [2] and hotspots. In hot-spots, e.g. buildings and shopping centers, a high density of IoT devices exist; while outside them, a low density of devices exists. Then, a Poisson cluster process (PCP), which takes the correlation between locations of devices into account, suits well for the distribution process of devices in LPWA IoT networks [22, 23]. Furthermore, characterization of the 3D interference from partial overlapping of short IoT packets in time, frequency, and code domains has not been investigated in prior works. Our results on analytical characterization of the 3D interference are expected to attract attention in a diverse set of problems, from IoT networks in which devices have cheap oscillators [13], to satellite communications in which achieving fine synchronization is more complicated due to propagation delay [24]. Finally, there is lack of a unified approach investigating the tradeoff between economic viability, i.e. the required investment cost in the access network, and quality of service for IoT devices, i.e. reliability and durability of communications. Enabling IoT connectivity requires deployment of access points (APs) and allocation of frequency resources, which increase the network costs. On the other hand, the experienced delay, consumed energy, and success of IoT applications have strong couplings with reliability of data transfer, which is in turn a function of provisioned network resources. This work aims at filling the aforementioned research gaps. B. Contributions Here, we address an important problem, not tackled previously: network design and operation control in coexistence

scenarios with grant-free radio access. The main contributions of this work include: • Analytical modeling: – Provide a rigorous analytical model of reliability for heterogeneous LPWA IoT networks, powered by stochastic geometry analysis of the 3D interference, in terms of provisioned network resources, and characteristics of the traffic. – Provide an analytical model of battery lifetime for IoT devices in terms of device’s communication parameters, and performance of the network. – Analyze the tradeoffs among network cost, battery lifetime, and reliability of communications. Present the operation regions in which tuning device’s communication parameters, increases both reliability and battery lifetime, offers a tradeoff between them, and decreases both of them. • Optimized design: – Present cost-optimized resource provisioning for the access network, and lifetime-optimized operation control for IoT devices, subject to reliability constraints. • Scalability analysis: – Present scalability of the access network by investigating the point up to which the amount of provisioned resources (energy consumption of devices per packet) should be scaled to compensate the increase in traffic load, interference, or QoS requirement. Investigate the scaling bounds and limitations. Fig. 1 represents a graphical illustration of contributions. The remainder of paper has been organized as follows. System model and problem description are presented in the next section. Modeling of KPIs is presented in section III. Section IV presents the optimized resource provisioning and operation control strategies. Simulation results are presented in section V. Concluding remarks are given in section VI. II. S YSTEM M ODEL AND P ROBLEM D ESCRIPTION A. System Model A set of IoT devices, denoted by Φ, have been distributed according to different spatial PCPs in a wide service area. ∆ Φ comprises of K subsets, Φk for k ∈ K = {1, · · · , K}, where each subset refers to a specific type of IoT service. Traffic from different subsets differ in the way they use the time-frequency radio resources, i.e. in frequency of packet generation 1/Tk , signal bandwidth wk , packet transmission time τk , number of replicas1 transmitted per packet nk , and transmit power Pk . Subscript k refers to the type of IoT devices. For PCP of type-k IoT traffic, the (λk , υk , f(x)) tuple characterizes the distribution process in which, λk is the density of the parent points and υk is the average number 1 Practical motivations for modeling such replicas can be found in state of the art IoT technologies like NB-IoT and SigFox in which, coverage extension and resilience to interference are achieved by repetitions of transmitted packets [2, 5]. When it is not the case, nk = 1 can be used.


Provisioned resources: Density of APs, system BW


Provisioned resources: Density of APs, system BW

Analysis

Analysis Optimized resource provisioning

Eco. viability of access network





Optimized operation control

Probability of success in a transmission




(a)

(b)

Fig. 1: (a) optimized resource provisioning in the access network; (b) optimized operation control for IoT devices. Violet-colored boxes represent the contributions. 1000

AP IoT device (type 1) IoT device (type 2) IoT device (type 3)

y-axis (meter)

500

0

-500

-1000 -1000

-500

0

500

1000

x-axis (meter)

(a) Graphical illustration of locations of devices for K = 3 which (b) A snapshot of received traffic from 4 transmitters (K = 3) at the represents heterogeneity in distribution processes of locations. receiver which represents differences in communication characteristics.

Fig. 2: Graphical description of the system model. Performance analysis and optimization with heterogeneity in communication characteristics as well as distribution processes of devices have not been studied before. of daughter points per parent point2 , as defined in [23]. Also, f(x) is an isotropic function representing scattering density of the cluster points around the cluster center, e.g. a normal distribution: √ f(x) = exp(−||x − x0 ||2 /(2σ 2 ))/ 2πσ 2 , (1) where σ is the variance of distribution and x0 is the location of the parent point. A frequency spectrum of W is shared for communications, on which the power spectral density of noise is denoted by N . We aim at collecting data from a subset of K, denoted by ϕ, where |ϕ| ≤ |K|, and treat transmissions from other devices as interference. When |ϕ| = |Φ|, the system model is in match with grant-free radio access over cellular 2 In PCP deployment, we have clusters of devices, where each cluster models a hot-spot. λk represents density of such clusters of devices, i.e. the parent points. υk represents the average number of devices in each cluster, i.e. the daughter points. Finally, f(x) represents how devices are distributed in each cluster.

networks in which, only authenticated devices receive grantfree access to the resource pool. When |ϕ| < |Φ|, the system model is in match with grant-free radio access over unlicensed band in which, coexisting technologies may reuse the same resource simultaneously. Devices in ϕ may also share a set of semi-orthogonal codes denoted by ϖ with cardinality |ϖ|, which reduces the interference from other devices reusing the same radio resource with a different code by factor of Q. Examples of such codes are semi-orthogonal spreading codes in LoRaWAN technology [2]. A list of frequently used symbols has been presented in Table I for ease of reading. III. A NALYTICAL M ODELING OF KPI S A. Modeling of Reliability In heterogeneous grant-free radio access networks, transmitting devices are asynchronous and may have different radio

TABLE I: Frequently used symbols Symbol

Description (k refers to IoT type.)

A Ψk Ψ; K Φk ; Φ ϕ; |ϕ| LIΨ x, y, z τk Tk γth Pk λk υk f(x) m, Ω g(·); δ; h Qj W ϖ Ps ; Po N, N nk L(k), L(k) λa Notation Notation

Service area Distribution process of locations of interfering devices ∪k∈K Ψk ; Set of IoT device types: {1, · · · , K} Set of type-k devices; ∪k∈K Φk Subset of interest out of K; Cardinality of ϕ Laplace functional of interference from Ψ Points in the 2D plane Packet transmission time reporting period Threshold SINR Transmit Power (IoT device) Density of parent points Avg. number of nodes per cluster Distribution function of daughter points of a parent point Parameters of Nakagami-m channels Pathloss function; Pathloss exponent; fading Interference rejection factor System bandwidth Number of shared (semi-)orthogonal codes Success probability; Outage probability Noise; Noise power Number of transmitted replicas per packet Battery lifetime (device, application) Density of APs a: Symbol, {a, A}: parameter, a: vector {A(·), A(·), A(·)}: Function

resource usage patterns3 , and hence, the received packets at the receiver can have partial overlapping in time-frequency. In order to derive an analytical model of reliability in such networks, we first derive analytical expressions for the received 3D interference in subsection III-A1, and for probability of success in subsection III-A2. These models are then employed in deriving reliability of communications in subsection III-A3. 1) Interference Analysis: We assume a type-i device, i ∈ ϕ, has been located at point z in a plane, and its respective AP has been located at the origin. In order to derive probability of success in data transmission from the device to the AP, we need to characterize the received interference at the AP. A common practice in interference analysis is to determine its moments, which is possible by finding its generating function, i.e. the Laplace functional [22, 25]. Towards this end, let us introduce three stationary and isotropic processes: i) (1) (1) Ψ(1) = ∪k∈K Ψk , where Ψk represents the PCP containing locations of type-k interfering nodes which are transmitting with a similar code to the code4 of transmitter of interest; ii) (2) (2) Ψ(2) = ∪k∈K Ψk , where Ψk represents the PCP containing locations of type-k interfering nodes which are transmitting data with a different code (or no code, in case k ∈ / ϕ) than the (j) transmitter of interest; and iii) Ψ = ∪j∈{1,2} Ψk . For an AP located at the origin, the Laplace functional of the received

interference at the receiver is given by: [ ] LIΨ (s) = E exp(−sIΨ ) (2) ∏ ∏ [∏ ] =E (j) Lh (sQj Pk g(x)) , j∈{1,2}

ph (q) =

x∈Ψk

( mq ) m 1 ( )m q m−1 exp − , Γ(m) Ω Ω

(3)

where ( Γ is the )Gamma function. Then, using Laplace table, Lh sQj Pk g(x) is derived as: ( )−m Lh (sQj Pk g(x)) = 1 + ΩsPk g(x)/m . (4) By inserting (4) in (2) and considering the fact that the received interferences from different devices are independent, we have: [∏ ] ∏ (∏ ) LIΨ (s) = Ex,y , (j) u(x, y) j,k

y∈Θk

x∈θy

where k ∈ K, j ∈ {1, 2}, the set of parent points of type-k is denoted by Θk , and the set of transmitting nodes which are (j) daughter points of y is denoted as θy . Also, Ex represents expectation over x, and ( )−m u(x, y) = 1+ΩsQj Pk g(x−y)/m . The received interference over the packet of interest can be decomposed into two parts: i) interference from transmitters belonging to the cluster of transmitter, i.e. daughter points of the same parent point; and ii) other transmitters. Let us denote the Laplace functional of interference from the former and latter transmitters as L†IΨ (s) and L‡IΨ (s) respectively. Then, we have: LIΨ (s) = L‡IΨ (s)L†IΨ (s). (5) To proceed further, we recall a useful lemma from [26]. Lemma 3.1: If X ⊂ R2 is assumed to be a PPP with intensity function µ(x), for any Borel function b: R2 → [0, 1], we have: ∫ [∏ ] ( ) Ex b(x) = exp − (1 − b(x))µ(x)dx . (6) x∈X

3 The

packet transmission time, signal bandwidth, and packet generation rate differ from one another. 4 Note: as mentioned in the system model, devices in ϕ share a set of semi-orthogonal codes for interference management.

k∈K

where Qj Pk g(x) is the average received power due to a typek transmitter at point x, Q1 = 1, Q2 = Q, and Q is the rate of rejection of interference between two transmissions with different multiple access codes, as defined in section II-A. Also, h is the power fading coefficient associated ( with the )channel between the device and the AP, and Lh sQj Pk g(x) is the Laplace functional of the received power. We consider the following general path-loss model g(x) = 1/(α1 + α2 ||x||δ ), where δ is the pathloss exponent, and α1 and α2 are control parameters. When h follows Nakagami-m fading, with the shaping and spread parameters of m ∈ Zi and Ω > 0 respectively, the probability density function (PDF) of the power fading coefficient is given by:

R2

Proof Presented in [26]. (j)

Using Lemma 3.1, and conditioning on Θk and θy , one has:

L‡IΨ (s) (7) [ ∏ ] ∫ ∏ { ( )} exp -ˆ υk,j [1-u(x, y)]f(x)dx , = Ey j,k

R2

y∈Θk

(∑ = exp - λk j,k

∫

{ ( 1- exp -ˆ υk,j

R2

∫

[1-u(x, y)]f(x)dx

)}

) dy .

R2

In (7), the average numbers of interfering type-k devices in each cluster are denoted as υˆk,j = υk ρk,j , in which, ρk,j is the 3D activity factor of devices, and is derived as the product of device’s activity factors in time, frequency, and code domains, i.e. ρk,j = ρtk,j ρfk,j ρck,j . In a single carrier, single-code, time-synchronized system, ρk,j is equal to one. k 1 In our system, for k ∈ ϕ, ρk,1 is derived as nTkkτk w W |ϖ| , and

k |ϖ|−1 ρk,2 is derived as nTkkτk w W |ϖ| . In these two expressions, the first fraction represents the percentage of time in which device is transmitting, i.e. the time activity factor, the second fraction represents the ratio of system bandwidth that device occupies in each transmission, i.e. the frequency activity-factor, and the third fraction represents the code-domain activity factor, which is clearly different for j = 1 and j = 2. For k ∈ / ϕ, in which devices don’t utilize semi-orthogonal codes for communications, there is no type-k device using an orthogonal code to the code of transmission of interest. Then, ρfk,1 = 0, and ρfk,2 = 1, and time/frequency activity factors are the same as the prior case. Following the same procedure used for deriving L‡IΨ (s), one can derive L†IΨ (s) as: ∏ [ ∏ ] Ey Ex [ (8) L†IΨ (s)= (j) u(x, y)] j∈{1,2} x∈θy ) ∫ ∫ ( ) (∑ = exp - j υî,j 1-u(x, y) f(x)dx f(y)dy.

R2

R2

Now, by inserting (7) and (8) in (5), the Laplace functional of interference is derived. 2) Probability of Success in Transmission: Let N denote the additive noise at the receiver. Using the above derived interference model, probability of success in packet transmission of a type-i device, located at z, to the AP, which has been located at the origin, is: ps (i, z) = Pr(Pi hg(z) ≥ [N + IΨ ]γth ) (9) ∫ ∞ m-1 ∑ 1 γth mq (a) = exp(− )q ν dPr(IΨ +N ≥ q) ν! 0 ΩPi g(z) ν=0 ν (b) ∑m-1 (−1) = [LIΨ (s)LN (s)](ν) s= γth m , ν=0 ΩPi g(z) ν! ν

∂ where [F (s)](ν) = ∂s ν F (s), (a) follows from [25, Appendix C] and equation (3), and (b) follows from [27, ∂n Lemma 3.1] and the fact that L(tn f(t)) = (−1)n ∂s n F (s). Furthermore, LIΨ has been characterized in (7) and (8), and LN (s) is the Laplace transform of noise. In order to get insights on how coexisting services affect each other, in the following we focus on m = 1, i.e. the Rayleigh fading,

and present a closed-form approximation of the success probability. In section V, we will evaluate tightness of this approximation. Theorem 3.2: For m = 1, success probability in packet transmission could be approximated as: ∑ ∑ [ ( Qj Pk γth )] ) ps (i, z) ≈ PN exp − λk υˆk,j H(z, 1, ΩPi (

× exp −

∑

j∈{1,2} k∈K

j∈{1,2}

υî,j H(z, f∗ (x),

Qj γth ) ) , (10) Ω

( ) where f∗ (·) = conv f(·), f(·) , ∫ ( g(x) f∗ (x)dx, H z, f∗ (x), ξ) = x∈R2 g(x) + g(z)/ξ ( ) PN = exp − N γth /[ΩPi g(z)] .

(11) (12)

Proof Appendix A. H(z, f∗ (x), ξ) and H(z, 1, ξ) could be derived in closed-form for most well-known pathloss and distribution functions, as follows. Corollary 3.3: For g(x) = α||x||−δ , 2

H(z, 1, ξ) = ||z||2 ξ δ 2π 2 csc(2π/δ)/δ.

(13)

Proof By change of coordinates, i.e. x → (r, θ), we have: ∫ −δ ( α||x|| H z, 1, ξ) = dx, −δ −δ 2 + α||z|| /ξ x∈R α||x|| ∫ ∞ 1 = 2π rdr. 1 + (r/||z||)δ /ξ 0 Solving this integral by using [28, Eq. 3.352], (13) is derived. Corollary 3.4: For g(x) = α||x||−4 , and f(x) given in (1), [ ||z||2 ||z||2 ||z||2 H(z, f∗ (x), ξ) = 2 √ ci( 2 √ ) sin( 2 √ )− 4σ ξ 4σ ξ 4σ ξ ] ||z||2 ||z||2 si( 2 √ ) cos( 2 √ ) , 4σ ξ 4σ ξ where si(·) and ci(·) are well-known sine and cosine integrals, as follows: ∫ ∞ ∫ ∞ sin(t) cos(t) si(x) = − dt, ci(x) = − dt. t t x x Proof Appendix B.

( Remark Analysis of H√ z, f∗ (x), ξ) shows that it can be well 2 ≫ 1. For theorem 3.2 in which approximated by (1 for ξ||z|| 4σ 2 ∗ ξ = Qj γth /Ω, H z, f (x), ξ) ≈ 0 for j = 1 because Q1 = ( ∗ ∆ Q√ ≈ 0; and H z, f (x), ξ) ≈ 1 for j = 2 when z ≫ z0 = 4 2σ Ω √ because Q2 = 1. 4 γ th Remark From theorem 3.2, one sees that probability of success, ps (i, z), is a function of ||z|| rather than phase of z. Then, hereafter we use p(i, z) to denote probability of success for communication distance of z.

Until now, we have derived probability of success as a function of communication distance to the respective AP. In the following, we investigate success probability when multiple APs are accessible in the service area. Regarding the fact that theorem 3.2 provides probability of success as a function of communication distance, given the distribution process of APs, the expected communication distance to the neighboring APs, and hence, probability of success in data transmission could be derived. In a homogeneous deployment of APs with density λa , the PDF of distance from a random point to the ℓth nearest AP, denoted by dℓ is given by [29]: Pdℓ (r) = exp(−λa πr2 )2(λa πr2 )ℓ /[r(ℓ − 1)!]. Then, one can derive the average probability of success in packet transmission from a typical type-i device as: ∏ℓmax ∫ ∞ ( ) Ps (i) = 1 − 1 − ps (i, r) Pdℓ (r)dr. (14) ℓ=1

0

Theorem 3.5: For f(x) given in (1), and g(z) = α||z||−4 , we have: Ps (i) ≈ 1 −

∏ℓmax [ ] X2 2 X0 exp( 1− √ 2 )G(X3 , ℓ) , ℓ−1 ℓ=1 4X1 X1

B. Battery Lifetime Performance (Durability) For most reporting IoT applications, packet generation at each device could be seen as a Poisson process [30]. Then, one can model energy consumption of a device as a semiregenerative process where the regeneration point has been located at the end of each successful data transmission epoch [4]. For a given device of type-i, let us denote the distance to ℓth neighboring AP as dℓ , the stored energy in batteries as E0 , static energy consumption per reporting period for data acquisition from environment and processing as Est , circuit power consumption in transmission mode as Pc , and inverse of power amplifier efficiency as η. Then, the expected battery lifetime is derived as [4]: L(i) =

E0 Ti , Est + βi Ec + βi ni (ηPi + Pc )τi

(16)

where Ec represents the average energy consumption in listening after each trial for ACK reception, and βi represents the average number of trials and is derived as: ] max Bi [ max ∑ [ ℓ∏ ]ni [ ℓ∏ ]ni [j−1] βi = j 1− 1−ps (i, dℓ ) 1−ps (i, dℓ ) , j=1

ℓ=1

ℓ=1

where ps (i, r) has been derived in theorem 3.2. Motivated by battery lifetime definition for an IoT device, we can further define battery lifetime for an IoT application. The battery lifetime of an IoT application could be defined as the length ( ) (λa π)ℓ N γth of time between the reference time and when application where X0 = exp − υî,2 , X1 = , (ℓ − 1)! ΩPi α is considered to be nonfunctional. The instant at which an 2 ∑ π X2 γth Qj Pk 0.5 π ) csc( ) + λa π, X3 = √ , application becomes nonfunctional is dependent on the level X2 = λk υˆk,j ( ΩPi 2 2 2 X1 of correlation between gathered data by neighboring devices. j,k In critical applications with sparse deployment of sensors, ∫∞ G(X3 , ℓ) = X2 2 (z-X3 )(ℓ−1) exp(−z 2 )dz, and is derived for where losing even one node deteriorates the performance 2X1 or coverage, the shortest individual battery lifetime (SIBL) any ℓ in the form of error function, e.g. for ℓmax = 2: may characterize the application lifetime. When correlation √ amongst gathered data by neighboring nodes is higher, the G(X3 , 1) = −( π(erf(X3 ) − 1))/2, √ longest individual battery lifetime (LIBL), or average indi2 G(X3 , 2) = exp(−X3 )/2 + (X3 π(erf(X3 ) − 1))/2. vidual battery lifetime (AIBL) may characterize the application lifetime. Using AIBL definition, the expected battery Proof Appendix C. lifetime for IoT application of type k can be approximated as 3) Reliability of Communications: Now, we have the re- L(i) ≈ L(i) βi =βî , where: quired tools to derive an analytical model of reliability of ∑Bi [ ][ ]ni [j−1] βî = j 1-[1-Ps (i)]ni 1-Ps (i) , (17) IoT communications. Once a type-i device has a packet to j=1 transmit, it transmits nk replicas of the packet, and listens for ACK from the AP(s). If No ACK is received in a bounded and L(i) and Ps (i) have been derived in (16) and theorem 3.5 listening window, device retransmits the packet, and this respectively. procedure could be repeated up to Bi − 1 times, where the C. Access Network’s Cost bound may come from the fair use of the shared medium [2, The access network’s cost can be modeled as sum of 7] or expiration of data. If data transmission is unsuccessful in Bi attempts, we call it an outage event. Then, the probability spectrum, infrastructure, and operation costs [31]. Then, the of outage for a typical type-i device in such setting could be total annual cost in a service area of A, is derived as: approximated as: [ ]ni Bi , Po (i) = 1 − Ps (i) where Ps (i) has been derived in theorem 3.5.

