QoE-driven Resource Allocation for Live Video

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2882441, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000. Digital Object Identifier ...

QoE-driven Resource Allocation for Live Video Streaming over D2D-underlaid 5G Cellular Networks JIHYEOK YUN1 , MD. JALIL PIRAN2 (Member, IEEE), and DOUG YOUNG SUH.1 , (Member, IEEE) 1

Department of Electronic Engineering, College of Electronics and Information, Kyung Hee University, Yongin, Korea (e-mail: [email protected], [email protected]) 2 Department of Computer Science and Engineering, Sejong University, Seoul, Korea (e-mail: [email protected])

Corresponding author: Doug Young Suh (e-mail: [email protected]) and Md. Jalil Piran (e-mail: [email protected]). "This research was supported by Korea Electric Power Corporation. (Grant number:R18XA02)"

ABSTRACT Recently, device-to-device (D2D)-underlaid fifth generation (5G) cellular networks have received plenty of attention because of their ability to save network resources and reduce energy consumption. Most existing algorithms for multimedia services over D2D networks consider only the signal-to-noise ratio (SNR) and ignore temporal requirements, which do not provide optimum performances. To overcome this issue, we propose a framework for a cross-layer D2D link control system, which guarantees the quality of service (QoS) and improves the quality of experience (QoE) for live video streaming with different priorities and delay constraints. In this framework we considered three techniques, including priority-based video transmission, flexible communication mode switching of user equipment (UE), and subset-based relay assignment. According to the live video generation period, our system dynamically adjusts the ratio between cellular and D2D mode durations in each unit of the communication period for each user individually. Our proposal also considerably reduces the duration and frequency of video playback freezing by delivering at least the minimum service quality of the delivered video to the sink users even in the shadow area and thereby improves the QoE for all users while minimizing energy consumption. System-level simulation shows that the proposed algorithm outperforms other methods in terms of the average mean time to failure (MTTF), average peak signal-to-noise ratio (PSNR), and average energy consumption. INDEX TERMS Device-to-device communication, resource allocation, live video streaming, QoS, QoE.

I. INTRODUCTION

In D2D networks, devices communicate among each other directly without network infrastructures, such as an access point (AP) or base station (BS). Since the development of fourth generation (4G) long-term evolution-advanced (LTEA) technologies, cellular networks have been supporting D2D communications, even in licensed networks. Meanwhile, in unlicensed networks, D2D technologies are widely used in Bluetooth and Wireless-Fidelity (Wi-Fi) Direct. With significant advances in wireless access technologies and the rapid proliferation of media-rich devices and applications, there has been a fundamental change in wireless network traffic. According to a forecast by CISCO, video traffic has already surpassed 55% of wireless network traffic and is expected to reach 70% by 2019 [1]. Moreover, expectations are changing more dramatically because of the popularizaVOLUME x, 20xx

tion of smart cellular phones, surveillance services, and video communications on vehicles [2]. Additionally, the demand for sharing videos is rapidly increasing. Therefore, cellular networks are confronted with problems when handling huge traffic and guaranteeing acceptable QoS and QoE for realtime video services. A practical solution for this problem is taking advantage of the proximity-based D2D underlaying cellular communication paradigm, in which devices share a licensed spectrum for D2D communication without interfering with cellular communications. Because of the geographical proximity, D2D communications increase the network capacity and are convenient for context-aware applications and energy saving. Owing to these benefits, mobile operators have deployed D2D communications in 4G LTE-A in the 3rd Generation Partnership Project (3GPP) Release 12. Furthermore, D2D communications are considered essential parts of 1

2169-3536 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2018.2882441, IEEE Access Jihyeok Yun et al.: QoE-driven Resource Allocation for Live Video Streaming ...