Ctot =c1 λa A+c2 λa AEcons +c3 W, (15)

(18)

where c1 [e/AP] is the annual cost per AP excluding the energy cost, c2 [e/Joule] is the annual cost for energy, and c3 [e/Hz] is the annualized spectrum cost. c3 is zero for IoT

solutions over the unlicensed spectrum. Also, Econs is the energy consumption per unit time, and depending on the IoT technology, could be modeled in different ways. A proposed model for Econs is as follows: ∑ Econs = Pr + Pa Λk , k∈ϕ

in which Pr is the load-independent power consumption, e.g. power consumption in listening to the channel and processing, Pa is the load-dependent power consumption, e.g. power consumption in forwarding received data to the core network and responding to the sender, and Λk is the arrival rate of packets of type-k devices. IV. O PTIMIZED R ESOURCE P ROVISIONING AND O PERATION C ONTROL A. Tradeoff Analysis From the reliability expression in (15) and theorem 3.5, one sees that probability of success for type-i traffic increases in (i) increase in the density of the APs, i.e. λa , which reduces the average communication distance; (ii) increase in communication bandwidth, i.e. W , which reduces the probability of collision; (iii) increase in device transmit power, i.e. Pi , and (iv) increase in number of replicas per packet, i.e. ni . (18) represents that network cost increases with increase in the density of the APs and bandwidth of communication (in the case of licensed spectrum). Then, there is a clear tradeoff between reliability of provided communications and the investment cost. Furthermore, regarding the fact that the impacts of bandwidth and AP density on the reliability of communications are not the same, we can seek for an optimal investment strategy to minimize the cost while complying with the reliability constraints (subsection IV-B). From the battery lifetime analysis in (16), one sees that lifetime of devices may decrease in ni and Pi because they increase the consumed energy per data transfer. On the other hand, when reliability of communication is lower than a threshold, increase in ni and Pi may decrease the need for listening to the channel for ACK arrival and retransmissions, and hence, increasing ni or Pi may increase the battery lifetime. Taking these facts into account, one sees there exists an operation point beyond which, increase in Pi and ni offers a tradeoff between reliability and lifetime, and before which, it increases both reliability and durability of communication. This observation will be evaluated using simulation results in section V. From the above discussion, one sees that the design objectives, i.e. network cost, battery lifetime, and reliability of communications, cannot be treated separately in resource provisioning and operation control problems because they are coupled in conflicting ways such that improvements in one objective may lead to deterioration of the others. Taking these couplings into account, in the following we aim at optimizing network-side resource provisioning and device-side operation control.

B. Optimized Resource Provisioning As mentioned above, increasing W and λa have different impacts on reliability of communications as well as costs of the access network. Then, an interesting research problem consists in deriving the optimized balance between densification and spectrum allocation to IoT traffic. This problem can be modeled as: ∆

minimize Ctot = [c1 A+c2 AEcons ]λa +c3 W, λa ,W

s.t.: Po (i) ≤

Preq o (i), ∀i

(19)

∈ ϕ.,

Preq o (i)

where is the required reliability level. The reliability constraint in (25) could be rewritten as the minimum required success probability in communications as follows: √ ∆ req 1 − ni Bi Preq (20) o (i) = Ps (i) ≤ Ps (i). By using the Ps (i) expression in theorem 3.5, and satisfying (20) with equality, we have: ∫ ∞ Preq X0 exp(-X5 r2 )2rdr s (i) = 0 ( ) √ 0.5 πλa π exp − υî,2 =∑ , (21) 2 N γth ˆk,2 ( PPkiγΩth )0.5 π2 csc( π2 )+λa π+ ΩP k λk υ iα in which, ℓmax = 1, δ = 2, and Q ≈ 1 have been assumed for brevity of expressions. Also, X5 is an auxiliary variable equal to the denominator of (21). This equation can be rewritten as: √ 0.5 πλa A1 A2 N γth exp(− ) = + λa + , (22) req Ps (i) W W πΩPi α where A1 = υi A2 =

∑ k

υk

ni τi ϖ − 1 wi , Ti ϖ

nk τ i ϖ − 1 Pk γth 0.5 π π wk λi ( ) csc( ). Tk ϖ Pi Ω 2 2

Solving (22) for λa , we derive an important expression for bandwidth-AP density tradeoff, as follows: λa =

N γth A2 πΩPi α + W √ 0.5 π A1 exp( W ) − Preq s

.

(23)

1

This expression sheds light on the interconnection between the required bandwidth and AP density in providing a required level of reliability for IoT communications. Using (23), the optimization √ problem in (19) reduces to a simple search over W ≥ log( 0.5Preq π )/A1 for minimizing: s

{

[c1 A+c2 AEcons ] max i∈ϕ

γth N A2 } πΩPi α + W √ +c3 W. 0.5 π A1 -1 Preq s (i) exp( W )

(24)

Fig. 3 represents the tradeoff presented in (23) as well as the objective cost function in (24), which is a quasi-convex function. One sees while the required density of APs decreases by increase in the system bandwidth; the network cost decreases to some point, and beyond which, network cost increases in the system bandwidth. The point at which change in behavior

TABLE II: Simulation Parameters

×10 -6

200

Network cost ( × c 1 )

Req. density of APs ( × Km -2 )

1.5

150 1

Optimized resource provisioning

100

0.5 50 0 5

10

0 20

15

System BW ( × 10 KHz)

Fig. 3: The optimized resource provisioning problem. Required ps (1, deg ) = 0.5, λ1 =1.6, and other parameters can be found in Table II. of cost function occurs is a function of cost parameters, i.e. c1 , c2 , c3 . This point moves towards the right-side of Fig. 3 by decrease in cc31 , i.e. by decreasing the cost of bandwidth in compassion with the AP. C. Optimized Operation Control Here, we aim at maximizing battery lifetime of IoT devices by optimizing their communication parameters. Using the battery lifetime definition in (16), one can define the optimization problem for deriving the optimized operating point of type-i IoT devices as follows: maximize L(i);

(25)

ni ,Pi

s.t.: Po (i) ≤

Preq o (i), ni

≤ nmax , Pi ≤ Pmax .

The reliability constraint in (25) could be rewritten as the minimum required success probability in communications as in (20). Now, we rewrite the reliability constraint, which was driven in (21), as follows: √ D0 1 − ni Bi Preq , (26) o (i) = √1 D1 + λa π + N γth Pi Ωα P i

where the auxiliary variables D0 and D1 are defined as: ( ) √ D0 = 0.5 πλa π exp − υî,2 , ∑ Pk γth 0.5 π 2 π D1 = λk υˆk,2 ( ) csc( ). k Ω 2 2 By simplifying (26), ni is derived as a function of Bi as follows: ⌈ ⌉ / √ D0 ) log(1 − ni = log( Bi Preq ) . o √1 D1 + λa π + N γth Pi Ωα Pi (27) Also, the constraint on ni is translated to a constraint on Pi as: √ γth D0 √ −D1 + D1 2 -4 NΩπ (λa π- nmax ) )2 Bi 1Preq o ∆ ( Pi ≥ Pmin = . D0 √ 2(λa π- nmax Bi req ) 1−

Po

Parameters

Value

Service area Pathloss Thermal noise power Distribution of devices Packet arrival of each device

20 × 20 Km2 x 133 + 38.3 log( 1000 ) −174 ( dBm/Hz PCP λi ×1e-6,200, Eq. (1) with σ=100) Poisson distributed with average reporting period (Ti ) of 300 s 100 ms 10 KHz, 100 KHz 1000 J, 10 mW, 0.1 J, 0.2 J 0.5 W, 1.5 W 1,1,0.5 Default: 21 dBm, 1, 5.5e-8 1, 0 c1 /2000, c1 /2270

Packet transmission time (τi ) Signal, system bandwidth E0 , Pc , Est , Ec Pr , Pa γth , |ϖ|, η Pi , n i , λ a ℓmax , Q c2 , c 3

Then, the optimization problem in (25) reduces to a simple search over Pmin ≤ Pi ≤ Pmax for minimization of: βî Ec + βî ni (ηPi + Pc )τi ,

(28)

in which ni has been found as a function of Pi in (27), βî has been found as a function of Ps (i) and ni in (17), and Ps (i) has been found as a function of Pi in (26). This operation control optimization problem is investigated numerically in the next section (Fig. 7). V. P ERFORMANCE E VALUATION In order to investigate usefulness of our findings in IoTnetwork planning and operation control, here we implement a MATLAB simulator for a heterogeneous IoT network. In our simulator, 2 types of IoT devices have been considered that differ in the distribution processes describing locations of their respective nodes, and communications’ parameters such as transmit power. Motivations for this setup are the coexistence of different IoT technologies over the public ISM spectrum, e.g. SigFox and LoRaWAN [7], and the coexistence of heterogeneous IoT services in grant-free enabled cellular networks [6]. For type i, the distribution process of locations is characterized by PCP(λi , υi , f (x)), where λi is the density of cluster points (in Km−2 ), υi = 200 is the average number of nodes in each cluster, and distribution of cluster nodes around the cluster center, i.e. f (x), is modeled by a normal distribution with standard deviation of 100 meters. The √ reliability constraint is described as ps (i, deg ), where deg = 1/(πλa ) is equivalent to the cell-edge communication distance in the case of grid deployment of APs. The packet arrival process at each node follows a PPP of rate T1i . The other parameters can be found in Table II. A. Validation of Derived Analytical Expressions First, we investigate tightness of the derived analytical expressions. Towards this end, we consider a coexistence scenario, comprising of devices belonging to type-1, which are of our interest, and type-2, which are interferes and their transmit power is 4 dB higher than type-1 devices. Fig. 4 represents probability of success in packet transmission for

p s (1,d eg): Reliability

p s (1, d): Probability of success

1

0.9

0.8 Simulation Analytic Impact of Type-1 (other clusters) Impact of Type-1 (same cluster) Impact of Type-2 (all) Impact of noise

0.7

0.6

acceptable region

0.5

0 20

System BW (KHz) 0

500

1000

1500


0.8

4000 0.4

0 10 -2

B1 = 8

0.2 0 10 -1

10 0

600

Cost ( × c 1 )

Battery lifetime ( × reporting period) Probability of collision (1-p s(1,d eg))

Probability of collision in transmission

1

Network cost ( × 0.1 c 1 )

0.6

2000

-1

(a) Specifying the region in which reliability constraint is satisfied.

Fig. 4: Validation of analytical and simulation results. Device distribution: K=2, λ1 =0.19, λ2 =3.8, υ1 =1200, υ2 =30, P1 =21 dBm, and P2 =25 dBm.

6000

10

0

2000

d: Distance from the AP (meters)

8000

10 0

10

0.5

Cost, Lifetime

1

400

minimum cost in the acceptable region

acceptable region

200 0 20 10 0

10

Bandwidth (KHz)

10 -1

0


(b) Finding minimum-cost investment strategy in the acceptable region.

Fig. 6: Optimized deployment strategy (λ1 =4.6, required ps (1, deg )=0.7).

λa : Density of APs ( × Km 2 )

Fig. 5: Tradeoff between network cost, reliability, and battery lifetime (K = 1, λ1 =6.4). For λa ≤ 0.4 Km2 , the maximum number of retransmissions, i.e. B1 = 8, is met.

network design objectives are coupled such that improvement in one objective deteriorates the other(s). C. Optimized Deployment Strategies

type-1 as a function of distance from the AP. One sees that the analytical model matches well with the simulation results. We have further depicted the contributions of noise, interferences from the same and other clusters of type-1 devices, as well as interference from type-2 devices. Regarding the fact that transmit power of type-2 devices is 4 dB higher than type-1 devices, it is clear that interference from type-2 traffic (plusmarked curve) is the most limiting factor. B. Analysis of Performance Tradeoffs Fig. 5 represents the tardeoff between cost of the access network (left y-axis), durability of communications (right yaxis), and reliability of communications. The x-axis represents the density of APs. We observe that by increase in density of APs, probability of collision in packet transmission decreases which results in less required number of retransmissions. Then, it is clear that battery lifetime can be significantly saved by provisioning more resources for IoT traffic, which on the other hand, increases network costs. We observe that the

Fig. 6 represents the resource provisioning problem which has been partially touched in subsection IV-B. First, Fig. 6a illustrates reliability of communications for different AP density-system bandwidth configurations. One sees by increase in both AP density and system bandwidth, probability of success in communications increases. Taking 0.7 as the required success probability, the density-bandwidth region in which reliability constraint is satisfied has been characterized in Fig. 6a. In Fig. 6b, network cost has been depicted as a function of provisioned network resources, i.e. AP density and system bandwidth. Now, it is straightforward to search over the acceptable region, mapped from Fig. 6a, to find the resource provisioning strategy which minimizes the network costs, as depicted in Fig. 6b. 1) Optimized Operation Control: Fig. 7 represents the interplay among success probability, battery lifetime, number of replica transmissions (ni ), and device transmit power (Pi ). The x-axis in Fig. 7a and Fig. 7b represents P1 for circlemarked curves, and n1 for cross-marked curves. In these


2500


2000 1500 optimized operation point for type-1

1000 500 0 1

2

3

n 1 : Nr of replicas,

4

5

6

7

8

9

20

Required AP density (× Km-2 ) Required BW (× 10 KHz) Required n 1

15

Required P 1

10 0

10 10

-1

5 10 -2 0.3

0 0.4

0.5

0.6

0.7

0.8

0.9

1

p s (1,d eg): Required reliability

10

P1 : Transmit power ( × 126 mW)

Fig. 8: Scalability analysis (K = 1, λ1 =3.2).

(a) Battery lifetime for type-1 devices 0.7

10 1

Required BW, n 1 , P1

Required AP density ( × Km -2 )

Type-1 battery lifetime ( × reporting period)

3000

0.5

0.4

s

eg

1

s

eg

1

s

eg

1

s

eg

1

s

eg

1

s

eg

1

s

eg

1

s

eg

1

Probability of success (Type 2)

Probability of success (Type 1)

replica transmissions, i.e. n1 , increases both battery lifetime and reliability for n1 ∈ {1, 2}, offers a tradeoff between battery lifetime and reliability for n1 ∈ {3, 4}, and decreases 0.5 both reliability and battery lifetime for n ≥ 5. These results 0.3 0.4 confirm importance of the derived results in this work, as they p (1,d ) vs. P ; Sc1 p (1,d ) vs. n ; Sc1 shed light to the operation point after which, it is not feasible 0.3 0.2 p (1,d ) vs. P ; Sc2 to trade battery lifetime in hope of reliability. p (1,d ) vs. n ; Sc2 0.2 2) Scalability Analysis: The analytical model of reliability p (2,d ) vs. P ; Sc1 0.1 p (2,d ) vs. n ; Sc1 has been found in (15) as a function of: i) transmit power, ii) 0.1 p (2,d ) vs. P ; Sc2 number of replica transmissions, iii) density of APs, and iv) p (2,d ) vs. n ; Sc2 0 0 bandwidth of communications. Fig. 8 represents the rate at 1 2 3 4 5 6 7 8 9 10 which, the amount of provisioned resources at the networkn 1 : Nr of replicas, P1 : Transmit power ( × 126 mW) side, or energy resources at the device-side, should be scaled (b) Probability of success in transmission for type-1 and type-2 devices to comply with the increase in the level of required reliability. Fig. 7: Optimized operation control (K = 2, λ2 =2.4, λ1 =2.4 It is clear that transmit power of devices could be increased in Sc1 and λ1 =1.2 in Sc2). In circle-marked curves, n1 = 1 up to a certain level in order to combat noise. However, and P1 is varying. In plus-marked curves, P1 = 126 mW and beyond a certain point, increase in the transmit power cannot increase the success probability because it cannot compensate n1 is varying. the impact of interference. On the other hand, one sees that increase in the number of replicas per packet could also be figures, Sc1 and Sc2 differ in density of type-2 devices, which leveraged to increase reliability of communications. However, is 2.4 in Sc1, and 1.2 in Sc2. One observes in Fig. 7a that there is a saturation point in scenarios with higher densities battery lifetime is a quasi-concave function of both Pi and of nodes, where increasing number of replicas increases ni . Furthermore, in Sc1, where density of nodes is higher traffic load significantly, and may even reduce reliability of than Sc2, battery lifetime decreases significantly by increase communications. Example of such event was observed in Fig. in the number of replica transmissions. In both scenarios, we 7b for n1 ≥ 5. Finally, the rate of increase in reliability see that the energy-optimized operation strategy for type-1 of communications by increasing the number of APs, which devices is to send 2 replicas per data packet to maximize their reduces the communications’ distance, and increasing the battery lifetimes. Fig. 7b represents the success probability bandwidth, which decreases the collision probability, could for type-1 and type-2 traffic as a function of n1 and P1 . We be observed in Fig. 8. observe that success probability for type-1 devices increases VI. C ONCLUSION to a point beyond which, the resulting interference from extra A tractable analytical model of reliability in large-scale transmitted replicas starts deteriorating the performance. On the other hand, increase in the transmit power for type-1 heterogeneous IoT networks has been presented as a function devices, increases the success probability for this type and of IoT traffic intensity and access network’s resources. This severely decreases the performance of type-2 devices. It is also model has been employed to analyze the impacts of resource worthy to note that in Fig. 7b, success probability increases in provisioning at the network-side and operation control at the n1 till n1 = 4, however, from the battery lifetime analysis in device-side on reliability and battery lifetime of IoT devices. Fig. 7a, it is evident that battery lifetime decreases in n1 for The derived expressions illustrate the rate of increase in n1 ≥ 3. To conclude, we see that increase in the number of reliability and battery lifetime achieved by increasing the 0.6

bandwidth of communications and number of APs. Our analyses indicated that depending on the operating point, increasing transmit power and number of replica transmissions may increase both reliability and battery lifetime, offer a tradeoff between them, or decrease both of them, i.e. it can turn from being a friend to foe. Then, we have developed reliabilityconstrained, cost- and battery-optimized resource provisioning and operation control strategies for the access network and IoT devices respectively. Finally, we have presented the scalability analysis to figure out the bounds up to which, increasing the provisioned resources at the network-side, or increasing energy consumption of IoT devices per packet transfer, can compensate the impact of increase in number of devices or their required QoS. The tightness and tractability of the derived expressions promote use of them in IoT network planning and operation control. A PPENDIX A P ROOF OF T HEOREM 3.2 When m = 1, (9) reduces to: ) ∫ (∑ [ ] ps (i, z) = exp −λk 1 − exp(−ˆ υk,j U(y)) j,k

( N γth m ) × × exp − ΩPi g(z) ∫

where U(y)=

R2

∫

R2

R2

( ∑ ) exp − j υî,j U(y) f(y)dy,

g(x − y) f(x)dx, g(x − y) + G(z)

G(z) = ΩPi g(z)/[γth Qj Pk ]. Then, using Jensen’s inequality, we have: ∫ (∑ ) ps (i, z) ≈PN exp −λk υˆk,j U(y) dy k,j 2 R ∫ ∑ × exp(− υî,j U(y)f(y)dy), j R2 ∫ ∫ (∑ ) g(y) ≈PN exp -λk υˆk,j dy f(v)dv 2 g(y)+G(z) 2 R R k,j ∫ ∫ (∑ ) g(x-y) × exp - υî,j f(x)dxf(y)dy g(x-y)+G(z) j R2 R2 ∫ ( ∑ ) g(y) ≈PN exp − λk υˆk,j dy k,j R2 g(y) + G(z) ∫ ∫ (∑ ) g(v) f(v+y)dvf(y)dy , × exp - υî,j g(v)+G(z) j R2 R2

in which, v is an auxiliary variable equal to x − y. Using the isotropic property of f(x), i.e. f(y) = f(−y), and another change of variables, we have: ∫ ) ( ∑ g(y) dy ps (i, z) ≈PN exp − λk υˆk,j k,j R2 g(y) + G(z) ∫ (∑ ) g(v) × exp υî,j f∗ (v)dv , j v∈R2 g(v)+G(z)

where: ∫

f(v-y)f(y)dy = conv(f(v), f(v)) = f∗ (v). ∆

y∈R2

By using H(·), as defined in (11), we have: ∑ ∑ [ ( γth Qj Pk )] ) λk υˆk,j H(z, 1, ps (i, z) ≈ PN exp − ΩPi (

× exp −

∑

j∈{1,2} k∈K

j∈{1,2}

υî,j H(z, f∗ (x),

Qj γth ) ) . Ω

A PPENDIX B P ROOF OF C OROLLARY 3.4 For f(x) given in (1), f∗ (x) is derived as: f∗ (x) = exp(−||x − x0 ||2 /(4σ 2 ))/4πσ 2 . Then, we have: ( 2π H z, f∗ (x), ξ) = 4πσ 2

∫

∞

r=0

exp(−r2 /(4σ 2 )) dr r4 1 + ξ||z|| 4

∫ ξ||z||4 ∞ exp(−r2 /(4σ 2 )) √ = rdr 2 2 2 4 2σ r=0 [ ξ||z|| ] + r ∫ ∞ 4 exp(−a1 x) (e) ξ||z|| = dx 4σ 2 x=0 (a2 )2 + x2 √ √ √ 2[ 2 ξ||z|| ξ||z|| ξ||z||2 (f) ci( ) sin( ) = 2 2 4σ 4σ 4σ√2 √ 2 ξ||z|| ξ||z||2 ] − si( ) cos( ) , 4σ 2 4σ 2

where in (e), we have used change of variables x = r2 , (f) has been derived using table of integrals in [28, Eq. 3.352], and a1 = 4σ1 2 and a2 = ξ||z||4 are auxiliary variables.

A PPENDIX C P ROOF OF T HEOREM 3.5 Using theorem 3.2, corollary 3.4, and corollary III-A2 we have: ( r4 N γth ) ( ) ps (i, r)dPdℓ (r) ≈ exp exp -ˆ υi,2 × ΩPi α √ ( ∑ γth Qj Pk π 2 π ) csc( ) × exp λk υˆk,j r2 ΩPi 2 2 j,k

2(λa πr2 )ℓ dr, r(ℓ − 1)! ≈2rX0 r2(ℓ−1) exp(-[X1 r4 +X2 r2 ])dr, (29) exp(-λa πr2 )

where X0 :X4 have been defined in theorem 3.5. Inserting (29) in (14), Ps (i) is derived as: Ps (i) ≈ ≈ ≈ ≈

ℓ∏ max ℓ=1 ℓ∏ max ℓ=1 ℓ∏ max ℓ=1 ℓ∏ max ℓ=1 ℓ∏ max

∫

∞

1-

X0 r2(ℓ−1) exp(-[X1 r4 + X2 r2 ])2rdr,

0

∫

∞

1-X0

x(ℓ−1) exp(−[X1 (x +

0

1-X0 exp( 1- √

X2 2 ) 4X1 2

X0 X1

ℓ−1

exp(

∫

∞ X2 2 2X1

(y-

X2 2 ) 4X1 2

∫

X2 2 X2 2 ) )])dx, 2X1 4X1 2

X2 (ℓ−1) ) exp(-X1 y 2 )dy, 2X1 ∞

X2 2 2X1

2 X2 (z- √ )(ℓ−1) e-z dz, 2 X1

X3 2 )G(X3 , ℓ), (30) X1 X1 ℓ−1 ℓ=1 √ X2 x = r2 , y = x + 2X , and z = X1 y are auxiliary variables. 1 ≈

1- √

X0

exp(

R EFERENCES [1] C. Mavromoustakis, G. Mastorakis, and J. M. Batalla, Internet of Things (IoT) in 5G mobile technologies. Springer, 2016. [2] W. Yang et al., “Narrowband wireless access for low-power massive internet of things: A bandwidth perspective,” IEEE Wireless Commun., vol. 24, no. 3, pp. 138–145, 2017. [3] A. Laya, L. Alonso, and J. Alonso-Zarate, “Is the Random Access Channel of LTE and LTE-A Suitable for M2M Communications? A Survey of Alternatives,” IEEE Communications Surveys Tutorials, vol. 16, no. 1, pp. 4–16, Jan. 2014. [4] A. Azari et al., “Network lifetime maximization for cellular-based M2M networks,” IEEE Access, vol. 5, pp. 18 927–18 940, Sept. 2017. [5] 3GPP TR 45.820, “Technical Specification Group GSM/EDGE Radio Access Network; Cellular system support for ultra-low complexity and low throughput Internet of Things (CIoT),” Tech. Rep., Aug 2015. [6] 3GPP TSG RAN WG1 , “Overall solutions for UL grant free transmission ,” Tech. Rep., June 2017. [7] M. Lauridsen, B. Vejlgaard, I. Z. Kovacs, H. Nguyen, and P. Mogensen, “Interference measurements in the european 868 MHz ISM band with focus on LoRa and SigFox,” in IEEE Wireless Communications and Networking Conference, 2017. [8] M. Masoudi et al., “Grant-free Radio Access IoT Networks: Scalability Analysis in Coexistence Scenarios,” in IEEE Global Communications Conference, 2018. [9] H. Tabassum et al., “Non-orthogonal multiple access (NOMA) in cellular uplink and downlink: Challenges and enabling techniques,” arXiv preprint arXiv:1608.05783, 2016. [10] Z. Ding et al., “A survey on non-orthogonal multiple access for 5G networks,” IEEE Journal on Selected Areas in Communications, vol. 35, no. 10, pp. 2181–2195, 2017. [11] R. Karaki, A. Mukherjee, and J. F. Cheng, “Performance of autonomous uplink transmissions in unlicensed spectrum LTE,” in IEEE Global Communications Conference, Dec 2017, pp. 1–6. [12] P. Popovski, et al., “Ultra-Reliable Low-Latency Communication (URLLC): Principles and Building Blocks,” arXiv preprint arXiv:1708.07862, 2017. [13] A. Azari, P. Popovski, G. Miao, and C. Stefanovic, “Grant-Free Radio Access for Short-Packet Communications over 5G Networks ,” in IEEE Global Communications Conference, 2017. [14] G. Miao et al., “ E2MAC: Energy Efficient Medium Access for Massive M2M Communications,” IEEE Trans. on commun., vol. 64, no. 11, pp. 4720–4735, 2016. [15] Z. Zhang, X. Wang, Y. Zhang, and Y. Chen, “Grant-free rateless multiple access: A novel massive access scheme for internet of things,” IEEE Communications Letters, vol. 20, no. 10, pp. 2019–2022, 2016.

[16] A. T. Abebe and C. G. Kang, “Comprehensive grant-free random access for massive & low latency communication,” in IEEE International Conference on Communications, 2017. [17] L. Liu, E. G. Larsson, W. Yu, P. Popovski, C. Stefanovic, and E. De Carvalho, “Sparse signal processing for grant-free massive IoT connectivity,” arXiv preprint arXiv:1804.03475, 2018. [18] A. Aminjavaheri, A. RezazadehReyhani, R. Khalona, H. Moradi, and B. Farhang-Boroujeny, “Underlay control signaling for ultra-reliable low-latency IoT communications,” arXiv preprint arXiv:1711.02248, 2017. [19] Z. Li et al., “2D time-frequency interference modelling using stochastic geometry for performance evaluation in low-power wide-area networks,” arXiv preprint arXiv:1606.04791, 2016. [20] Y. Zhang, K. Peng, Z. Chen, and J. Song, “SIC vs. JD: Uplink NOMA techniques for M2M random access,” in IEEE International Conference on Communications, May 2017, pp. 1–6. [21] R. Abbas, M. Shirvanimoghaddam, Y. Li, and B. Vucetic, “Grant-Free Massive NOMA: Outage Probability and Throughput,” arXiv preprint arXiv:1707.07401, 2017. [22] V. Suryaprakash et al., “On the modeling and analysis of heterogeneous radio access networks using a poisson cluster process,” IEEE Trans. on Wireless Commun., vol. 14, no. 2, pp. 1035–1047, 2015. [23] C. Saha, M. Afshang, and H. S. Dhillon, “Poisson cluster process: Bridging the gap between PPP and 3GPP HetNet models,” in IEEE Information Theory and Applications Workshop , 2017, 2017. [24] C. Wang, “Modeling and analysis of synchronization schemes for the TDMA based satellite communication system,” 2012. [25] R. K. Ganti and M. Haenggi, “Interference and outage in clustered wireless ad hoc networks,” IEEE Trans. on Inf. Theory, vol. 55, no. 9, pp. 4067–4086, 2009. [26] J. Moller and R. P. Waagepetersen, Statistical inference and simulation for spatial point processes. CRC Press, 2003. [27] F. Baccelli, B. Blaszczyszyn, and P. Muhlethaler, “An aloha protocol for multihop mobile wireless networks,” IEEE Trans. on Inf. Theory, vol. 52, no. 2, pp. 421–436, 2006. [28] I. S. Gradshteyn and I. M. Ryzhik, Table of integrals, series, and products. Academic press, 2014. [29] M. Haenggi, “On distances in uniformly random networks,” IEEE Trans. Inf. Theory, vol. 51, no. 10, pp. 3584–3586, 2005. [30] 3GPP TSG GERAN 46, “USF capacity evaluation for MTC,” Tech. Rep., May 2010. [31] S. Tombaz, A. Vastberg, and J. Zander, “Energy- and cost-efficient ultrahigh-capacity wireless access,” IEEE Wireless Commun., vol. 18, no. 5, pp. 18–24, Oct. 2011.

Paper D: Grant-Free Radio Access for Cellular IoT A. Azari, P. Popovski, C. Stefanovic, and C. Cavdar Submitted to IEEE Transactions on Wireless Communications, 2018. The conference version of this paper has been published in IEEE Globecom 2017 [63].

141

Grant-Free Radio Access for Cellular IoT ˇ Amin Azari1 , Petar Popovski2 , Cedomir Stefanović2 , Cicek Cavdar1 1 KTH Royal Institute of Technology; 2 Aalborg University Email:{aazari,cavdar}@kth.se, {cs,petarp}@es.aau.dk

Abstract—Radio resource management plays a vital role in experienced delay, reliability, and energy consumption of Internet of things (IoT) communications. Towards enabling massive IoT connectivity in an energy efficient way, this article focuses on leveraging grant-free radio access, and outlines several key transmission/reception techniques that promote the performance of the considered radio access scheme. The proposed solutions aim at providing a low energy consumption and communication delay profile by relaxing the signaling burden at the cost of sending several packet copies at the transmitter-side and more complex signal processing at the receiver-side. Specifically, the timing and frequency offsets, as well as transmission of multiple replicas of the same packet, are exploited as three sources of diversities to trigger successive interference cancellation at the receiver-side. The system performance is investigated by using a model based on stochastic geometry which leads to closedform expressions for the key performance indicators, such as reliability and battery lifetime. The derived expressions are then employed in optimizing the number of replica transmissions. The evaluation results indicate that the proposed data transmission and reception schemes can significantly prolong battery lifetime of IoT devices by removing the need for connection establishment and reducing the number of retransmissions. The obtained results also indicate existence of traffic-load regions, where grant-free radio access outperforms the grant-based one, which is used in LTE and NB-IoT systems. Then, the results pave the way for enabling intelligent grant-based/free operation mode switching in 5G networks. Index Terms—5G, Asynchronous access, Battery lifetime, Grant-free access, Radio access management.

I. I NTRODUCTION Internet of things (IoT), including massive machine-type communications (mMTC) and ultra-reliable low-latency communications (URLLC), is one of the major drivers of 5G, expected to be integrated in cellular networks by 2020 [1]. The characteristics of mMTC include: extremely high density of nodes, short payload sizes, and limited computation and communication capabilities. Moreover, devices in most of IoT applications are battery powered, necessitating long battery lifetimes [2]. Thus, in contrast to the existing cellular traffic, the mMTC traffic requires support for massive concurrent access with high energy efficiency. The continuing growth of IoT market has encouraged mobile network operators (MNOs) to investigate evolutionary and revolutionary radio access technologies for accommodating the IoT traffic in cellular networks [3]. Evolutionary schemes aim at enhancing access procedures of existing LTE networks [4]. On the other hand, revolutionary solutions aim at fundamental revision of the cellular access procedures in order to provide ultra-low delay and/or energy consumption profiles [5]. In the existing cellular networks, devices first contend over random access channel (RACH) to reserve radio resources,

and then send data over the granted resources. As in most IoT applications the actual data to be transmitted is in order of bits, this connectivity procedure incurs unnecessary energy consumption in overhead signaling and idle listening to the base station (BS) [2]. As a result, battery lifetime of connected devices will be much less than the 5G requirement, which requires more than 10 years of battery lifetime for IoT devices [6]. Moreover, the massive number of potential connections that should be sustained concurrently is likely to cause congestion of the access network [7]. Capacity limits of RACH in serving IoT traffic and a survey of improved solutions can be found in [8], where among the proposed solutions, access class barring and capillary networking have been adopted in standardization [9]. The development of LTE for low-cost mMTC has been initiated in release 12, and continued in release 13 with introduction of LTE category M (LTE-M) and narrow-band IoT (NB-IoT) [10]. By reducing the maximum system bandwidth to 1.4 MHz (as opposed to 20 MHz for LTE) and data rate to 1 Mbps, LTE-M has targeted IoT applications with small to medium size bidirectional data transfer requirements. In NB-IoT, bandwidth for communications and data transmission rates have been decreased to 200 KHz and 200 Kbps, respectively, in order to improve the link budget, and hence, reduce the required energy for data transmission. However, LTE-M and NB-IoT still suffer from the required overhead signaling for synchronization, resource reservation, and listening to the control channel [11]. Power consumption in data transmission state is usually higher than the other operation states. However, regarding the fact that a device usually spends a longer time in listening to the BS than in the transmission state, energy consumption in the non-transmission state can be much higher than the transmission state [2]. The limitations of grant-based radio access management for realizing ultra-low latency [12] and ultra-long battery lifetime [13, 14] have motivated researchers to look into grant-free radio access for IoT communications. A. Literature Study A potential solution to alleviate energy consumption in synchronization, connection establishment, and resource reservation is to enable a non-orthogonal radio access scheme for IoT communications [13]. This is due to the fact that the overhead signaling is mainly used for keeping orthogonality in code/time/frequency domains among simultaneously active users. Non-orthogonal radio access has become a hot topic in recent years as part of the efforts which are shaping the new radio (NR) for 5G [15, 16]. In literature, non-orthogonal radio access has been leveraged to increase the network

throughput [17], the communication reliability [12], and to reduce the service delay [17]. A comprehensive study on major non-orthogonal radio access schemes that have been nominated for 5G is presented in [18]. In [18], a categorization of radio access schemes is made based on the feature which enables multiple access: a) multiple access using codebooks, e.g. power-domain codebooks in Sparse Code Multiple Access (SCMA), which resolves collisions based on the received power levels; b) multiple access using sequences, e.g. Multi User Shared Access (MUSA), in which complex spreading sequences are shared amongst devices and used in non-orthogonal access; and c) multiple access using interleaver/scrambler like Resource Spread Multiple Access (RSMA), which has been proposed as a candidate for grantfree access in future LTE releases [19]. Further, the common assumption of saturated-buffers used for measuring capacity of communication channels has been revisited in mMTC, and many-user multiple access channel has been introduced [20]. In [21], a random access code for multiple access channel has been introduced, and its achievable regions of spectral and energy efficiency have been investigated. In [22], an advanced compressed sensing technique for massive grant-free radio access connectivity has been proposed, and its performance in efficient detection of simultaneously active devices has been investigated. The capacity of wireless channels, shared by many users with finite block-lengths, has been partially investigated in [23], and is still an open problem, especially when it comes to non-orthogonal IoT communications. To keep the device energy consumption, complexity, and cost as low as possible, narrowband/ultra-narrowband approaches with frequency hopping for interference resilience have been implemented in some IoT technologies, such as SigFox and LoRaWAN [24]. SigFox, introduced in 2009, is the first widely announced IoT technology which focuses on providing low-cost, low-power communications over the ISM spectrum. SigFox handles IoT communications over ultranarrowband signals, i.e. 100 Hz bandwidth, with semi-random frequency hopping over 600 KHz bandwidth [24]. The LoRa wide area network (LoRaWAN), introduced in 2015, handles IoT communications over narrowband signals, i.e. 125 KHz bandwidth, with random frequency hopping over 3 channels in the ISM spectrum [24]. Furthermore, LoRaWAN leverages the chirp spread spectrum (CSS) modulation to maintain immunity against potential interference over the ISM spectrum. Both SigFox and LoRaWAN technologies implement the grant-free radio access for data transmissions, i.e., once a packet is triggered at the device, it is transmitted without any handshaking, resource reservation, or authentication process. This new grant-free radio access paradigm in IoT connectivity, which can be called time-frequency asynchronous ALOHA, enables connectivity for extremely low-complexity transmitters equipped with cheap oscillators [25, 26]. In this case, transmitted packets from different devices are vulnerable to partial time-frequency overlapping as transmissions happen over non-orthogonal radio resources. In order to improve the interference resilience, each device may transmit several replicas per data packet. This diversity can be exploited by the receiver-side through a) decoding of packets by combining

their (partially) interference-free replicas; and b) removal of replicas of decoded packet through interference cancellation, enabling potential decoding of new packets. Such successive interference cancellation (SIC)-based receivers for asynchronous ALOHA systems have been investigated in [25, 27]. Specifically, the solution in [27] exploits timing offsets and replica “diversity”, where the former is due to the fact that different devices start data transmission in different time instances and the latter is due to the fact that devices select replica positions independently. The proposed receiver in [27] requires complete knowledge of replicas’ positions. The authors of [27] suggest that finding position of replicas is possible by either exploiting correlation or robust header encoding. However, the performance of the proposed solutions degrades by increase in the traffic load, and thus, the amount of interference. Furthermore, using blind correlation in search of replicas significantly increases complexity of receiver, and hence, increases the required time to detect and decode the packets. Recall the fact that due to the cheap price of many IoT devices, they will be equipped with cheap oscillators. Thus, carrier frequency offset (CFO) in IoT communications will be inevitable. In [28], the frequency and time offset of devices have been used for simultaneous detection of multi RFIDtags. Motivated by this work, here we propose to exploit CFO as another source of diversity for contention resolution, and develop a SIC-based receiver for grant-free access, which can support decoding of a multitude of low-complexity IoT devices. Furthermore, we analytically analyze the performance impact of replica transmission on reliability of communications. This analysis is subsequently employed in design of a replica transmission control scheme. The simulation results confirm that the grant-free radio access scheme, augmented by the proposed replica transmission control and SIC-based receiver, can significantly shorten the access delay and prolong battery lifetime of IoT devices.

B. Contributions In [13], our preliminary results on leveraging CFO for contention resolution were presented, assuming channel inversion power control for IoT devices and a constant number of replica transmissions per packet. This paper presents substantial improvements over [13] by removing those assumptions, presenting an optimized replica-transmission control scheme, exploiting stochastic geometry for deriving analytical expressions, and validating analytical expressions by simulations. The main contributions of the paper include: •

• •

Development of a replica-based transmission scheme for IoT devices in grant-free radio access, which aims at maximizing both reliability and battery lifetime. Development of a collision resolution scheme utilizing time/frequency asynchronism. Development of closed-form statistics of two-dimensional (i.e., time-frequency) interference using stochastic geometry tools, and derive closed-form expressions for the delay, reliability, and battery lifetime.

The remainder of the text is structured as follows. The system model is described in the next section. Section III presents the proposed receiver design. Performance indicators and tradeoffs are analytically modeled in Section IV. The replica-transmission control scheme is presented in Section V. Simulation results are presented in Section VI. The concluding remarks are given in Section VII.

Frequency !

!"

Max drift

•

Identification of operating regions in terms of traffic load in which grant-free radio access outperforms the grantbased access. Presentation of performance tradeoffs in asynchronous grant-free radio access protocol design.

#$" Carrier frequency

Max drift

•

Time

Fig. 1: Received virtual frames at the BS. Replicas with the same color refer to a single data packet.

III. T HE P ROPOSED C ONTENTION R ESOLUTION S CHEME II. S YSTEM M ODEL We consider a single cell serving a multitude of asynchronous IoT devices over a shared spectrum of W Hz. When a device has a packet to transmit, it assumes a virtual frame (VF) consisting of M parts/slots, each with duration of Tp , where Tp is the time-duration of packet transmission of the device1 . Then, a packet is sent immediately at the first slot of the VF, and the N − 1 replicas of the packet are sent in N − 1 randomly selected slots out of M − 1 remaining slots, as depicted in Fig. 1. Beside data, each packet also includes a preamble of length Nzc . The cardinality of the set of available preambles, denoted by |S|, as well as the length of preambles, are two system design parameters. Nzc limits the |S|, and increases both accuracy of timing synchronization for finding the position of packets, and the probability of collision [29]. The latter is due to the fact that increase in N increases the packet size, and hence, the packet transmission time. The choice of N , i.e. number of replica transmissions per packet, determines the data transmission scheme. In general, N is a number chosen in {1, · · · , M }. If N is chosen randomly, or is fixed for all devices, we refer to it as grant-free radio access scheme without replicatransmission control. Otherwise, it is called grant-free scheme with replica-transmission control, to be discussed in Section V. A quasi-static fading channel model is assumed, which means channel gain is constant over a packet. As low-complexity sensors with cheap oscillators will be an essential part of future networks, CFO will be inevitable. Indeed, it is expected that in ultra-narrowband systems, the maximum level of CFO Fm is, by definition, expected to be several orders larger than the signal bandwidth w [30, section 3.2.2]. We denote the actual carrier frequency that the i-th transmitter uses for data transmission, and its drift from the intended carrier frequency as fi and ∆fi = fi −f , respectively. While frequency drifts in different wakeup epochs of operations of a device are expected to be independent, ∆fi is, in essence, constant during one virtual frame [31]. The timing offsets, CFOs, and number of transmitted replicas are independent among devices, and the transmissions of different devices are uncoordinated. Thus, overlapping of packet replicas sent by devices’ is inevitable, as depicted in Fig. 1. 1 The boundaries of slots and the slots’ lengths, i.e. T , are asynchronous p among different nodes.