5G cellular networks since the number of devices and users, along with their mobile traffic, will significantly increase [3]. In contrast to conventional D2D techniques, such as Bluetooth and Wi-Fi Direct, network-assisted D2D technology operates on licensed cellular networks, for which various useful algorithms for network channel reuse and controlling interferences have been developed; these operate in the channel-reuse mode [4]–[6]. In [4], an overview of networkassisted D2D communication, with respect to the evolution of wireless networks from the first generation (1G) to 5G, has been provided. The authors surveyed emerging research topics and the resource allocation algorithm for D2D-underlaid cellular networks. Notably, the channel-reuse scenario with interference management received plenty of attention and was analyzed for various network environments. Similarly, in [5], the authors surveyed the features of existing technologies for different research objects, such as throughput enhancement and saving energy. In particular, they analyzed channel reusing-based cooperative D2D communications and provided a network model for the D2D communication type. As discussed in [5], the commercialized standard for LTE-A including network-assisted D2D and relay-based D2D technologies on cellular networks is promising for future cellular networks with remarkably improved resource efficiency. On the other hand, Vo. et al. [6] presented a video streaming service approach using D2D communication for 5G networks as a target scenario and introduced an applicationaware optimal rate allocation method. In [4], the authors introduced the effect of the evolution of 5G cellular networks with D2D communication and elucidated topologies and types of D2D communication that are expected to be used in the future. As claimed in [6], 5G cellular networks that support a massive number of devices at high density, flexible spectra, bandwidth reuse, and high reliability, will be commercialized. Therefore, research into communication technologies reflecting characteristics of 5G networks are receiving considerable attention [7], [8]. In particular, the realization of real-time video delivery services using D2D communication is expected with the emergence of 5G cellular networks. Recently, many studies have focused on determining the optimal allocation of resources to improve the QoS/QoE of the video service, which is sensitive to delay and stability, using D2D communication. These studies include content dissemination using D2D vehicle-tovehicle (D2D-V2V) communications [9], D2D autonomous driving control for cooperative intelligent transportation systems [10], and cost-efficient real-time video streaming over D2D communications [11]. As described in this section, research on D2D communication has enabled highly-efficient utilization of network resources in dense 5G networks while also enabling video services over cellular-D2D networks. However, since the amount of traffic flow of real-time and live video services is increasing explosively, there is significant demand for D2D management systems that provide service-specific solutions to guarantee users’ QoS/QoE. Therefore, this paper proposes 2

a novel relay assignment algorithm and streaming method for enhancing the QoE of live video services over 5G cellularD2D networks. The proposed algorithm is compared with conventional channel allocation algorithms, which do not consider the characteristics of live video traffic. The remainder of this paper is organized as follows. In Section II, we survey works related to video streaming or video distribution scenarios in cellular-D2D communication environments. Section III presents some challenges that need to be addressed for the live video relaying scenario and then briefly introduces our contributions. Section IV provides system scenarios for the cellular-D2D communication considered in this paper. Section V presents problem formulation with notations for the optimized D2D relay channel allocation algorithm. In Section VI, we explain our scenario in greater detail. In Section VII, the proposed system is evaluated in terms of the MTTF, PSNR, and energy consumption by using a system-level simulation. Finally, we draw our conclusions in Section VIII. II. RELATED WORK

In this section, we review some existing techniques that have been proposed for delay-sensitive video streaming over D2D communications. According to the similarities between our proposed method and these related works, we separate these studies into five different categories. 1) QoE-driven video streaming over D2D communication Vo et al. [6] introduced a scenario in which D2D devices cooperate to assist video content delivery of the BS in a 5G network. The proposed algorithms in [6] include D2D relay assignment and rate allocation methods for cooperative video transmission using D2D communication. These algorithms exhibit improved QoS/QoE of the video streaming services in terms of the 1) continuity of playback, 2) PSNR, 3) quality fluctuation for smooth playback, and 4) energy consumption overhead by the relay based on D2D communication. However, only the UE whose target content is already cached can serve as a relay node. Therefore, the scenario introduced in [6] targeted a video-on-demand (VoD) service. 2) Transmit power and distance between D2D pairs Improvement in resource utilization through the direct routing of D2D traffic was addressed in [12], [13], where the network efficiency was enhanced by utilizing resource allocation based on inter-recipient transmission. These studies focused on the throughput gain using power control under energy and total rate constraints. The power constraints included in these algorithms are factors that should be considered when devices with power limitations act as relays. 3) Cross-layer design for D2D link setup algorithm Research on the multi-source content-sharing system VOLUME x, 20xx