At the receiver end, we sample the arriving signal at rate Fs , searching for the (potentially collided) replicas, see Fig. 2. The choice of Fs incurs a system performance tradeoff, because a higher sampling rate increases the collision resolution capability by increasing the receiver cost. Once the presence of the signal is detected, to which we refer to as an event, the receiver jointly processes samples organized in a time frame with length of Tf . Tf is a design parameter related to the tolerable delay in data processing. The samples in the time frame are processed using a periodogram module, which aims at finding periodic components in the signal, and returns the detected carrier frequencies, as depicted in Fig. 3. Denote the number of detected carrier frequencies (which are determined by CFOs of the contending devices) as K (see in Fig. 2). Then, the samples from the time frame are demodulated and correlated with the preamble2 K times. Each of the correlations may return peaks. Denote by Yj (n) the output of correlation of the preamble with Xi,j , where Xi,j is the demodulated version of Xi (n) by f +∆fj . In Yj (n), the respective peak of j-th CFO is located in the correct timing offset (i.e. at the starting point of the packet replica), while the respective peaks of other CFOs are shifted due to mismatch between demodulated frequency and their frequencies, as depicted in Fig. 4. The details of Fig. 4 and the reasons behind these time-shifts will be further discussed in Section III-B; here we note that the level of shifted peaks can be as high as the original peak, or even higher, if the length of preamble sequence is short, which may be the case in IoT applications with short packet lengths. The task of the peak detection module in Fig. 2 is to report the set of detected peaks. The task of decision making module is to detect and remove shifts of time offsets of already detected peaks and to report the set of K found time offsets (respective to the K found CFOs). Then, the respective demodulated sequence of each carrier frequency, e.g. Xi,j (n) for fi , is truncated from τi,j to the length of a packet and is fed to the SIC module along with its carrier frequency and time offset, i.e. the (Zi,j (n), fi,j , τi,j ) tuple for j ∈ {1, · · · , K}, is fed to the SIC module. A. The SIC Module The SIC module continuously receives and saves demodulated sequences related to processed events, and their respective carrier frequencies and time offsets. Then, it tries 2 For

|S| >1, we take correlation with the set of preambles.

(!)

" (#)

Sampling Windowing

Peak Detection

Periodogram

DM(*)

DM(* 0 1*% )

DM(* 0 1*' )

DM(* 0 1*$ )

/*$ /*% /*'

&$ (#)

",' (#)

",% (#)

",$ (#)

Correlation &' (#)

Peaks Detection

Preamble(s)

(-",$ # , *",$ , +$,. )

(*",$ , +",$ )

Cut sequence

SIC Module

Decision Making

(*",', , +",' )

(-",' # , *",' , +$,' )

Fig. 2: The proposed receiver design. i is the event index. DM(f ) represents demodulation with frequency f . Periodogram Power Spectral Density Estimate

Power/frequency (dB/Hz)

20 0

∆ f1 =-150 Hz

∆ f3 =100 Hz

-20

∆ f2 =-50 Hz

-40 20

20

0

0

-20

-20

-60

-80 -40

-40 0.1

0.2

0.3

19.7

-100 2

4

6

8

10

12

19.8 14

B. Processing of Detected Peaks

19.9 16

18

Frequency (kHz)

Fig. 3: Leveraging CFO for collision resolution. This figure represents the output of periodigram where three frequency components remain after demodulation with the carrier frequency.

1

Y1 (n)

of the other replicas become known, and hence, these are removed. If the packet replica cannot be decoded correctly, the SIC module tries to find its replicas by performing search in the set of previously processed and stored events containing the same carrier CFO, i.e. CFO is used as the signature of VFs. If the other replicas are also in collision, the SIC module can combine them. Thanks to the derived set of CFOs and time offsets, we have the time/frequency map of collisions, and hence, it is possible to figure out the level of interference that each replica is suffering from and use selection combining (SC) or equal gain combining (EGC) in order to improve the performance. The former consists of merging successfully received parts of replicas together to construct the original packet, while the latter combines replicas with equal gain. After combining detected replicas that belong to a packet, the SIC module tries again to decode the packet. If decoding succeeds, the receiver removes all replicas of the decoded packet, which lowers the level of interference for the other demodulated sequences (i.e., processed events) and provides for easier decoding of other packets. If decoding fails, the demodulated sequence is stored for further processing, while the receiver shifts the detection zone in time and tries to decode newly arrived events. In case that the subsequent decoding trials lower the level of interference in the previously stored collisions, new decoding attempts are made.

∆f 1 =-150 Hz time offset=1000 indices T 1 ={1000,1180,1200,

0.5

We first elaborate on the shift of peak location in taking correlation of preambles with the demodulated signals, when CFO is inevitable. If we take cross correlation of a preamble sequence, i.e. P (n), n ∈ {0, · · · , Nzc -1}, with itself, the result is a sequence of length 2Nzc − 1, i.e. P (m), m ∈ {1, · · · , 2Nzc − 1}, with a peak at Nzc + 1. Now, assume P (n) has been transmitted with carrier frequency f , and has been demodulated with f + ∆f , i.e. the CFO is ∆f . Denote by P˜ (n) the demodulated signal, i.e., P˜ (n) = P (n)ej2π∆f nTb , ∀n ∈ {0, Nzc − 1},

1380,1640}

Y2 (n)

0 1

∆f 2 =-50 Hz time offset=1100 indices T 2 ={920,1100,1300}

0.5

Y3 (n)

0 1

∆f 3 =100 Hz time offset=1160 indices T 3 ={780,960,1160}

0.5 0 0

500

1000

1500

2000

2500

3000

n: time index

Fig. 4: Shift in position of peaks in Y1 (n):Y3 (n) for Nzc = 45, Tb = 1 ms, and Fs Tb =20. y−axis represents outputs of correlation of Xi,1 (n):Xi,3 (n) with the preamble, as depicted in Fig. 2.

to decode each sequence. If the sequence, i.e. the supposedly contained packet replica, is decoded successfully, the locations

where Tb is the bit duration. Taking cross correlation of P (n) with P˜ (n), the peak location moves periodically between −⌊Nzc /2⌋ and ⌊Nzc /2⌋, as discussed in [32]. Given Tb and Nzc as characteristics of the system, the level of shift in the position of peak, denoted by Q, can be derived as a function of ∆f , as depicted in Fig. 5. The decision making module (in Fig. 2) decides which subset of detected peaks represents the correct time offsets of the replicas. In this module, a successive peak detection (SPD) function is used. Denote the set of received peaks from the peak detection module as {T1 , · · · , TK }, where Tj is the subset of detected peaks in Yj (n), and Yj (n) is the result of correlation of Xj (n) with the preamble, as in Fig. 2. The SPD function searches over Tj :s and makes a map of peaks and their shifted positions. For example, given pj as a candidate peak position3 in Tj , SPD checks Tj+k to see if it contains a peak 3 p represents time offset of a peak w.r.t. the reference time in processing j of the respective event.

( ) pj − Q ∆fj+k − ∆fj , ∀k ∈ {1, · · · , K} \ j,

or not. If pj ’s repetitions can be found in Tk :s, then pj is validated, else it is removed from Tj . An example of this procedure has been illustrated in Fig. 4 for K=3, where the set of CFOs is {−150, −50, 100} Hz, and the set of timeindex offsets from a reference time is {1000, 1100, 1160}. In this example, T1 ={1000, 1180, 1200, 1380, 1640} and T2 ={920, 1100, 1300}. In T2 , 1100 represents the correct starting time of signal with ∆f2 =-50 Hz; however, 920 and 1380 represent the shifted positions of peaks in Y1 (n), which has been shifted -80 indices. Referring to Fig. 5, one sees that the expected shift in position of peak for ∆f2 ∆f1 =Q(100)=0.088Nzc Fs Tb =80. Furthermore, it can be seen that the peaks in subfigure 3 are the shifted versions of peaks in subfigure 2, where the level of shift is Q(150)=140. In order to decrease the complexity of SPD, one may increase the length of preambles, i.e. Nzc , which eliminates the side peaks [29]. On the other hand, increase in Nzc may increase energy consumption of devices in transmission of longer preambles. Hence, Nzc regulates a tradeoff between energy consumption of end-devices and complexity of receiver in contention resolution. Investigation of this tradeoff is beyond the scope of this work, and the interested reader can refer to [33] for more info. C. The access-delay/processing-delay tradeoff The grant-free radio access scheme aims at reducing the signaling overhead in hope of decreasing both the access delay and the energy consumption of data transmissions. In this scheme, the probability of collision of data transmissions increases with the increase in the traffic load. The receiver proposed in this section is able to trigger successive interference cancellation in case of collisions, and recover collided packets to a certain extent. However, this improvement is achieved at the cost of increased processing delay at the receiver, as well as waiting for the reception of several packet replicas. Thus, we expect that in medium to high traffic load regimes, where the probability of collision increases, the grant-based radio access may outperform the grant-free radio access regarding energy consumption and communication delay. This switchover point is investigated in the following sections. IV. A NALYTICAL P ERFORMANCE E VALUATION In this section, we investigate performance of the grant-free radio access scheme. A. Reliability Analysis In order to model reliability, we first derive an analytical model for interference. Let us assume a device has been located at point z, and the BS has been located at the origin of the 2D plane. Devices are assumed to be distributed in the area based on a Poisson point process (PPP) with density λ. The received interference from other devices at the BS is characterized with its Laplace functional, enabling investigation of its moments [34].

Q( ∆ f): Peak Drift ( × NzcF s T b )

at

0.5

0

-0.5 -1/Tb

-0.5/T b

0.5/Tb

0

1/Tb

∆ f: CFO (Hz)

Fig. 5: The y-axis represents the expected drift in the peak location when CFO is ∆f .

We introduce a stationary and isotropic process Ψ, which represents the PPP containing locations of interfering nodes. For a BS located at the origin, the Laplace functional of the received interference at the receiver is given by [35]: [ ] [∏ ] ¯ t g(x)) , (1) LI (s) = E exp(−sI) = E Lh (sξP x∈Ψ

¯ t g(x) is the average received power due to an interwhere ξP ferer located at point x, ξ¯ models the partial time/frequency overlapping, to be discussed in the following text, h is the power fading coefficient associated( with the)channel between ¯ t g(x) is the Laplace the device and the BS, and Lh sξP functional of the received interference power. We assume a path-loss model of form g(x) = G||x||−δ , where δ is the pathloss exponent, and further assume that h represents Rayleigh fading with parameter µ, whose probability density function (PDF) is given by: ( ) ph (q) = µ exp −µq . (2) ( ) ¯ t g(x) is derived as: Using Laplace table, Lh sξP ¯ t g(x)) = Lh (sξP

1 ¯ t g(x)/µ . 1 + sξP

(3)

By inserting (3) in (1) and considering the fact that the received interferences from different device subsets are independent, we have: ] [∏ 1 LI (s) = Ex ¯ t g(x)/µ . x∈Ψ 1+sξP To proceed further, we recall a useful lemma from [36]. Lemma 1: If X ⊂ R2 is assumed to be a PPP with intensity function γ(x), for any Borel function b: R2 → [0, 1], we have: ( ∫ ) [∏ ] Ex b(x) = exp − (1 − b(x))γ(x)dx . (4) x∈X

Proof Presented in [36].

R2

Using Lemma 1 produces: ( ) ∫ ¯ t g(x) sξP LI (s) = exp −υ dx , ¯ R2 µ + sξPt g(x)   ∫ −υ  = exp dx µ ¯ t G||x||−δ + 1 sξP 2 R∞  ∫ −2πυ  = exp rdr , µ ¯ t Gr−δ + 1 sξP

is derived as: (5)

0

where υ ≤ λ denotes the density of interfering devices. When Fm ≥ w, i.e. when the maximum potential CFO is greater than the signal bandwidth, two simultaneous transmissions may interfere with each other, or may have no time/frequency overlapping areas. In this case, we approximate the received interference in two modes, including complete overlapping (ξ¯ = 1) and non-overlapping (ξ¯ = 0). From (5), LI (s) = 0 for ξ¯ = 0, and is a function of density of interfering nodes, NT w i.e. υ, for ξ¯ = 1. Furthermore, υ is modeled as Nλf Tr p 2F , m in which Nf is the number of available carrier frequencies, the first fraction represents load over each carrier, the second fraction represents the percentage of time in which device is transmitting, i.e. time-domain activity factor, and the third fraction represents the frequency-domain activity factor. For Fm ≤ w, i.e. when two simultaneous transmissions interfere each other with probability 1, we set the frequency-domain NT activity factor to 1, i.e. define υ as Nλf Tr p . Then, we derive the average fraction of overlapping in the frequency domain, ¯ as a function of probability distribution of CFO. In the i.e. ξ, following, we continue the analysis by assuming that ∆f is chosen uniformly from (−Fm , Fm ), i.e. ∆f ∼ U (−Fm , Fm ). One may extend our following results by replacing the uniform distribution with any other distribution of interest. Based on our assumption, the amount of overlapping in the frequency domain follows a uniform distribution as: ξ ∼ U (w − Fm , w). Now, one can derive the fraction of overlapping in the frequency domain as: ∫ ξ 1 w ξ Fm ξ¯ = = dξ = 1 − . (6) w w w−Fm Fm 2w

ps (z) = Pr (Pt hg(z) ≥ [θ + I]γth ) ( ) ∫ ∞ γth qµ = exp − )dPr(I + θ) ≥ q P g(z) 0 t = LI (s)Lθ (s) s= γth µ Pt g(z) ) ( ∫ ( ) ∞ µΘγth ||z||δ 2πυ (a) rdr exp − = exp − r δ 1 Pt G ) ξγ 1 + ( ||z|| 0 ¯ th ) ( ) ( 2 δ ¯ th ) δ2 2π csc( 2π ) exp − µΘγth ||z|| , = exp -υ||z||2 (ξγ δ δ Pt G (7) where Θ denotes the power spectral density of noise, and (a) follows from [37, Lemma 3.1]. Obviously, ps (z) = ps (||z||), i.e. success probability is a function of distance to the BS due to the isotropy of the path-loss function. Now, we have the required modeling to investigate probability of success for a VF, containing N replicas of the packet. By assuming independent processing of replicas at the BS, success probability in a VF transmission for a device located at distance z from the BS can be approximated as: N

Plb s (z) ≈ 1 − [1 − ps (z)] .

(8)

The expression in (8) is a lower-bound on the success probability, achieved by joint processing and combining of replicas, e.g. by equal gain combining or selection combining. On the other hand, by considering joint processing and combining of replicas leveraging channel-state and interference information, the success probability could be upper-bounded as: Pub (9) s (z) ≈ ps (z) γth ← 1 γth . N

In Section VI, we show by simulation results that the achievub able success probability falls in [Plb s (z), Ps (z)]. B. Delay Analysis Using the success probability modeling in (8)-(9), the experienced delay from packet arrival at a device located in distance z from the BS to successful reception at the BS can be modeled as: ∑B D(z) = [M Tp + Tack + Tw ](1 − Ps (z))i−1 Ps (z) i=1

− Tack − Tw ,

(10)

where we assumed that if a device does not receive the ACK within Tack seconds, it starts retransmitting after Tw seconds backoff time, and where B represents the maximum allowable number of attempts. C. Device Energy Consumption Analysis We proceed by deriving the success probability in a packet transmission. Let us denote by θ the additive noise at the receiver, and by γth the threshold signal to interference and noise ratio (SINR). Using the above derived interference model, success probability in packet transmission for the device located at point z, to the BS, located at the origin,

Here, we model the impact of system and traffic parameters on the energy efficiency of IoT devices in uplink communications. Denote by Λ = λπRc2 /Tr (Hz) the aggregated arrival rate of data packets at devices in a cell of radius Rc . Also, power consumptions of each device in the active, listening and transmission modes are denoted as Pc , Pl , and P¯t = αPt + Pc , respectively, in which Pc is the circuit power consumed by

electronic circuits, Pt is the transmission power, and α is the inverse power amplifier (PA) efficiency. The average energy efficiency in uplink communications in terms of bit-per-joule can be approximated as the ratio between amount of useful transmitted bits in a given interval to the consumed energy in that interval, as follows: ∫ Rc min{B, 1/Ps (z)}Λ[D − Doh ] 2z [ ] E= dz (11) P¯t N Tp + Pc [M − N ]Tp + Pl Tack Rc2 0 ∫ Rc 2z Λ[D − Doh ] 0 min{B, 1/Ps (z)} R 2 dz c ] , = [¯ Pt N Tp + Pc [M − N ]Tp + Pl Tack in which D denotes the average packet size, and Doh denotes number of overhead bits in a packet. When N = 1, δ = 4, and number of retransmissions after collision is unbounded, the energy efficiency expression can be further simplified as: ∫R Λ[D − Doh ]/Rc2 0 c exp(−az 2 − bz 4 )2zdz [ ] E= P¯t N Tp + Pc [M − N ]Tp + Pl Tack √ √ ∫ bR +ℓ exp(a2 /4b)Λ[D − Doh ]/[Rc2 b] ℓ c exp(−y 2 )dy [ ] = ¯ Pt N Tp + Pc [M − N ]Tp + Pl Tack √ √ exp(a2 /4b)λπ[D − Doh ]/[Tr b] π ] × = [¯ Pt N Tp + Pc [M -N ]Tp + Pl Tack 2 ( √ ) × erf( bRc + ℓ) − erf(ℓ) , µΘγth ||z|| a , a = υ(γth ) δ 2πδ csc( 2π where ℓ = 2√ . δ ), and b = Pt G b We can further extend the analysis, and derive the expected battery lifetime of devices in grant-free communications. In reporting applications, the packet generation process at each device can be modeled as a Poisson process, and hence, energy consumption of each device can be seen as a semi-regenerative process where the regeneration points are located at the end of successful data transmission epochs [4]. Denote the battery capacity of a device at the reference time as E0 , and the average time between two data reporting as Tr . Now, we define the expected battery lifetime at the regeneration point as the product of reporting period and the ratio between remaining energy and the average energy consumption per reporting period, as follows: 2

2

δ

E0 Tr [ ], 1 } P¯t N Tp + Pc (M -N )Tp + Pl Tack Est + min{B, Ps (z) (12) where Est is the average static energy consumption in each reporting period for data gathering and processing.

L(z) ≈

D. Analysis of Performance Tradeoffs In the following we investigate the tradeoff between energy consumption/battery lifetime of IoT devices and costs of the access network. Recall that the access network’s costs include deployment and operational expenses. In (12), we observed that the expected battery lifetime increases by increasing success probability in data transmission. On the other hand, • Eq. (7) indicates that decreasing communication distance contributes in increasing the success probability with exponent 2, when it comes to interference (the left side

of Eq. (7)), and with exponent 4, when it comes to noise (the right side of Eq. (7)). Hence, by increase in density of BSs, i.e. shortening device-BS distances, reliability of communications increases significantly, which in turn increases costs of the access network. • Eq. (7) further shows that success probability can be increased by increasing the available radio resources, i.e. system bandwidth, as the latter decreases υ, i.e. the chance of collision. These extra radio resources also mandate more investments in the access network. • Furthermore, as noted in Section III-B, detecting replicas in collision events requires either long synchronization preambles or sophisticated receivers to perform processing of derived peaks from cross correlations. The former increases the packet size, and hence, increases the collision probability, which in turn implies less energyand spectral-efficiency, as well as shorter battery lifetime. The latter increases complexity of the receiver, as well as the decoding delay. From an overall system perspective, we aim at minimizing the costs of the access network, maximizing the energy efficiency of communications, minimizing the delay in data transmission, and prolonging battery lifetime of devices. However, the above discussion shows that these objectives cannot be treated separately, as they are coupled in conflicting ways. Thus, it is important to investigate the bounds up to which trading resources for maximizing a specific performance metric is feasible. In the following section, we investigate the positive/negative effects of replica transmission on reliability and battery lifetime of devices, and devise a replica-transmission control scheme. V. T HE R EPLICA -T RANSMISSION C ONTROL S CHEME From (8)-(9), reliability in data transmission increases with the number of replica transmissions. On the other hand, (7) shows that increase in the number of transmitted packets increases the traffic load, and hence, decreases the success probability in transmission. Thus, the number of transmitted replicas per packet has a significant impact on reliability performance, and hence, is investigated in this section. From Eq. (7), one observes that devices which are closer to the BS have a higher success probability in data transmission. This is due to the fact that packets from other devices, which have been located further far from the BS, have a lower chance to be recovered from collisions. Thus, the reliability performance can be improved by quantizing the long-term experienced path-loss in communications with the BS, and optimizing the number of replica transmissions in each pathloss regime. We denote such an adaptive scheme as ADP. In the case of a 2D deployment of IoT devices in the service area, the threshold path-loss values could be mapped to threshold distance values. In this case, the far away located devices transmit more replicas per data packet in order to enhance their communications’ reliability at the cost of increase in the interference for nearby devices; however, transmission of nearby devices is immune to such interference to some extent due to the propagation loss. Now, the replica transmission

{N1 ,N2 ,dth }

R0

dth

subject to: N1 ≤ N2 ∈ N, R0 ≤ dth ≤ Rc .

(13)

In (7), we have derived ps (z) for a homogeneous setting in which all devices send the same number of replicas per data packet. For a network which benefits from ADP(N1 , N2 , dth ), this expression can be updated as follows: (∫ ) dth −2πυ1 ps (z) = exp rdr (14) r δ 1 R0 1 + ( z ) ξγ ¯ th (∫ ) ( ) Rc −2πυ2 µΘγth z δ × exp r δ 1 rdr exp − P G t dth 1 + ( z ) ξγ ¯ th ( ) = exp − υ1 [H(dth ) − H(R0 )] − υ2 [H(Rc ) − H(dth )] δ ( µΘγth z ) × exp − , Pt G where υ1 = υ|N =N1 , υ2 = υ|N =N2 , and υ has been derived in Section IV-A as a function of N . Also, for δ = 4, we have: 1 x2 1 H(x) = πatan( √ )/ √ . 2 ¯ ¯ z ξγth ξγth z 2

TABLE I: Simulation Parameters

Parameters

Default value

Cell outer and inner radius Number of devices Path-loss Thermal noise power W ; w; Fm ; Fs Pc ; Pl ; αPt Tb ; Tp ; Tack ; Tw Threshold SINR (γth ) D; Doh Dsyn-ti ; Dsyn-fr Esyn-ti ; Esyn-fr ; Est E0 M ; Nzc ; |S| Grant-based: NRA Grant-based: TRA ; TRes

3000, 50 (meter) 20000 d 133 + 38.3 log( 1000 ) -174 (dBm/Hz) 600; 200; 200; 4000 (Hz) 40; 80; 139 (mW) 0.01; 0.5; 1; 2 (sec) 1 100; 50 (bits) 2; 1 (sec) 2Pl ; Pl ; 5Pc (Joule) 5000 (Joule) 1; 2N ; 23 20 2; 1 (sec)

1

Probability of success

control problem reduces to finding the optimized threshold values and respective number of replica(s), such that the impact of increased success probability for far-away located devices worth the decrease in success probability for other devices. Let us consider a 2D deployment of IoT devices in a service area of inner and outer radius of R0 and Rc , with a single threshold distance value, i.e. dth . In this case, the proposed ADP scheme reduces to ADP(N1 , N2 , dth ), in which, devices located closer than dth send N1 replicas per packet, and other devices send N2 replicas per packet. Now, the optimization problem is written as: ∫ Rc ∫ dth [ ]N2 [ ]N1 1 − ps (z) dz min 1 − ps (z) dz +

(Ana) N=1 (Sim) N=1 (Ana-LB) N=2 (Ana-UB) N=2 (Sim-SC) N=2

0.8 0.6 0.4 0.2 0 0

500

1000

1500

2000

2500

Distance from the BS (meter)

Fig. 6: Validation of analytical expressions. The ADP scheme, which has been proposed here for enhancing the grant-free radio access, is similar to the coverage class notion in NB-IoT systems. In NB-IoT, a set of coverage classes are introduced, each associated with a number of signal repetitions, which are assigned to users based on their experiencing path-loss in communications [38]. These signal repetitions enable reliable communications of devices deployed in remote or isolated areas with the BS. While NB-IoT represents a big step towards prolonging battery lifetime in cellular networks, it still suffers from the required overhead signaling for access reservation and scheduled data transmission [11, 38]. One must note that in NB-IoT, all downlink/uplink transmissions of a class, including random access, scheduling, data transmission and acknowledgment, are repeated with the repetition order of the respective class. Thus, the required signaling, and its inevitable repetition in NB-IoT, may also affect the delay/energy performance of coexisting IoT devices, especially the ones classified with a lower repetition order [38]. In the next section, we investigate performance of the proposed grant-free radio access scheme, study the performance impacts of design parameters of the ADP scheme, and compare the results against the grant-based radio access scheme. The simulation results confirm that the

proposed grant-free radio access is a promising solution to be integrated in the legacy cellular networks, in order to address their shortcomings in providing long battery lifetime IoT connectivity. VI. P ERFORMANCE E VALUATION The system model implemented in this section assumes 20000 IoT devices distributed according to a spatial Poisson point process in a cell, with one BS at the center. The simulation parameters are listed in Table I, where Dsyn-ti , Dsyn-fr , Esyn-ti , and Esyn-fr represent spent delay and energy in time and frequency synchronization respectively. As a benchmark, we implement a simplified narrowband version of grant-based IoT communications over cellular networks. In this benchmark: (i) devices become synchronized by receiving system information over the broadcast channel (BCH); (ii) devices contend over the random-access channel (RACH) for resource reservation from a pool of NRA orthogonal preambles; (iii) successful devices receive response within TRes seconds over the downlink control channel (DCCH); (iv) the scheduled devices transmit their packets collision-free over the uplink shared channel

0.6

N=1; TiSy; FrSy N=2; TiSy; FrSy N=1; TiSy; FrAs N=2; TiSy; FrAs N=1; TiAs; FrAs N=2; TiAs; FrAs Grant-based

0.3

0.2

Energy Efficiency (Bit/mJoule)

Energy Efficiency (Bit/mJoule)

0.4

0.1

0

N=1; TiAs; FrAs N=2; TiAs; FrAs Grant-based ADP(1,2,1000) ADP(1,2,1500) ADP(1,2,1950) ADP(1,2,2500)

0.5 0.4 0.3 0.2 0.1 0

0

1

2

3

4

5

6

0

1


(a) Energy efficiency analysis

4

5

6

(a) Energy efficiency analysis N=1; TiAs; FrAs N=2; TiAs; FrAs Grant-based ADP(1,2,1000) ADP(1,2,1500) ADP(1,2,1950) ADP(1,2,2500)

15

Delay (Sec)

Delay (Sec)

10

3

20


15

2


5

10

5

0

0 0

1

2

3

4

5

6

0

1



10000 8000 6000 4000 2000 1

2

3

3

4

5

6

(b) Delay analysis Battery Lifetime ( × reporting period)

Battery Lifetime ( × reporting period)

(b) Delay analysis 12000

0

2


4

5

6


(c) Battery lifetime analysis

Fig. 7: Performance comparison of i) asynchronous grant-free radio access scheme (without the replica-transmission control), ii) synchronous, and iii) grant-based schemes

(USCH); (v) uplink and downlink channels are separated in the frequency domain; and (vi) uplink RA and data channels are multiplexed in time. A. Validation of Analytical Results Fig. 6 compares the analytical reliability expression derived in (8)-(9) with the simulations results, for Λ = 4.5 and γth = 3. For N = 1 replicas per packet, when there is actually no combination of replicas, the analytical expression in (8), or equivalently (9), is in quite match with the simulation results. For N = 2, one sees that the success probability

12000 N=1; TiAs; FrAs N=2; TiAs; FrAs Grant-based ADP(1,2,1000) ADP(1,2,1500) ADP(1,2,1950) ADP(1,2,2500)

10000 8000 6000 4000 2000 0

1

2

3

4

5

6


(c) Battery lifetime analysis

Fig. 8: Performance analysis of grant-free radio access with the replica-transmission control scheme

obtained by simulation results is indeed between the lowerand upper-bounds derived in (8) and (9), respectively. Further, increasing number of replica transmissions for all devices increases reliability for devices located close to the BS, and beyond a certain distance decreases it. This is due to the fact that increase in N increases the traffic load; the adverse effect of an increased traffic load on the performance of remote devices that experience a higher level of path-loss is shown in (7).

Probability of success

Battery lifetime (×1000 reporting period)

0.8

10 9 8 7 6

Λ=1.7 Λ=2.5 Λ=3.3 Λ=4.9

5 4 1000

1500

2000

2500

2750

3000

dth : Threshold distance

Battery lifetime ( × reporting period)

(a) Battery lifetime vs. threshold distance for ADP (1,2,dth ) 9000 8000 7000 6000 5000

Λ=2.5 Λ=3.3 Λ=4.5 Λ=5.7

4000 3000 1

2

3

4

5

N2 : Number of replica transmissions beyond dth


(b) Battery lifetime vs. number of replicas per packet for ADP (1,N2 ,2500) 1 0.8 N=1; TiAs; FrAs N=2; TiAs; FrAs ADP(1,2,1000) ADP(1,2,1500) ADP(1,2,1950) ADP(1,2,2500)

0.6 0.4 0.2

fraction of users in each region 0.0278 | 0.0833 |

0.1389 | 0.1944

| 0.2500

| 0.3056

0 0

500

1000

1500

2000

2500

3000

Probability of success w/o combining Probability of success due to combining Total probability of success

0.6

0.4

0.2

0 0.1

.9

1.7

2.5

3.3

4.1

4.9

5.7


Fig. 10: Impact of selection combining of replicas on success probability for ADP(1,2,2500), γth = 3. Fig. 7a represents bit-per-joule energy efficiency of IoT devices in uplink communications. Obviously, asynchronous grant-free access increases energy efficiency by decreasing the consumed energy in connection establishment and synchronization. The decrease in energy efficiency by increasing number of replicas could be justified by using the success probability evaluation in Fig. 6, which shows that increasing number of replicas from 1 to 2 for all devices, decreases success probability for devices located beyond 1200 m from the BS, i.e. for 83% of devices, since cell radius is 3000 m. Fig. 7b represents the evaluation of service delay. In the low traffic load regime, waiting for random access opportunities in the grant-based approach imposes a high access delay in comparison with the grant-free schemes. On the other hand, with the increase in the traffic load, there is a switchover point beyond which the experienced delay in grant-free schemes surpasses the grant-based scheme due to the collisions and backoff after collisions. Finally, Fig. 7c investigates the battery lifetime performance. As it could be expected from the results presented in Fig. 7a, the asynchronous grant-free scheme outperforms the benchmark schemes in battery lifetime performance significantly, especially in low to medium traffic load regimes.

Distance to the BS (meter)

(c) Success probability vs. distance from the BS (Λ = 2.5)

Fig. 9: Tuning design parameters for ADP(N1 , N2 , dth )

B. Comparison of Data Transmission Schemes In Fig. 7, FrAs, FrSy, TiAs, and TiSy denote frequency and time asynchronicity/synchronocity, respectively. In other words, TiSy means that the devices are slot synchronized, while FrSy means that CFOs of the devices can take equallyspaced discrete values, i.e. the devices are sub-channel synchronized, where the channels are spaced each 200 Hz (w=200 Hz). This figure evaluates performance of the proposed asynchronous grant-free data transmission scheme without replicatransmission control, which is denoted in the legend by [N=1,2,TiAs, FrAs]. The benchmarks are synchronous grantfree radio access, and grant-based radio access schemes.

C. Performance Evaluation of Transmission Control Scheme

the

Proposed

Replica-

The bottleneck of grant-free communications is the high traffic load regime in which, success probability for devices, especially the ones located far away from the BS, decreases. The proposed replica-transmission control scheme in Section V aimed at addressing this problem. Fig. 8 represents performance of ADP(1,2, dth ) for different dth values. From Fig. 8a, one sees that ADP(1,2,2500) outperforms the others in energy efficiency, especially the asynchronous grant-free scheme with fixed N , in medium to high traffic load regimes. Similar results can be seen in Fig. 8b for the service delay, in which APD(1,2,2500) outperforms the grant-based approach even in the high traffic load regime. The figure also shows that the battery lifetime performance has been also significantly improved by controlling the number of replica transmissions. It is also interesting to see the performance impacts of dth and

N2 on the ADP(N1 ,N2 ,dth ). First, Fig. 9a investigates changes in battery lifetime performance by changing the threshold distance. Furthermore, Fig. 9b represents performance of the ADP scheme by changing the value of N2 . It is clear that N2 = 2 achieves the highest battery lifetime for different traffic load regimes. Fig. 9c evaluates the success probability of different data transmission schemes by changing the communication distance. It is clear that the ADP(1,2,2500) can significantly increase the minimum experienced success probability in the cell, i.e. 100-200% improvement for regions beyond dth , which constitute 30% of the service area. Finally, Fig. 10 represents the impact of post processing of replicas, i.e. selection combining, when all replicas are in collision. Fig. 10 represents that in ADP(1,2,2500), around 8% of the total success probability is achieved by selection combining, when all replicas are in collision. As in ADP(1,2,2500), 30% of devices transmit 2 replicas per data packet, one sees that replica combining improves the success probability for devices beyond dth by 27%. VII. C ONCLUSIONS An asynchronous grant-free radio access transmission/reception scheme has been proposed for low-complexity IoT devices. The scheme aims at providing a low delay and energy consumption profile for short packet communications by removing the synchronization/reservation requirements at the cost of sending several packet copies at the transmitter side and more complex signal processing at the receiver side. Closed-form expressions of key performance indicators have been derived. It has been shown that by tuning the transmission parameters, one can achieve very long battery lifetime as well as reduced access delay, for low-complexity IoT devices. Also, the regions of the traffic load in which grant-free/grant-based radio access performs favorably have been investigated, and it has been shown that there is a switchover traffic load behind which, grant-free access outperforms the other solutions. The simulation results have verified the performance of the proposed solutions, and promote integration of them in future cellular networks, along with the legacy grant-based schemes in LTE and NB-IoT systems. Finally, we note that the proposed approach has the potential to be used in other asynchronous communications systems, e.g., in satellite communications. R EFERENCES [1] M. R. Palattella et al., “Internet of things in the 5G era: Enablers, architecture, and business models,” IEEE J. Sel. Areas Commun., vol. 34, no. 3, pp. 510–527, March 2016. [2] G. Miao et al., “E 2 -MAC: Energy efficient medium access for massive M2M communications,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4720–4735, Nov 2016. [3] Ericsson, Huawei, NSN, and et al., “A choice of future M2M access technologies for mobile network operators,” Tech. Rep., March 2014. [4] A. Azari et al., “Lifetime-aware scheduling and power control for M2M communications in LTE networks,” in IEEE VTC, 2015, pp. 1–5. [5] C. Bockelmann et al., “Towards Massive Connectivity Support for Scalable mMTC Communications in 5G networks,” IEEE Access, May 2018. [6] Nokia Networks, “Looking ahead to 5G: Building a virtual zero latency gigabit experience,” Tech. Rep., 2014.

[7] F. Cao and Z. Fan, “Cellular M2M network access congestion: Performance analysis and solutions,” in IEEE Wireless and Mobile Computing, Networking and Communications conference, Oct 2013, pp. 39–44. [8] A. Laya, L. Alonso, and J. Alonso-Zarate, “Is the random access channel of LTE and LTE-A suitable for M2M communications? a survey of alternatives.” IEEE Commun. Surveys Tuts., vol. 16, no. 1, pp. 4–16, Dec 2014. [9] I. Leyva-Mayorga et al., “Performance analysis of access class barring for handling massive M2M traffic in LTE-A networks,” in IEEE ICC, May 2016, pp. 1–6. [10] 3GPP TS 45.820, “Cellular system support for ultra-low complexity and low throughput internet of things (ciot),” Tech. Rep., 2015, (Rel. 13). [11] P. A. Maldonado et al., “Narrowband IoT data transmission procedures for massive machine-type communications,” IEEE Network, vol. 31, no. 6, pp. 8–15, Nov 2017. [12] P. Popovski et al., “Ultra-Reliable Low-Latency Communication (URLLC): Principles and Building Blocks,” arXiv preprint arXiv:1708.07862, 2017. [13] A. Azari, P. Popovski, G. Miao, and C. Stefanovic, “Grant-Free Radio Access for Short-Packet Communications over 5G Networks ,” in IEEE Global Communications Conference, 2017, pp. 1–5. ´ Morin, M. Maman, R. Guizzetti, and A. Duda, “Comparison of the [14] E. device lifetime in wireless networks for the internet of things,” IEEE Access, vol. 5, pp. 7097–7114, April 2017. [15] H. Tabassum et al., “Non-orthogonal multiple access (NOMA) in cellular uplink and downlink: Challenges and enabling techniques,” arXiv preprint arXiv:1608.05783, 2016. [16] Z. Ding et al., “A survey on non-orthogonal multiple access for 5G networks: Research challenges and future trends,” IEEE Journal on Selected Areas in Communications, vol. 35, no. 10, pp. 2181–2195, July 2017. [17] R. Karaki, A. Mukherjee, and J. F. Cheng, “Performance of autonomous uplink transmissions in unlicensed spectrum LTE,” in IEEE Global Communications Conference, Dec 2017, pp. 1–6. [18] H. Kim et al., “Multiple Access for 5G New Radio: Categorization, Evaluation, and Challenges,” arXiv preprint arXiv:1703.09042, 2017. [19] R1-163510 , “Candidate NR Multiple Access Schemes ,” Tech. Rep., April 2016, 3GPP TSG RAN WG1 Meeting 84, Busan, Korea. [20] X. Chen and D. Guo, “Many-access channels: The gaussian case with random user activities,” in IEEE International Symposium on Information Theory (ISIT), 2014, pp. 3127–3131. [21] Y. Polyanskiy, “A perspective on massive random-access,” in 2017 IEEE International Symposium on Information Theory, June 2017, pp. 2523– 2527. [22] P. Popovski et al., “Ultra-Reliable Low-Latency Communication (URLLC): Principles and Building Blocks,” arXiv preprint arXiv:1708.07862, 2017. [23] X. Chen, T. Chen, and D. Guo, “Capacity of gaussian many-access channels,” IEEE Transactions on Information Theory, vol. 63, no. 6, pp. 3516–3539, Feb 2017. [24] W. Yang et al., “Narrowband wireless access for low-power massive internet of things: A bandwidth perspective,” IEEE Wireless Commun., vol. 24, no. 3, pp. 138–145, May 2017. [25] R. D. Gaudenzi et al., “Asynchronous contention resolution diversity ALOHA: Making CRDSA truly asynchronous,” IEEE Trans. Wireless Commun., vol. 13, no. 11, pp. 6193–6206, Nov 2014. [26] Z. Li, S. Zozor, J. M. Drossier, N. Varsier, and Q. Lampin, “2D time-frequency interference modelling using stochastic geometry for performance evaluation in Low-Power Wide-Area Networks,” in IEEE International Conference on Communications, 2017, pp. 1–7. [27] F. Clazzer et al., “Exploiting combination techniques in random access MAC protocols: Enhanced contention resolution ALOHA,” arXiv preprint arXiv:1602.07636, 2016. [28] K. Fyhn et al., “Multipacket reception of passive UHF RFID tags: A communication theoretic approach,” IEEE Trans. Signal Process., vol. 59, no. 9, pp. 4225–4237, June 2011. [29] A. Azari, S. N. Esfahani, and M. G. Roozbahani, “A new method for timing synchronization in OFDM systems based on polyphase sequences,” in 2010 IEEE 71st Vehicular Technology Conference, May 2010, pp. 1–5. [30] M. T. Do, “Ultra-narrowband wireless sensor networks modeling and optimization,” Ph.D. dissertation, Lyon, INSA, 2015. [31] J. Fang et al., “Fine-grained channel access in wireless LAN,” IEEE/ACM Trans. on Networking, vol. 21, no. 3, pp. 772–787, Aug 2013.

[32] M. Hua et al., “Analysis of the frequency offset effect on Zadoff–Chu sequence timing performance,” IEEE Trans. Commun., vol. 62, no. 11, pp. 4024–4039, Oct 2014. [33] A. Bana et al., “Short packet structure for ultra-reliable machinetype communication: Tradeoff between detection and decoding,” arXiv preprint arXiv:1802.10407, 2018. [34] V. Suryaprakash, J. Moller, and G. Fettweis, “On the modeling and analysis of heterogeneous radio access networks using a poisson cluster process,” IEEE Trans. on Wireless Commun., vol. 14, no. 2, pp. 1035– 1047, Oct 2015. [35] M. Haenggi, R. K. Ganti et al., “Interference in large wireless networks,” Foundations and Trends in Networking, vol. 3, no. 2, pp. 127–248, 2009. [36] J. Moller and R. P. Waagepetersen, Statistical inference and simulation for spatial point processes. CRC Press, 2003. [37] R. K. Ganti and M. Haenggi, “Interference and outage in clustered wireless ad hoc networks,” IEEE Trans. on Inf. Theory, vol. 55, no. 9, pp. 4067–4086, Aug 2009. [38] A. Azari, G. Miao, C. Stefanovic, and P. Popovski, “Latency-energy tradeoff based on channel scheduling and repetitions in NB-IoT systems,” in IEEE Global Communications Conference, 2018.