using cooperative relay communication is ongoing. The authors in [14]–[17] proposed D2D link assignment and channel allocation algorithms to maximize the energy/resource efficiency of a D2D relay-based content sharing system. They evaluated the performance in terms of the energy/spectral efficiency and reliability on the content-sharing scenario. However, since the proposed algorithms are cache-based, they cannot satisfy the delay constraints of an on-time live video content-sharing system. Xing et al. [14] proposed a system model for cooperative content distribution using multi-source D2D communication. Their method finds the D2D relay sets that minimizes the outage probability of content distribution and maximizes system throughput in a combined multi-source and channel reusing scenario. Alternatively, the authors of [15], [16] presented a one-source, multi-hop D2D communication method as a system model. The relay assignment algorithm, proposed in [15] improves the system throughput and energy efficiency. The authors of [16] proposed a relay assignment algorithm with the primary purpose of minimizing energy consumption. In [17], the authors considered delay constraints by using an efficient rate allocation algorithm for D2D communication between multi-sources and a destination UE. The algorithm that include a delay constraint, as shown in [17], is closer to the optimization method for real-time streaming than the algorithms in [14]– [16]; however, since these studies also assumed cached video, the induced delay is unacceptable for live video services. 4) Delay-aware D2D mode selection In [18], [19], the authors considered delay constraints to determine the optimal D2D and cellular link sets with minimal dropping and blocking probabilities. The dynamic mode selection and resource allocation algorithm in [18] enhanced the average end-to-end delay under a dropping probability constraint on orthogonal frequency-division multiple access (OFDMA) cellular networks with D2D communications. In [19], the resource control problem in D2D communications was addressed using the constrained Markov decision process (CMDP), which characterizes the dynamic interference between D2D and cellular links based on their varying backlogged states, the dynamic route selection, and the coupled interactions between uplink and downlink resource allocations. The above studies, however, did not provide QoS parameters which could be used to show the suitability of their algorithms for real-time video streaming. Additionally, Wang et al. [20] proposed game-theoretic rate control for real-time video relaying over 5G wireless networks while considering the QoS. They analyzed the correlation between the PSNR and energy VOLUME x, 20xx

consumption as a performance measure when delivering real-time video over relays. To consider realtime scenarios, they defined the deadline of each video frame as the time interval between the time the frame is transmitted from the source and the time the frame should be played on the sink UE. If a frame arrives at the sink UE after the deadline, it is useless and is dropped without any processing. In their algorithm, the two phases of cellular and D2D modes were alternated within a fixed duration. 5) Scalable video over D2D communication In [17], [21], the authors considered delay constraints in real-time video relaying over cellular-D2D communications. For QoS adaptation to the time-varying channel condition, they used scalable video coding (SVC)-encoded video, in which video is multi-layered. In [20], the authors used temporal scalability. Alternatively, in [17], [21], the authors used spatial scalability, which requires about 25% more computation and about 10% more bitrate than temporal scalability. However, the dynamic bitrate range of the video with spatial scalability is several times wider than that of temporal scalability. A wider dynamic range helps satisfy the tight delay constraints of live videos, even under harsh conditions. However, SVC-based spatial scalability does lead to a disadvantageous rate overhead. In comparison to spatial scalability, temporal scalability is more accessible to adaptive video streaming systems but has less adaptability because of the lower dynamic bitrate range [22]. Thus, the temporal scalability that causes intermittent on-time playback failure significantly degrades the QoE, as compared to the spatial method, which guarantees seamless playback of a low-quality video layer. III. PROBLEM STATEMENT AND OUR CONTRIBUTIONS

In the previous section, we summarized techniques that are similar to our proposed scenario. In this section, we discuss the technical issues and challenges associated with those techniques in greater detail in order to state the target problem of this paper. 1) QoE-aware live video streaming with cross-layer design Vo et al. [6] evaluated the standard deviation of the PSNR and the enhanced average PSNR to demonstrate the QoE improvement for their channel allocation algorithm. Their algorithms and evaluation methods showed excellent effects on D2D video streaming. However, since they considered cooperative video delivery with pre-defined users acting as relays, this scenario is similar to the low-latency VoD service scenario. The algorithms presented in [17], [21] include rate control with SVC, which supports the spatial scalabil3