Paper E: Self-organized Low-power IoT Networks: A Distributed Learning Approach Amin Azari and Cicek Cavdar Accepted in IEEE International Global Communication Conference (Globecom), 2018

155

Self-organized Low-power IoT Networks: A Distributed Learning Approach Amin Azari and Cicek Cavdar KTH Royal Institute of Technology, Email: {aazari, cavdar}@kth.se Abstract—Enabling large-scale energy-efficient Internet-ofthings (IoT) connectivity is an essential step towards realization of networked society. While legacy wide-area wireless systems are highly dependent on network-side coordination, the level of consumed energy in signaling, as well as the expected increase in the number of IoT devices, makes such centralized approaches infeasible in future. Here, we address this problem by selfcoordination for IoT networks through learning from past communications. To this end, we first study low-complexity distributed learning approaches applicable in IoT communications. Then, we present a learning solution to adapt communication parameters of devices to the environment for maximizing energy efficiency and reliability in data transmissions. Furthermore, leveraging tools from stochastic geometry, we evaluate the performance of proposed distributed learning solution against the centralized coordination. Finally, we analyze the interplay amongst energy efficiency, reliability of communications against noise and interference over data channel, and reliability against adversarial interference over data and feedback channels. The simulation results indicate that compared to the state of the art approaches, both energy efficiency and reliability in IoT communications could be significantly improved using the proposed learning approach. These promising results, which are achieved using lightweight learning, make our solution favorable in many low-cost low-power IoT applications. Index Terms—Coexistence, IoT, Reliability, Battery lifetime, Low-power wide-area network.

I. I NTRODUCTION From the first to the fourth generation (4G) of wireless networks, a majority of resources in telecommunications have been dedicated to optimize communication systems with respect to the physical channel constraints, such as noise and interference [1]. Thanks to the large communication bandwidth and advanced hardware/software used in both 4G base stations and user devices, 4G networks are able to offer highspeed, seamless, and reliable connectivity to users. Compared to the previous generations, the fifth-generation of wireless networks (5G) has an increased focus on providing connectivity for energy/complexity/cost constrained smart devices, i.e. Internet-of-things (IoT) [2]. The long-term envision is to provide low-cost, large-scale, and ultra-durable connectivity for everything which benefits from being connected. Until now, the design and optimization of communication networks have been based on statistical models for arriving traffic as well as physical constraints like noise and interference. User devices, mostly smart-phones with a daily charging routine, listen frequently (in the order of sub-seconds) to their serving base stations (BSs), which are responsible for managing the connections, sending connection instructions, and scheduling

radio resources. As complexity, scale and heterogeneity of wireless networks, especially due to the IoT traffic, availability of statistical models for arriving traffic in the network-side and ability of energy-limited devices in frequent listening to the access network become infeasible [3]. The state-of-theart wide area IoT enabling solutions could be categorized as evolutionary and revolutionary solutions. The former includes solutions that aim at accommodating IoT traffic in existing cellular infrastructure, e.g. LTE category 1 and M [4, 5]. The latter includes solutions which aim at enabling IoT communications in a narrow bandwidth with decreased signaling between devices and the access network [6]. Examples of such solutions are NB-IoT (inside 3GPP), and SigFox and LoRa (outside 3GPP). SigFox and LoRa, the two dominant IoT solutions over the unlicensed band, benefit from a simplified connectivity procedure, called grant-free access, which removes need for pairing, synchronization, and access reservation. Thanks to the reduced signaling in grant-free access, these solutions are able to offer ultra-long battery lifetimes in IoT communications [7]. Beside solutions over the unlicensed band, the grant-free access is expected to be also included in future releases of the 3GPP LTE [8]. While energy consumption in the grant-free radio access mode is extremely low thanks to the removal of signaling procedures, the reliability of communications in this mode is a bottleneck [9, 10]. For example, the ever-increasing coexistence of communications technologies over the ISM band, and lack of dynamic control over operation of IoT devices using these radio resources, result in no performance guarantee for IoT communications in this band [9]. Fig. 2 represents interference measurements in the European 868 MHz ISM band in Alborg, Denmark [9]. One sees in this figure that in use-cases like business park, the ISM band suffers from a high level of interference in some sub-bands, which on the other hand, means a high probability of collision on them. On the other hand, one sees that in case of smart transmission sub-channel selection, there are sub-bands which suffer from sporadic interference, and hence, probability of success over them is much higher. Similar problem, but in the codedomain, could be seen in spreading factor distribution of LoRa wide area networks (LoRaWAN), as discussed in [11]. In recent years, there is an ever increasing interest in leveraging machine learning tools for characterizing large-scale networks, where there is no statistical model for describing their behaviors, as well as for operation control of independent nodes which have limited information about their environments and

get information only from interactions with their environments [12]. In [13], network-side reinforcement learning has been proposed for overload control in LTE systems serving massive IoT traffic. In [14], self-organized clustering and clusteredaccess for massive IoT deployments have been investigated. In [15], use of multi-arm bandit (MAB) for IoT networks has been proposed, where devices learn how to avoid subchannels suffering from a high level of static interference. To realize self-organized IoT networks able to adapt themselves to the environment, here we investigate communication in coexistence scenarios, in which the choice of communications parameters, including data rate, sub-channel, transmit power, and number of repetitions, affects both capacity and battery lifetime of the network. Our aim is to enable low-cost IoT devices to increase reliability of their communications, while keeping their energy consumptions as low as possible. In order to investigate application of the derived results in practice, we further present a distributed learning approach for operation control in LoRa technology, and compare the results against the analytical results from solving the equivalent centralized optimization problem. The performance evaluation results indicate significant decrease in energy consumption and increase in probability of success in communications. The main contributions of this work include: • Present a lightweight learning approach designed based on internal and external regret for increasing energy efficiency and reliability of IoT communications, respectively, with reduced network intervention. • Develop an analytical model for performance evaluation of the distributed learning solutions by leveraging tools from stochastic geometry. • Present distributed learning for operation control of IoT devices utilizing LoRa technology. Evaluate reliability and energy efficiency of communications utilizing the proposed and benchmark solutions. Highlight tradeoffs between reliability of communications against unintended/adversarial interference and energy efficiency. The remainder of this paper has been structured as follows. System model is investigated in the next section. The proposed learning approaches are presented in section III. In section IV, distributed learning is employed for operation control of LoRa devices, and its performance is compared against the centralized optimized solution. Simulation results are presented in section V. Concluding remarks are given in section VI. II. S YSTEM M ODEL AND P ROBLEM F ORMULATION A multitude of IoT devices, denoted by set Φ, have been distributed in a wide service area. Different IoT devices differ in radio resource usage pattern, i.e. in average reporting period, signal bandwidth, transmit power, and data rate (packet transmission time). A frequency bandwidth of W is shared for communications, on which the power spectral density of noise is denoted by N . We aim at collecting data from a subset of IoT devices1 , Φs ⊂ Φ, and treat traffic from other devices as 1 For example, one may consider coexistence of LoRa and SigFox in an environment, where LoRa receiver treats SigFox signal as interference.

Fig. 1: Interference measurement in the ISM band [9]. Left: c business park, Right: hospital complex. (⃝2017 IEEE) interference. The problem to be tackled is operation control for devices of interest, i.e. members of Φs , by considering operations of all other devices into account. Assume at time t, the ith device from Φs has data to transmit. Then, the operation control problem could be written as follows: max

pi ,ci ,hi ,mi

F (Reli , EEi )

s.t: pi ∈ P, ci ∈ C, hi ∈ H, mi ∈ M,

(1)

in which F (·) represents the objective function in terms of reliability (Rel) and energy-efficiency (EE) of communications. Regarding different QoS requirements of different IoT applications, definition of F (·) may differ from one IoT application to the others. Here, we focus on a weighted sum of objectives, i.e. F (Reli , EEi ) = (1 − β)Reli + βEEi , where 0 ≤ β ≤ 1 offers a tradeoff between reliability and energy efficiency. Also, pi , ci , hi , and mi represent the selected transmit power, code2 , sub-channel, and number of transmitted replicas per packet3 . Furthermore, X represents the set of available values for Xi . In the sequel, we aim at solving this optimization problem. III. S ELF O RGANIZATION AS A S OLUTION A centralized solution to the problem in (1) is very complex4 , and not applicable to low-power IoT devices which require less frequent signaling with the access network in order to save energy. Thus, instead of solving the problem in a centralized manner, we leverage distributed online learning. In online learning, each device, also called hereafter agent, aims at maximizing its objective function F (Reli , EEi ) by choosing the best actions Ai = {pi , ci , hi , mi } ∈ A, given the rewards (ACK/NACK) of its previous actions. After 2 The

transmit code also determines the data rate [16]. the NB-IoT and SigFox, several replicas are transmitted per data packet for range extension and resilience against interference, respectively [6]. 4 In section IV, we investigate this optimization problem analytically to get insight on complexity order of the centralized solution. 3 In

choosing the action at time t, i.e. Ai (t), agent receives the reward, denoted by ξ(t) ∈ {1, 0}, where 1 and 0 represent acknowledgment and no acknowledgment respectively. This type of learning is commonly described as multi-agent multiarm bandit (MAB) in the literature [17]. In MAB learning, an agent chooses one of the K arms at each time to play, and receives a reward afterwards. The agent’s aim is to maximize its self-accumulative return or equivalently, minimize its self-accumulative regret5 . The MAB problem offers a tradeoff between exploration and exploitation, where the former indicates decision epochs in which agent tries different actions even if their previously observed rewards are less than the others, and the latter indicates decision epochs at which agent acts greedy based on the previous rewards. Due to its widespread applications in gambling, robotics, and etc., MAB learning has been well investigated in literature, and efficient solutions have been proposed to minimize agent’s regret. In the sequel, we investigate solutions to the IoT-device’s operation control problem in environments dealing with stochastic interference over the data channels and no interference over the feedback channel (modeled by stochastic MAB), as well as environments dealing with stochastic interference over the data channels and adversary interference over the feedback channel (modeled by adversary MAB).

Algorithm 1: Pseudo-code of UUCB1 . 1 2

Initialization: Zk (1)=0, Tk (1)=1, ∀k ∈ A; for t = 1, 2, · · · do √ - Update index: bk (t) = Zk (t) + α log(t)/Tk (t); - Take action: arg maxk∈A bk (t) → A(t); - Receive reward: ξ(t) ∈ {0, 1}; - Update reward: Zk (t+1)=Zk (t), ∀k ∈ A\A(t); ZA(t) (t+1)=ZA(t) (t)+ˆ z (t); - Update counter: TA(t) (t+1)=TA(t) (t)+1; Tk (t+1)=Tk (t), ∀j ∈ A\A(t); - return A(t);

Algorithm 2: Pseudo-code of UEXP3. 1 2

Initialization: Wk (1) = 1, ∀k ∈ A; for t = 1, 2, · · · do Wk (t) - Define Dist.: pk (t)=(1-ρ) ∑|A| j=1

Wj (t)

ρ + |A| , ∀k ∈ A;

- Take action: A(t) ∼ {p1 (t), · · · , p|A| (t)}; - Receive reward: ξ(t) ∈ {0, 1}; - Update weight: Wk (t+1)=Wk (t), ∀k ∈ A\A(t); ˆ

WA(t) (t+1)=WA(t) (t) exp( |A|pρξ(t) (t) ); A(t)

- return A(t);

A. Learning for Stochastic MAB For stochastic MAB, the MAB in which each arm’s rewards are drawn from a probability density function (PDF), upper confidence bound (UCB) index policies perform close to optimally [17]. The aim of UCB index policies is to select the arm with the largest upper confidence bound for the expected return. Then, each agent maintains an index function for each arm, which is a function of the past rewards of this arm, and in each decision epoch, selects the arm with the maximal index. Among UCB algorithms, the UCB1 algorithm [18], attains a regret growing at O(log n) in the stochastic MAB, where n is the number of rounds [19]. B. Learning for Non-stochastic MAB Having insights on optimized learning in stochastic MAB, we can investigate learning in non-stochastic MABs, where arms’ rewards are not drawn from a specific PDF. An example of non-stochastic MAB is the adversarial environment, in which, an adversary can interrupt the rewards (the feedback messages). Furthermore, in IoT solutions over the unlicensed spectrum, the feedback channel is also affected by the interference from coexisting technologies, and hence, the rewards might be interrupted. Efficient learning approaches for adversary settings have been proposed in literature, among them, the exponential-weight algorithm for exploration and exploitation (EXP3) is a promising approach [20]. On each decision epoch t, EXP3 chooses an action, out of A, according to a set of respective distributions, i.e. Ai (t) ∼ {p1 (t), · · · , p|A| (t)}. EXP3 assigns each action a probability mass function based 5 Regret indicates difference between reward of a non-optimal and the optimal action.

on mixing the estimated cumulative reward and the uniform distribution, where the former incurs the exploitation mode, and the latter incurs √the exploration mode. EXP3 attains a regret growing as O( n) in the adversarial MAB [19]. C. Light-weight Learning for Low-power IoT Networks Recall the optimization problem in (1), in which the aim is to maximize the reliability and energy efficiency of devices. As a distributed solution to this problem, here we incorporate both reliability and energy efficiency in the learning process for stochastic and non-stochastic settings. Let’s start with the stochastic setting, where at the end of each successful transmission, the device receives an acknowledgment. Once the acknowledgment is received, the accumulated reward of the respective action is incremented by one in UCB1 algorithm [18]. Now, denote by Ei and Emin , the consumed energy in packet transmission using action i, and the minimum consumed energy amongst actions achieved a successful packet transmission respectively. In our proposed learning solution, we modify the reward achieved by choosing action k as: ˆ = ξ(t)(1-β) + ξ(t)βEk /Emin , ∀k ∈ A, ξ(t)

(2)

in which ξ(t) ∈ {0, 1} represents the acknowledgment, β is a design parameter offering tradeoff between reliability an energy efficiency, and t represents the time index. Based on this updated reward function, we present the updated UCB1 (UUCB1 ) algorithm in Algorithm 1. Following the same approach, and by updating the reward function in EXP3 [20], we present the updated EXP3 (UEXP3) algorithm in Algorithm

2. In these algorithms, α and ρ are the design parameters, which offer tradeoff between exploration and exploitation in the UUCB1 and EXP3 respectively. Furthermore, the device index, i.e. i in the underscript, has been dropped. Mapping (2) to the objective function in (1), one sees that F (Reli , EEi ) in (1) has been modeled by the modified reward function, ˆ = F (Reli , EEi ). Furthermore, the first term in (2) i.e. ξ(t) represents the external regret, while the second term represents the internal regret. In the following section, we employ the proposed learning scheme in operation control of IoT devices connected through LoRa technology, and compare the results against the results of a centralized optimized solution. IV. D ISTRIBUTED L EARNING FOR I OT O PERATION C ONTROL : A L O R A T ECHNOLOGY E XAMPLE A. Communication Using The LoRa Technology LoRa, the physical layer of LoRaWAN, aims at enabling low-power low-rate long-distance communications. Communication in LoRa occurs in 3 sub-channels in the public ISM band; each with bandwidth (BW) of 125 KHz. High resilience to noise and interference is the key to operate efficiently in the ISM band. Towards this end, the chirp spread spectrum (CSS) modulation has been used in LoRa, which enables signals with different spreading factors (SFs) to be distinguished and received simultaneously, even if they are transmitted at the same time on the same channel. The spreading factors, ranging from 7 to 12, denote the number of chirps used to encode a bit, and hence, determine the data rate: R(c) = c×BW×µ , ∀c ∈ C = {7, · · · , 12}, where µ is the code2c rate. Based on [16], the required SNRs for correct detection of signals with spreading factors {7, · · · , 12} are γthN = {-6,-9,12,-15,-17.5,-20}, respectively. Then, one sees that by increase in the SF index, data rate decreases and resilience to noise increases. Finally, LoRaWAN supports the following transmit powers for communication: {2, 5, 8, 11, 14} dBm [21].

2) Centralized Optimized Operation Control: In order to save space, in this version we present the formulation for |H|=|P|=1, i.e. we consider a single-channel single-transmit power level LoRa network in which, we aim at distributing spreading factors amongst devices. Furthermore, we assume interference is only coming from the coexisting LoRa devices. Regarding the fact that by increase in the SF index, data rate decreases, probability of collision increases, and resilience to noise increases, nodes located closer to the BS are expected to choose lower-index SFs and vice versa [11]. Then, the SF allocation problem is equivalent to finding the optimized density of nodes, which are using different spreading factors in each region of the service area. Let us divide the service area to a set of rings, each with inner and outer radius of rj,1 and rj,2 respectively, where in each of them, density of nodes which are using each SF is assumed to be constant (j ∈ J). By extending the results in [22, Section III.A], one can derive6 the Laplace functional of interference from devices distributed on the jth ring, denoted by Φj,c ⊂ Φ, with transmit power Pt , reporting period Trep , packet length D, spreading factor C, and transmission time T (c) = D/R(c), as: ∫ rj,2 ) ( λj,c Tc /Trep rdr , LΦj,c (s)= exp − 2π 1 +1 −δ rj,1 sPt Gr in which Gr−δ is the pathloss. Now, the Laplace functional of received interference from all devices ∏ using spreading factor c could be written as LΦc (s) = j∈J LΦj,c (s). Let N and IΦ denote the additive noise and interference from set ϕ of devices at the receiver. Using the above derived interference model, probability of success in packet transmission for a device located at distance z to the BS, using spreading factor c ∈ C, is derived as: ps (c, z)=Pr(Pt hGz −δ ≥ γc N )Pr(Pt hGz −δ ≥ γthI IΦc ), γth =LΦc (s) , LN (s) γc I s=

B. Operation Control in LoRa Consider a LoRa gateway in a 2D plane with multitude of devices, distributed according to a Poisson point process (PPP) with density λ in r ≤ R, where r is the distance to the BS located at the origin and R is the boundary of service area for the gateway. The IoT devices aim at data transfer to the gateway on average each Trep seconds. Recalling the operation control problem in (1) the LoRa operation control consists in solving the following problem: max F (Reli , EEi )

pi ,ci ,hi

=

∏

exp

=

∏

1) Distributed Learning for Operation Control: One can directly apply the presented Algorithms 1 and 2 in section III, in order to solve the optimization problem in (3). In this case, the set of actions, i.e. A, includes 90 pairs of actions, each including a transmit power, sub-channel, and spreading factor.

Pt Gz −δ

∫

rj,2

rj,1

j∈J

(

exp -λj,c

j∈J

s=

Pt Gz −δ

) ( N γc (c)z δ ) -2πλj,c Tc /Trep , rdr exp r δ 1 Pt G 1+( z ) γth I

) ( N γc z δ ) Tc [Q(rj,2 )-Q(rj,1 )] exp , Trep Pt G

where γc = γthN (c), and for δ = 4, we have7 : Q(x) = πatan( √

(3)

s.t: pi ∈ P = {2, 5, 8, 11, 14}dBm, ci ∈ C = {7, 8, 9, 10, 11, 12}, hi ∈ H = {1, 2, 3}.

(

(4)

1 1 x2 )/ √ . γth I z 2 γth I z 2

Now, the optimization problem in (3) reduces to: ] ∫ rj,2 ∑ ∑ λj,c [ Tc max (1 − β) ps (c, z)dz + β , λj,c λ T1 rj,1 j∈J c∈C ∑ s.t: λj,c = λ.

(5)

c∈C

6 Details

will be presented in the journal version. 7 Similar expressions could be found using table of integrals for other pathloss exponent values.


V. P ERFORMANCE E VALUATION

0.48 0.46 0.44 Algorithm 1 Centralized optimized strategy

0.42 0.4 0.38 0.36 0

10

20

30

40

50

60


Fig. 2: Learning for SF selection control: comparison of Alg1 and the centralized optimized solution. TABLE I: Parameters for performance analysis [16]. Parameters

Values

Service area Aggregated packet arrival rate: σ = Nd /Trep Packet length: D Number of sub-channels Bandwidth: W Code rate: µ Threshold SNR: γthN Threshold SIR: γthI Power consumption: Pt , Pc , η Learning parameters: α, β, ρ

Circle of radius 2 Km 12.5 (Sc1), 2.5 (Sc2,3) per seconds 100 bytes (Sc1), 20 bytes (Sc2,3) 1 (Sc1,2), 3 (Sc3) 125 KHz 4/5 {-6,-9,-12,-15,-17.5,-20} dB 6 dB {8,14} dBm, 10 dBm, 2 0.1, 0.5, 0.4

By solving this problem, the density of usage of each SF is identified as a function of distance to the BS. From (5), one sees that solving the centralized optimization problem, even in the simplified form where we assumed a 2D PPP distribution with a single-level transmit power and a singlechannel LoRa network without external interferer, is highly complicated. In the following, we compare performance of the centralized optimized solution against the distributed learning approach. 3) Comparison of Solutions: Fig. 2 compares probability of success in transmission for Algorithm 1 and the result of centralized optimized strategy for the case: C = {7, 10}, γth N ={-6,-15} dB, Pt =14 dBm, Nd = 1000, Trep = 200 sec, and D= 100 bytes. Other parameters could be found in Table I. The x-axis represents index of transmitted packets. Here, each node decides to send data over SF 7 or 10, i.e. |A| = 2. One sees that after a few number of transmissions, the learning’s results become close to the centralized solution. These results also promise a valuable improvement in the battery lifetime due to the fact that without need for listening to the BS and signaling, we have configured communication parameters of IoT devices in a distributed form. Detailed energy and reliability performance evaluations are presented in the following section.

In this section, we present the simulation results in the context of operation control for the LoRa technology. We assume a massive number of LoRa nodes have been distributed according to a PPP in a cell of radius 2 Km. Our aim is to distribute 6 spreading factors and two transmit power levels amongst them. The simulation parameters have been presented in Table I. In this table, Sc1, Sc2, and Sc3 refer to three different scenarios in which, simulations have been carried out. In the following figures, Alg1 refers to the Algorithm 1, Alg2 refers to the Algorithm 2, EqLoad refers to the centralized algorithm proposed in [11], in which number of devices using a SF is proportional to its data rate, and RandSel refers to the algorithm in which SFs are selected randomly. The aforementioned schemes differ in the way they choose the SF, while all of them choose Pt = 14 dBm as the transmit power. In contrast, Alg1(PC) refers to the Algorithm 1 in which, devices have freedom to choose their spreading factors and transmit powers out of C and P, respectively. Fig. 3 represents reliability and energy efficiency performance evaluations of different schemes versus index of transmitted packets in a scenario in which there is no external interference. One sees that Alg1 is converging after a few transmissions and is able to have the superior performance in success probability and energy efficiency, even in comparison with the centralized solution proposed in [11]. Fig. 4 represents how different SFs have been allocated to devices in different regions of the cell by following Alg1 (left) and the EqLoad scheme (right) [11]. Fig. 5 represents the reliability and energy consumption performance evaluations in the same setup as for Fig. 3, with the only difference that here external stochastic interference has been considered. The probability of occurring interference on each SF differs from the others. One sees that huge increase in probability of success and decrease in energy consumption could be achieved by using the proposed learning approaches. Furthermore, this figure presents the tradeoff between reliability and energy efficiency, which can be controlled by the design parameter, β. In this regard, one sees that Alg1(PC) achieves an acceptable success probability in data transmission with an ultra-low energy consumption profile by using β = 0.5 (the solid green line). On the other hand, by decreasing β to 0.01, one sees that probability of success has been significantly improved while energy consumption has been increased in comparison with the previous state. One must note that the energy consumption results in this section represent the energy consumption per packet transmission trial, and hence, the ultimate decrease in energy consumption of devices using our learning approaches will be much higher due to the following facts. (i) The probability of success achieved using the learning approaches is higher than the other schemes, which on the other hand implies less number of required retransmissions. (ii) The learning approaches do not need frequent listening to the BS and receiving control data from them, which on the other hand implies significant

0.4

1500

1000

1000 SF7 SF8 SF9 SF10 SF11 SF12

500 0 -500

y-axis (meter)

0.35

-1000

0

10

20

30

40

50

0.06 0.05 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

0.04 0.03 0.02 0.01 0

10

20

30

40

50


Fig. 3: Learning for power and SF selection control without external interference (Sc1).

reduction in energy consumption for coordination. Fig. 6 represents the energy consumption and reliability results in an adversarial setting where, the adversary affects 50% of the feedback messages sent by the BS. In other words, with probability of 0.5, an ACK message is substituted by a NACK message and vice versa. One sees that in adversary environments, Alg2 outperforms the others in reliability and energy consumption performance measures. One must note that the energy consumption per packet transmission for Alg1 is lower than Alg2. However, due to the increased number of required retransmissions in Alg1, the total energy consumption of Alg1 will be higher than Alg2. Fig. 7 represents the energy efficiency and reliability results for the problem in which devices learn to send data over sub-channels suffering from different levels of stochastic interference. One sees that the Alg1(PC) scheme is able to achieve more than 50% decrease in energy consumption per data transfer, while increasing the probability of success by 30% in comparison with the benchmark schemes. VI. C ONCLUSION Distributed learning for IoT communications with reduced network intervention has been investigated. Reducing signaling between IoT devices and the access network results in decreasing energy consumption per data transfer for IoT devices as well as decreasing control over radio resource usage for the access network. In order to benefit from the former, and suffer as low as possible from the latter, here we have presented a light-weight distributed learning approach,

-1500

0 -500

-1500

-2000 -2000

-1000

0

-2000 2000 -2000

1000

-1000

x-axis (meter)

0

1000

x-axis (meter)

Fig. 4: The constellations of selected SFs for Alg1 (left) and EqLoad Algorithm (right). Probability of success in transmission

0.3

SF7 SF8 SF9 SF10 SF11 SF12

500

-1000

0.8

0.7

0.6 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel Alg.1(PC), β=0.01

0.5

0.4 0

20

40

60

80

Index of transmitted packets Consumed energy per packet (J)

Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel


2000

1500

y-axis (meter)


2000

12

×10 -3 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel Alg.1(PC), β=0.01

10 8 6 4 2 0

20

40

60

80


Fig. 5: Learning for power and SF selection control with stochastic external interference (Sc2).

to be implemented in the device-side. The presented approach leverages external and internal regrets, for minimizing energy consumption and collision probability, respectively, in data transmission over shared wireless channels. The proposed approach has been subsequently employed in operation control of IoT devices using LoRa technology, where analytical as well as simulation results have been derived to characterize performance of the proposed distributed learning approach versus the centralized optimized approach. The analytical and simulation results represent significant improvement in probability of success in data transmission as well as battery lifetime of devices by utilizing the proposed learning approach, even in adversarial setups. These results confirm that equipping IoT devices with lightweight learning enables them to adapt themselves effectively to the environment, and hence, makes communications more reliable and durable with

2000


R EFERENCES

0.5

0.45

0.4 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

0.35

0.3 0

20

40

60

80


×10

-3

12 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

10 8 6 4 2 0

20

40

60

80



Fig. 6: Learning for power and SF selection control in an adversarial setting with stochastic external interference (Sc2).

0.8 0.75 Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

0.7 0.65 0.6 0.55 0.5 0

20

40

60

80


Index of transmitted packets 6

×10 -3

5 4

Alg.1 Alg.2 Alg.1(PC) EqLoad RandSel

3 2 1 0 0

20

40

60

80


Fig. 7: Learning for power and sub-channel selection control with stochastic external interference (Sc3). All devices are using SF 9.

reduced network intervention.

[1] P. Popovski, “Will communications theory finally make itself redundant?” IEEE ComSoc Technology News, 2017. [2] C. Mavromoustakis, G. Mastorakis, and J. M. Batalla, Internet of Things (IoT) in 5G mobile technologies. Springer, 2016. [3] M. Kulin, C. Fortuna, E. De Poorter, D. Deschrijver, and I. Moerman, “Data-driven design of intelligent wireless networks: An overview and tutorial,” Sensors, vol. 16, no. 6, p. 790, June 2016. [4] A. Azari and G. Miao, “Network lifetime maximization for cellularbased M2M networks,” IEEE Access, vol. 5, pp. 18 927–18 940, 2017. [5] Nokia Networks, “LTE-M – optimizing LTE for the Internet of things,” Tech. Rep., 2015. [6] W. Yang, M. Wang, J. Zhang, J. Zou, M. Hua, T. Xia, and X. You, “Narrowband wireless access for low-power massive internet of things: A bandwidth perspective,” IEEE Wireless Commun., vol. 24, no. 3, pp. 138–145, 2017. ´ Morin, M. Maman, R. Guizzetti, and A. Duda, “Comparison of the [7] E. device lifetime in wireless networks for the internet of things,” IEEE Access, vol. 5, pp. 7097–7114, 2017. [8] R1-163510 , “Candidate NR Multiple Access Schemes ,” Tech. Rep., April 2016, 3GPP TSG RAN WG1 Meeting 84, Busan, Korea. [9] M. Lauridsen et al., “Interference measurements in the european 868 MHz ISM band with focus on LoRa and SigFox,” in IEEE WCNC, 2017, pp. 1–6. [10] M. Masoudi, A. Azari, E. A. Yavuz, and C. Cavdar, “Grant-free Radio Access IoT Networks: Scalability Analysis in Coexistence Scenarios,” in IEEE ICC, 2018. [11] F. Cuomo et al., “EXPLoRa: EXtending the Performance of LoRa by suitable spreading factor allocations,” in IEEE 13th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob), 2017, pp. 1–5. [12] C. Jiang, H. Zhang, Y. Ren, Z. Han, K. C. Chen, and L. Hanzo, “Machine learning paradigms for next-generation wireless networks,” IEEE Wireless Communications, vol. 24, no. 2, pp. 98–105, April 2017. [13] Y. J. Liu, S. M. Cheng, and Y. L. Hsueh, “eNB selection for machine type communications using reinforcement learning based markov decision process,” IEEE Transactions on Vehicular Technology, vol. 66, no. 12, pp. 11 330–11 338, Dec. 2017. [14] A. Azari, “Energy-efficient scheduling and grouping for machine-type communications over cellular networks,” Ad Hoc Networks, vol. 43, pp. 16 – 29, 2016. [15] R. Bonnefoi, L. Besson, C. Moy, E. Kaufmann, and J. Palicot, “Multiarmed bandit learning in IoT networks: Learning helps even in nonstationary settings,” in CROWNCOM, 2017. [16] O. Georgiou and U. Raza, “Low power wide area network analysis: Can LoRa scale?” IEEE Wireless Communications Letters, vol. 6, no. 2, pp. 162–165, 2017. [17] M. S. Talebi Mazraeh Shahi, “Minimizing regret in combinatorial bandits and reinforcement learning,” Ph.D. dissertation, KTH Royal Institute of Technology, 2017. [18] P. Auer, N. Cesa-Bianchi, and P. Fischer, “Finite-time analysis of the multiarmed bandit problem,” Machine learning, vol. 47, no. 2-3, pp. 235–256, 2002. [19] S. Bubeck et al., “The best of both worlds: stochastic and adversarial bandits,” in Conference on Learning Theory, 2012, pp. 1–23. [20] R. Kl´ıma et al., “Combining online learning and equilibrium computation in security games,” in International Conference on Decision and Game Theory for Security. Springer, 2015, pp. 130–149. [21] B. Reynders, W. Meert, and S. Pollin, “Power and spreading factor control in low power wide area networks,” in IEEE ICC, 2017. [22] A. Azari, M. Masoudi, and C. Cavdar, “Optimized Resource Provisioning and Operation Control for Low-power Wide-area IoT Networks,” arXiv preprint arXiv:1804.09464, 2018.

Paper F: Serving Non-Scheduled URLLC Traffic: Challenges and Learning-Powered Strategies Amin Azari, Mustafa Ozger, and Cicek Cavdar Submitted to IEEE Communications Magazine, August 2018.

165

1

Serving Non-Scheduled URLLC Traffic: Challenges and Learning-Powered Strategies Amin Azari, Mustafa Ozger, and Cicek Cavdar

Abstract—Supporting ultra-reliable low-latency communications (URLLC) is a major challenge of 5G wireless networks. Whilst enabling URLLC is essential for realizing many promising 5G applications, the design of communications’ solutions for serving such unseen type of traffic with stringent delay and reliability requirements is in its infancy. In prior studies, physical and MAC layer solutions for assuring the end-to-end delay requirement of scheduled URLLC traffic have been investigated. However, there is lack of study on enabling non-scheduled transmission of urgent URLLC traffic, especially in coexistence with the scheduled URLLC traffic. This study at first sheds light into the coexistence design challenges, especially the radio resource management (RRM) problem. It also leverages recent advances in machine learning (ML) to exploit spatial/temporal correlation in user behaviors and use of radio resources, and proposes a distributed risk-aware ML solution for RRM. The proposed solution benefits from hybrid orthogonal/non-orthogonal radio resource slicing, and proactively regulates the spectrum needed for satisfying delay/reliability requirement of each traffic type. A case study is introduced to investigate the potential of the proposed RRM in serving coexisting URLLC traffic types. The results further provide insights on the interplay between the reliabilities of coexisting traffic, uncertainties in users’ demands and channel conditions, and amount of required radio resources. Index Terms—5G, scheduled/non-scheduled coexistence, IoT, machine learning, proactive resource provisioning, URLLC.

I. I NTRODUCTION HE fifth generation of wireless networks (5G) has targeted accommodating a diverse set of wireless services, from enhanced mobile broadband (eMBB) to the Internet of Things (IoT) [1]. The latter itself could be categorized into two distinct service types, i.e. massive machine type communications (mMTC), which aims at connecting everything that benefits from being connected, and ultra-reliable lowlatency communications (URLLC), which requires successful data transmission within a strictly bounded short time interval [2]. URRLC is considered as an essential prerequisite of a new wave of services including drone-based delivery, smart factory, remote control, and intelligent transportation systems (ITS) [3]. In the uplink direction, enabling URLLC requires guaranteeing delay and reliability requirements for both scheduled transmissions of connected devices as well as non-scheduled transmissions. The non-scheduled transmissions could be (i) from connected/disconnected devices which have urgent data, usually of short payload size; (ii) from devices which ask access reservation for subsequent scheduled transmissions of

T

The authors are with the School of Electrical Engineering and Computer Science, KTH - The Royal Institute of Technology, Stockholm, Sweden (email: {aazari, ozger, cavdar}@kth.se).

critical data; or (iii) urgent device-to-device communications. The coexistence of scheduled and non-scheduled traffic could happen in many scenarios like vehicular networks, and monitoring and alarm systems [3, 4]. Then, the set of resources allocated to URLLC should be managed to be used for both scheduled and non-scheduled communications. Isolation among services via allocating orthogonal slices of resources is a common practice when dealing with ergodic objectives like throughput. However, in URLLC applications, due to the limitations of time-diversity and crucial need to large spectrum bands for providing frequency diversity, isolation through orthogonal slicing is challenging [3, 5]. Recently, serving coexisting eMBB/URLLC services over limited radio resources has attracted attentions, and non-orthogonal RRM for serving URLLC has been proposed [1, 6]. While the orthogonal RRM is itself a complex problem, the non-orthogonal RRM, which requires specifying shared/dedicated resources and scheduling rules over shared ones, is a much more complex problem. Furthermore, to comply with URLLC delay requirements, this problem is needed to be solved in very short time scales. Among candidate enablers for solving such large-scale, complex, and time-limited RRM problem, artificial intelligence (AI) is promising [7]. In the light of recent advances in computing and storage technologies, AI is making the leap from traditional pattern recognition use cases to governing complex systems through advanced machine learning (ML) approaches. While in previous decades, machine learning, as a major branch of AI, has experienced several up and down periods, this time its technology readiness level is so high that it has already penetrated to the design of many complex systems [7]. This motivates us to investigate usefulness of leveraging ML in RRM for serving the URLLC traffic. While the traditional ML consists in a single node which makes decisions in a centralized manner, the delay/reliability requirements of URLLC call for novel, scalable and distributed learning approaches. Here, we investigate RRM for serving coexisting nonscheduled with scheduled URLLC services. Our main focus is on serving non-scheduled short packet URLLC transmissions along with the scheduled URLLC transmissions. The results derived within this work could be easily extended to the coexistence of URLLC and non-URLLC traffic by modifying the delay/reliability constraints of the scheduled traffic. The main contributions of this work include: (i) introduce coexistence of scheduled and non-scheduled URLLC traffic as a challenge to be tacked for realization of URLLC; (ii) introduce a hybrid orthogonal/non-orthogonal multiple access (OMA/NOMA) scheme to serve coexisting URLLC traffic;

2

!

BS

downlink data /control/ACK

$

!

!

!

signaling FPT-NS

V2V-NS

scheduled transmission scheduled transmission

signaling

V2I-NS

BS!

V2V-NS

# !

"

(a) V2X communications within the ITS use case

(b) Communications exchanges

Fig. 1: An illustrative example of system model in the ITS use case, including scheduled (green dashed arrows) and non-scheduled (red dotted arrows) URLLC communications.

(iii) investigate a distributed learning-powered RRM solution, including intelligent resource provisioning, scheduling, and utilizing modules, for spectrum efficient yet reliable coexistence management; and (iv) investigate different sources of uncertainty in proactive resource scheduling, how they affect the decisions, and the amount of radio resource which must be sacrificed for compensating the performance loss due to such uncertainties. Finally, open challenges and potential solutions in RRM for URLLC coexistence are discussed. II. E NABLING URLLC: T HE ROLE OF E FFICIENT R ADIO R ESOURCE M ANAGEMENT In URLLC applications, we usually deal with infrequent bursty traffic, where the arrival time of the burst could not be predicted [1, 6]. Due to stringent delay constraints, once a URLLC packet is generated, the packet must be transmitted immediately without any delay [3, 6]. This abrupt transmission, which usually has a short payload size, could be an alarm message, or the first packet transmission (FPT) which will be followed by upcoming scheduled transmissions. Apart from the non-scheduled URLLC traffic, we also need serving scheduled URLLC traffic which is not necessary of short payload size [3, 8, 9]. The resource allocation to this traffic type should guarantee no packet drop at the device due to the expiration of data [8, 9]. Fig. 1(a) represents a graphical illustration of the scheduled/non-scheduled URLLC traffic coexistence in the ITS use-case. In this use case, intelligent vehicles communicate with each other (V2V) as well as with the communication infrastructure (V2I) for traffic efficiency and safety. One observes that the non-scheduled communications include (i) FPT of devices which transmit urgent data and reserve access for further messages [3] (refer to FPT-NS in Fig. 1(b)); (ii) urgent message from a device to the access network [6] (refer to V2I-NS in Fig. 1(b)); and (iii) urgent V2V communications [4] (refer to V2V-NS in Fig. 1(b)). One further observes that the non-scheduled transmissions are infrequent, and could have temporal correlation with the scheduled transmissions (the FPT-NS in Fig. 1(b)). While the above traffic types and requirements have been introduced in the ITS use case, they

are generic, and could be also found in other use cases like surveillance and alarm systems. Based on the above discussions, enabling URLLC requires tackling three major issues: (i) enabling ultra-reliability in communications, which on the other hand requires diversity; (ii) enabling low-latency communications, which limits time diversity and enforces frequency diversity; and (iii) coexistence management of scheduled and non-scheduled traffic, which is due to the fact that frequency resources could not be scaled with the amount of URLLC traffic. The coexistence management calls for an efficient resource management strategy, without which, scalability of communications systems in serving URLLC service is challenging. While in prior studies, the first two issues have been investigated, there is lack of research on the coexistence management within the URLLC services [6, 8]. Towards serving a coexistence of scheduled and non-scheduled URLLC traffic: • delay and packet error probability for the first transmission (non-scheduled traffic) should be reduced dramatically. Since critical data must be dispatched immediately, grant-free access for this traffic is mandatory. • queuing delay violation for scheduled traffic should be minimized to prevent packet drop due to expiration of critical data. This requires reliable and in time transmission of the generated packets. The first item, requires revolutionary multiplexing schemes breaking the barrier of interference-free orthogonal transmissions [1, 6]. The need for prompt access may occur in any time instant, e.g. while resources have been reserved for the scheduled-users, hence, serving these two traffic types together poses coexistence challenges. The main research question is to find an efficient RRM scheme to serve both scheduled and non-scheduled URLLC traffic over a limited set of radio resources. A. The Hybrid Multiple Access Solution The straightforward way of serving non-scheduled and scheduled traffic is to allocate them orthogonal sets of radio resources, i.e. OMA. The OMA can be very suboptimal in spectral efficiency because the set of allocated resources to

3

non-scheduled traffic could be unused most of the time due to the bursty nature of such safety-critical messages [1]. On the other hand, insufficient resource allocation to this traffic type results in reliability degradation. In order to compensate spectral inefficiency of OMA, NOMA multiplexes different traffic types over a set of shared resources to overcome spectral inefficiency of OMA. However, keeping isolation among URLLC services with NOMA is challenging due to potential temporal overloading in one service. Here, we consider a hybrid of OMA/NOMA, namely hyrid multiple access (HMA), in which, frequency resources, are divided into three groups: (i) resources allocated to serving the non-scheduled URLLC traffic; (ii) resources allocated to serving the scheduled URLLC traffic; and (iii) resources shared among them. It is clear that the HMA could be reduced to OMA and NOMA, and hence, it is the general form of multiple access coordination. Over the shared resources, we need to limit the aggregated received interference from the scheduled users. Furthermore, in order to prevent performance degradation for the scheduled traffic over the shared resources, we need to secure the reliability of data transmission even in presence of non-scheduled traffic. Towards this end, proactive or reactive strategies could be used. The former aims at making the transmission robust, e.g. by planning for occurring erasure in the channel (erasure coding), while the latter consists in cancellation of received interference from the non-scheduled traffic. Regarding the stringent delay requirement of URLLC, it is clear that the reactive strategy increases the delay violation probability for the scheduled traffic. Thus, risk-aware proactive RRM for minimizing reliability degradation is required. B. RRM for the Hybrid Multiple Access To make the HMA solution working, the RAN needs to control the traffic load dynamically, and adjust proactively the size of resource pools associated to each traffic stream. Furthermore, coordination among the BSs is needed to make sure the aggregated intra- and inter-cell interference will not cause an outage. Then, the RRM must be done in a multi-cell level to compensate the inter-cell interference impact on the outage probability. While a centralized solution seems promising for satisfying the reliability constraint even in cell-edges; the required signaling, including data gathering, forwarding to the center, and decision feedback, can potentially violate the URLLC delay constraints [9]. Even in a single cell scenario, due to the potential inter-dependence of the scheduled and non-scheduled traffic, keeping the outage risk for both traffic types lower than the URLLC requirement is not an easy task. All in all, regarding the high-mobility of users, their time/location-dependent distributions within the service area, and the bursty nature of the service requests, the RRM for assuring a minimum service requirement over a wide service area is highly challenging [5, 7, 8]. III. ML

IN

S UPPORT OF E FFICIENT RRM

By broadening the set of services and resources in 5G, control and troubleshooting of radio access network (RAN) by

human resources become too complicated and costly [7]. Furthermore, the stringent delay/reliability requirements of newly established URLLC services make them more challenging. Moreover, many decisions taken in serving users’ requests in legacy networks have temporal/spatial correlations, i.e. they are repeated in different locations and time instances. Then, it is wise to leverage the huge volumes of available data in cellular networks, and extract patterns of request arrivals, channel variations of users along mobility patterns, and outage statistics [7]. The derived patterns could be beneficial in reliable resource provisioning for future service requests. This brings the idea of leveraging ML tools in design and operation control of future wireless networks. ML aims at estimating a function, e.g. a rule or pattern, from a set of noisy data which has been generated by the true function. The recent advances in centralized and distributed computing, storage, and learning algorithms have improved the position of ML as a design and optimization tool significantly [9]. The Proposed ML-Powered RRM To manage network resources for serving scheduled and non-scheduled URLLC traffic, one needs to identify the set of dedicated resources to non-scheduled traffic, the set of shared resources, and the scheduling policy over the shared and dedicated resources. Once the set of resources and scheduling rules are determined, the scheduler allocates resources to users based on their rankings. The rank of each user depends on its delay budget and the expected level of received interference from the user over the shared resources. We propose to break the intelligent RRM problem into 3 sub-problems: radio resource provisioning (RRP), radio resource scheduling (RRS), and radio resource utilizing (RRU). Fig. 2 represents a graphical representation of the proposed hierarchical learning solution. In the proposed solution, a proactive resource provisioning based on training and prediction is done at the RAN control center. RRP at the RAN control center benefits from (i) transfer learning, i.e. making decision for a service area based on decisions taken in neighboring areas or previous time instances [10]; (ii) federated learning, i.e. leveraging the huge volumes of available records of requests/responses at the edges for extracting common rules and policy sets [9]; and (iii) risk-aware learning from past RRP decisions for outage minimization of coexisting services and traffic types. Distributing the learning task between BSs and the RAN control center aims at compensating partial observability of BSs in cellular networks. As BSs have access to local information, i.e. they observe only a limited part of the service area, and URLLC users could be highly mobile, resource provisioning at the BS-level could be less effective. For example, a BS cannot predict the arriving of a vehicle platoon which is currently in the neighboring service area. Then, distributed learning can further enable transfer knowledge of arrival of such demanding users, and sharing of respective effective resource scheduling policies. As an example, such learning may result in the knowledge for the RAN control center to activate some dormant BSs to balance the nonscheduled and scheduled traffic serving responsibilities among BSs, and hence, to minimize the probability of outage.

4

Action: RRP: (i) inter-service/cell resource allocation, (ii) BSs’ activity management.

Learning: (i) risk-aware online learning for RRP update, (ii) inter-BS transfer learning for RRS, (iii) federated learning of mobility pattern, etc.

Action: RRS: (i) inter and intra service resource scheduling.

Learning: (i) risk-aware online learning for RRS update, (ii) regression-based user/radio map update.

Action: RRU: (i) intelligent resource utilization for data transmission and reception.

Learning: (i) risk-aware online learning for RRU update.

RAN control center

Fig. 2: Distributed ML for radio resource management. The proposed RRM includes (i) RRP at the radio access network (RAN) control center, (ii) RRS at the BSs, and (iii) RRU at the end-devices.

Furthermore, in the proposed intelligent RRM solution, BSs decide how to control access of the scheduled users over the shared and dedicated resources to minimize the queuing delay violation for them, while complying with the maximum allowed interference constraint over the shared resources. This distributed control allows BSs to react to instant changes in the channel quality of connected users, and schedule them based on their remaining delay budgets. Finally, the intelligent usage of radio resources for reliable data transmission/reception is done at the device side. Devices can decide how to allocate their power resources to the set of shared and dedicated frequency resources to ensure the highest possible reliability. This is especially the case in non-scheduled URLLC transmissions, either to the BS or to another device/vehicle, where devices can benefit from reinforcement learning for improving reliability of their transmissions [11]. The proposed RRM requires a set of ML solutions at the RAN control center, BS, and device-side for addressing the challenges posed in serving the URLLC traffic, as described in Section I. In the following section, we describe some of these solutions, as well as the open problems. These solutions have been also depicted in the learning boxes of Fig. 2. IV. ML S OLUTIONS E MBEDDED IN THE P ROPOSED RRM A. Federated Learning of User and Radio Maps for RRP It is clear that the RRP for the URLLC traffic at the RAN control center requires knowledge about the outage risk of potential users. When we consider no a priori information about the users, the design is based on the worst case scenario, and hence, huge volumes of frequency resources are required to satisfy a level of QoS. Here, we investigate the combined use of radio and user maps in achieving an estimate of the outage risk of URLLC users. 1) Radio Map: They are expected to be an essential 5G component for enabling agile yet efficient resource allocation for mobile use cases like autonomous driving [12]. For a given

service area, radio map includes information like radio signal strength, delay spread, and coexisting interference level. Regarding the inevitable changes in the environment, radio maps need to be updated once a time. In the proposed hierarchical RRM solution, this is the role of the RAN control center to coordinate continuous update of the radio map, and to manage location-based inter-update time intervals. Furthermore, there might be time periods during which, no update of signal strength from a specific region is sent to the BSs. This could be due to temporal absence of users in a location for an extended period of time. In such cases, this is the role of the involving BSs to predict and update the received signal strength in that location, e.g. by leveraging temporal/spatial regression as described in [12]. 2) User Map: With the ever increasing number of connected devices and the volume of online data, the locations and mobility patterns of many vehicles, specially public transportation systems, are known or could be learnt [13]. Further than offline learning of mobility patterns, the RAN control center can coordinate information exchange among BSs in a service area to let them collaboratively construct and update a common user map. Having a map of users and mobility patterns, enables prediction of presence of users in different regions of the service area for an extended period of time [13, 14]. The map of users requires continuous update to be effective in URLLC applications. In our proposed structure, this is the role of the RAN control center to keep the consistency of the global user map, and coordinate the inter-update time intervals among the involving BSs. Once the position/speed/direction information of a user is known, and the radio map of the respective environment is given, we are able to estimate path-loss and shadowing components of its wireless communication channel. Then, the uncertainty of the channel scales down to the small-scale fading [14, 15]. Thus, we will be able to predict the received power at the BS as a function of transmit power of users.

5

Within the context of URLLC, this type of information is useful for early identification of a user which is expected to enter a crowded/low-coverage area. Thus, the RAN would be able to configure the network -manage the radio/energy resources- to lower the risk of communication outage for the respective users. The radio maps will further enable user management based on regional characteristics. This means that once devices enter a specific region, they could be granted higher-level support, e.g. dual connectivity, or extra resources for their transmissions could be reserved, or their non-scheduled transmissions could be handed over to multiple neighboring BSs. B. Risk-Aware Learning for RRS Introduction of URLLC into cellular networks mandates a transition from average-utility network design into risk-aware network design, where a rare event may incur a huge loss [2]. Instead of legacy ML algorithms, which aim at maximizing the average return of agents, or equivalently minimizing the average regret of agents, we aim at minimizing the risk of huge loss. The loss could be defined in different ways as a function of the return. When serving solely one traffic type, the riskaware learning-powered scheduling enforces maximizing the first moment of the return, i.e. reliability, while minimizing its higher moments, e.g. the variance. Minimizing higher moments of the return, i.e. optimizing the worst case return, could be achieved by defining the utility function as a nonlinear function of the return, R. A well-known example is exponential utility function in which, instead of maximizing the expected value of R, the objective is to maximize the expected value of exp(-R) [2]. In our proposed RRM solution, scheduling of shared radio resources is done by leveraging risk-aware learning at the BSs. Here, the scheduling task is performed such that the risk of outage for scheduled and nonscheduled traffic remains under the URLLC requirement. To exemplify the implementation of risk-aware MLpowered scheduling in our solution, let us present our solution in the context of Q-learning as a simple learning algorithm. In Q-learning, the long-term payoffs of state-action pairs are estimated by experiencing different actions in each state, which is achieved by following an ϵ-greedy policy. We transform the legacy Q-learning algorithm to a risk-aware learning algorithm by defining internal and external reward functions. The former, could be defined as the reliable data rate achieved by allocating a shared resource block to a scheduled user, and choses nonnegative values. The latter indicates the risk level of the non-scheduled traffic once the scheduling to the scheduled traffic is done, and can choose negative (for high risk) and positive (for low risk) values. The overall reward is defined the product of the two rewards. Then, the higher the risk of outage to the non-scheduled traffic, the overall reward will cause further regret to the system. Now, one can define a 3D scheduling matrix, i.e. the policy, and map the state and action concepts in the Q-learning to the respective items in the scheduling problem. The first dimension of the scheduling matrix corresponds to the set of states, and is mapped to the set of available shared resources in the scheduling problem. The

second corresponds to the set of actions, and is mapped to the set of quantized path-loss values of the scheduled users. The third corresponds to the quantized channel quality of the target non-scheduled URLLC user. For example, (x, [y1 , y2 ], [z1 , z2 ]) refers to scheduling a user with path-loss in [y1 , y2 ] dB, when there are 2 remaining shared resources and the predicted pathloss of the worst-case URLLC user with non-scheduled traffic is in [z1 , z2 ] dB. It is clear that for decreasing the risk of outage for the non-scheduled traffic, BSs prefer to schedule users experiencing higher path-loss values. By following this setup, it is straightforward to construct scheduling policies specific to the serving area of a BS by leveraging available training data at the RAN control center. Then, the respective BSs can schedule users based on these simulator-generated scheduling policies, and update them continuously. C. Transfer Learning for RRM Transfer learning is defined as reusing the knowledge achieved in solving of previous tasks, for speeding up solving of future problems [10]. In practice, this is a natural behavior of our brains. They are able to construct shared representations of challenges and strategies encountered in everyday problems, and reuse them in future problems to speed up finding of optimal solutions [10]. In cellular networks, many decisions in design, operation optimization, and troubleshooting are repeated in different locations and periods of time. Transfer learning allows spatial/temporal reuse of the learnt strategies, in order to enhance the learning process. One example of transfer learning is to leverage a fast yet comprehensive networks simulator at the RAN control center for training scheduler of BSs, which experience changes in their respective service area. Further forms of learning consist in exchange of decisions, strategies, and patterns among BSs in a service area. For example, one may consider the inter-dependence between arrivals of non-scheduled and scheduled service requests. When there is such interdependence, the arrival of a nonscheduled transmission may increase/decrease probability of receiving scheduled service request, and vice versa. Once such interdependence is learnt in part of the service area, e.g. for a platoon of vehicles, this information is transferred to involving BSs in the direction of the respective users. Another transfer learning example consists in preventing overloading in the service area, before it happens, based on the overloading events in other locations or previous time instances. In this case, once the control center anticipates arrival of URLLC requests, e.g. based on temporal/spatial correlation in the request arrivals, it can decouple the non-scheduled and scheduled resources to trade spectrum efficiency for achieving reliability. Decoupling the traffic could be done by handing over the non-scheduled traffic serving responsibility to the dormant neighboring BSs, and allocating all radio resources to the scheduled URLLC traffic where needed. D. Open Challenges and Potential Solutions The main challenge in utilizing intelligent RRM consists in large dimensionality and complexity. The heterogeneity of services, QoS requirements, and diverse set of available

2.5

×10 6

Outage probability for non-sched traffic

0.9999-reliable data rate (b/s) for sched.

6

2

1.5

1

0.5

No-learning (conservative design) Intelligent RRP & RRS with OMA Intelligent RRP & RRS with HMA

0 1e-1

1e-2

1e-3

1e-4

1e-5

1e-6

1e-7

Target outage probaility for non-scheduled traffic

(a) Intelligent RRP and RRS: reliable data rate for the scheduled traffic versus the outage probability for the non-scheduled traffic (single-shot transmissions). Increase in the required reliability highlights the merits of intelligent RRM and HMA.

10 -2

10 -3 PSD of Interference= [0,-164,-164] dBm/Hz

10

-4

10 -5 PA= [1,0,0]; OMA PA= [α, 1-α ,0] PA= [α, (1-α)/2, (1-α)/2] PA= [α, 1-α, 0] PA= [α, (1-α)/2, (1-α)/2]

10 -6

10 -7

PSD of Interference= [0,-172,-172] dBm/Hz

0

0.2

0.4

0.6

0.8

1

α : fraction of allocated power to RRB1

(b) Intelligent RRU: outage probability for different power allocation (PA) decisions. PA=[α, x, y] represents a decision in which, α fraction of transmit power is used for the dedicated RRB, and x and y fractions are used for the other two shared RRBs.

Fig. 3: Performance evaluation of the proposed RRM

resources make the RRM a very complex problem. This complexity has a significant impact on the convergence time of the learning algorithms. Furthermore, the high mobility of users demanding URLLC service, especially in the ITS use case, and the partial observability of the network by each BS, put further constraints to the RRM problem. Regarding the delay and complexity constraints, the role of on-device and edge learning for decision making must be highlighted. In order to further overcome the shortcomings of ML in timely finding optimal decisions, it is important to compare it against the way human beings learn. Human beings benefit from planning and reasoning for reducing the search space and explore the solution in most probable areas. In contrast, ML approaches treat the search space unbiased, and in the exploration phase, select all actions randomly to get an estimate of their respective payoffs. One way to reduce the search space, and speed up the learning process, is to add smart planning to the learning algorithms. Towards this end, one may benefit from the human-resources, and integrate them in the learning loop, to reduce the search space, and prevent BSs in taking risky/time-consuming actions during the exploration process.

traffic, and the learnt risk of outage in the service area, the ML-powered RRP and RRS modules are implemented and used for resource management, as described in Section IV. Fig. 3(a) represents the level of reliable data-rate achieved for the scheduled traffic, versus the required outage probability for the non-scheduled traffic. The dashed-curve represents a conservative RRM solution in which, resource allocation is done based on the risk to the cell-edge user. For easier interpretation of the results, the decoupled gains of RRM with OMA and HMA have been presented. The former represents the achieved reliable data rate by orthogonal allocation of resources to scheduled/non-scheduled traffic based on the online outage risk-level, in contrast with the conservative design (the dashed curve). The latter represents the achieved reliable data-rate by reuse of resources for scheduled traffic when possible. One further observes, the higher the outage risk, the higher room for spectrum efficiency by intelligent RRM. Furthermore, the benefits of HMA increase by increase in the required reliability level. Finally, we observe that uncertainty in the wireless channel, as well as user demand, mandates reserving more radio resources for guaranteeing reliability performance, which on the other hand reduces the overall spectrum efficiency.

V. ML-P OWERED RRM: A C ASE S TUDY

Now, let us examine the benefit of intelligent radio resource utilization. We assume a typical device with non-scheduled URLLC traffic experiences a distance-dependent path-loss of -80 dB, and needs to combat small-scale fading as well as path-loss for achieving reliable communications. Out of 5 RRBs, 3 could be used for non-scheduled transmission of this device, where the power spectral density (PSD) of scheduled traffic’s interference over them is denoted by [0, -172, -172] dBm/Hz. In other words, the first one is dedicated and the other two are shared with the scheduled traffic. Fig. 3(b) represents the outage probability versus the fraction of allocated power to the dedicated RRB, i.e. α. One sees that the best action is to distribute power equally on all RRBs. This is because the PSD of interference is comparable with the PSD of noise, i.e.

In this section, we demonstrate how to take advantage of intelligent RRM for spectrum-efficient yet reliable serving of scheduled and non-scheduled URLLC traffic. First, we investigate performance of intelligent RRP and RRS (Fig. 3(a)). Consider a single-cell scenario in which URLLC users generate and transmit non-scheduled URLLC traffic (infrequent) as well as scheduled URLLC traffic. The minimum and maximum experienced path-loss in the service area are denoted by -70 and -120 dB, respectively. We assume up to 5 radio resource blocks (RRBs) could be used for serving the non-scheduled URLLC traffic. Based on the outage risk of the most critical device with potential non-scheduled URLLC traffic, the required outage probability for the non-scheduled

7

-174 dBm/Hz. On the other hand, one observe that when the PSD of interference is [0, -164, -164] dBm/Hz, i.e. interference is much stronger than the noise, the best action is to use half of the power over the dedicated RRB, and divide the other half among the two shared RRBs. In both cases, one observes that an intelligent selection of α, could bring a huge gain for HMA in comparison with the OMA, i.e. transmission over the dedicated interference-free RRB (square-marked line). VI. C ONCLUSION Regarding the stringent reliability and delay requirements of URLLC, there is still lack of consolidated physical and MAC layer solutions enabling the URLLC service. One major challenge consists in addressing URLLC’s scalability and coexistence with other URLLC/non-URLLC services. Here, we have investigated coexistence of non-scheduled and scheduled URLLC services, and discussed challenges to be addressed for satisfying their stringent requirements. Then, we have proposed a hybrid multiple access solution for addressing such issues in a spectral/energy efficient way. In the heart of our proposed solution, we have further leveraged a distributed hierarchical ML approach for proactive RRM to different URLLC traffic streams. Finally, we have provided a case study to demonstrate the potential of leveraging ML in serving coexisting URLLC traffic over limited radio resources. The results indicated the gains that could be achieved by utilizing the proposed RRP, RRS, and RRU modules in comparison with the conservative non-learning operation. ACKNOWLEDGMENT This work is supported in part by the EIT Digital Project ACTIVE (Advanced Connectivity Platform for Vertical Segments), and in part by the Celtic Plus Project SooGreen (Service Oriented Optimization of Green Mobile Networks). R EFERENCES [1] P. Popovski, et al., “5G Wireless Network Slicing for eMBB, URLLC, and mMTC: A Communication-Theoretic View,” arXiv: 1804.05057, Apr. 2018. [2] M. Bennis, et al., “Ultra-reliable and low-latency wireless communication: Tail, risk and scale,” arXiv: 1801.01270, Jan. 2018. [3] I. Parvez, et al., “A survey on low latency towards 5G: RAN, core network and caching solutions,” IEEE Communications Surveys and Tutorials, vol. 10, no. 2, pp. 36-49, May. 2018. [4] H. Ye, G. Li, and B. Juang, “Deep Reinforcement Learning based Resource Allocation for V2V Communications,” arXiv: 1805.07222, May. 2018. [5] A. Kasgari, et al., “Stochastic optimization and control framework for 5G network slicing with effective isolation,” IEEE Conference on Information Sciences and Systems, 2018. [6] H. Ji, et al., “Introduction to ultra reliable and low latency communications in 5G,” arXiv: 1704.05565, Apr. 2017. [7] F. Calabrese, et al., “Learning Radio Resource Management in 5G Networks: Framework, Opportunities and Challenges,” arXiv: 1611.10253, May. 2018 [8] C. She, C. Yang, and T. Quek, “Radio resource management for ultrareliable and low-latency communications,” IEEE Communications Magazine, vol. 55, no. 6, pp. 72-78, June. 2017. [9] S. Samarakoon, et al., “Federated Learning for Ultra-Reliable LowLatency V2V Communications,” arXiv: 1805.09253, May. 2018. [10] B. Spector and B. Serge, “Sample-Efficient Reinforcement Learning through Transfer and Architectural Priors,” arXiv: 1801.02268, Jan. 2018.

[11] A. Azari and C. Cavdar, “Self-organized Low-power IoT Networks: A Distributed Learning Approach,” IEEE Global Communications Conference, arXiv: 1807.09333, Jul. 2018. [12] V. Chowdappa, et al., “Distributed Radio Map Reconstruction for 5G Automotive,” IEEE Intelligent Transportation Systems Magazine, vol. 10, no. 2, pp. 36-49, Apr. 2018. [13] R. Taranto, et al., “Location-aware communications for 5G networks,” IEEE Signal Processing Magazine, vol. 31, no. 6, pp. 102-112, Oct. 2014. [14] L. Muppirisetty, et al., “Proactive Resource Allocation with Predictable Channel Statistics,” arXiv:1806.05005, Jun. 2018. [15] R. Alieiev, et al., “Predictive Communication and its Application to Vehicular Environments: Doppler-Shift Compensation,” IEEE Transactions on Vehicular Technology, May. 2018.