ity of the video stream to satisfy the acceptable delay conditions and improve the quality of the real-time video. The authors used pre-cached relays to satisfying delay constraints of real-time VoD scenarios, but these relays could not satisfy the delay constraints of live video streaming scenarios. The scenario presented in [20] is closest to a live video streaming scenario. However, in the method proposed in [20], the D2D and cellular phases’ duration are fixed; this situation has less adaptability to the timevarying network condition than the flexible-duration phase method. However, the flexible phase duration has disadvantages regarding the resource efficiency, as compared to the fixed-duration method; this is the case because it involves unicast communication. This paper overcomes these disadvantages by using an efficient D2D link and resource management, improving the QoE of the live video streaming service. This paper focuses on the relation between energy consumption and the QoS/QoE of live video streaming over cellularD2D communications with flexible phase duration and spatial scalability of the video. 2) Evaluation parameters In [12], [13], evaluation parameters for relay-based cellular-D2D communications were simulated and concentrated in terms of the energy efficiency and enhanced throughput. In [18], [19], channel allocation algorithms considered delay constraints and evaluated the D2D communication performance in terms of the call blocking and dropping probability. Such QoS parameters are sufficient to show improved cellularD2D system performance for general traffic-targeted scenarios. However, these QoS parameters are not good enough to show the QoS of video streaming services. The studies in [6], [17], [20], [21], which considered seamless video streaming services in cellular-D2D communication systems, used the PSNR to evaluate the QoS. Then, they claimed that the QoE evaluation method considering all the PSNR, network resource efficiency, and energy consumption was enough to evaluate the QoE of video streaming services. However, these parameters were not enough to show the QoE intuitively, especially in delay-sensitive live video services. An important factor in the QoE evaluation of live video streaming services is related to seamless ontime replaying in the sink UEs. Therefore, intermittent on-time video replaying failures should be represented in the assessment. However, if on-time replaying fails, the PSNR values at those frames are non-zero values due to the similarity of successive video sequences. For this reason, it is not reasonable to evaluate on-time replaying failures using only the PSNR in the live video streaming service. Therefore, it is necessary to assess the time intervals of the failures in order to evaluate the seamless minimum quality guarantee of the video 4

service, while also showing the average video quality trend of the service using the PSNR. In this section, we present our contributions that help address the issues discussed above. 1) Priority-based flexible D2D mode control for QoEaware live video streaming We propose a type of flexible cellular-D2D mode control for non-pre-cached live video streaming, where all UEs participating in the system share content with each other while receiving content in real time for their own consumption without cached content. The delay constraints in our proposed method are applied to each user with independent conditions such as the channel states of relay candidates and available video layers. A BS carries out the channel allocation algorithm with the expected bitrate of the channel by using the signal-to-interference-plus-noise ratio (SINR) between the relay and the destination. The algorithm is repeated slot-by-slot until the unit period expires. If it is determined that more than one video layer can be transmitted using the expected bitrate during the time available for D2D communication, the target channel may be reused for the D2D pair. The proposed system determines the D2D pairs by considering the individual D2D allowance time of each UE and the channel condition. Then, each D2D pair performs unicast relaying. Energy overhead caused by the use of unicast relaying, instead of multicast relaying, can be overcome due to the temporarily optimized transitions in D2D communication. Scalable extension of the high-efficiency video coding (SHVC)-encoded traffic supports the availability of the dynamic-communication mode transition with a wide dynamic bitrate range of the video [23]. In the case of high-efficiency video coding (HEVC) [24], the length of cellular and D2D phases should be the same because the amounts of video traffic transmitted through the BS and the relaying D2D UE are the same. However, in the case of SHVC, the flexible length of the phases can support adaptive relay allocation. Layered video traffic, such as SHVC [25], has more traffic overhead than HEVC-encoded traffic. However, as shown in the simulation results of [17], an adaptive traffic size can derive enhanced system performance because the minimum rates of traffic replaying the video are spread out with high priority. Therefore, we propose layered video transmission ordering for priority-based D2D channel allocation. The priority between the layers is evident because the lower layer is needed to replay the upper layer. The lowest layer, which is the base layer (BL), is capable of replaying the video with the lowest bitrate and lowest quality. Therefore, we perform reordering according to the transmission priority of the layers. It is possible to play back a low-bitrate, low-quality video preferentially and VOLUME x, 20xx



then allocate extra network resources to high-bitrate, high-quality video transmission to prevent freezing of a live video replaying over D2D relaying with video layer reordering in a unit period. We assume that the duration of unit communication period follows the duration of a group of pictures (GOP, which is a videoencoding period that can be replayed) of the contents. 2) Subset-based relay assignment algorithm for maximizing evaluation parameters We perform a system-level simulation to calculate the average energy consumption of all UEs in the D2D clusters. In the simulation, the radius of the cluster varies widely. Then, we observe the trend of the QoS/QoE parameters (i.e., the PSNR and the MTTF) over the simulation. If video data packets arrive later than the expected time, or if the lost packets are not received on time, video playback is frozen. How often the video playback freezes is a critical measure of video quality. In this paper, a freezing event is considered to be a failure and the MTTF is used as a significant quality measure. Therefore, we evaluate the QoE performance of the proposed algorithm using the MTTF. The MTTF is not a characteristic parameter used for assessing video quality such as VOD. However, in the live video case, determining how often this failure occurs is truly meaningful when evaluating the QoE. To maximize the performance of the evaluation parameters, we define subsets for the management of all UEs participating in the live video relaying system and propose a service-based optimized relay assignment algorithm. As mentioned previously, the contributions enable quality-effective D2D-based live video playback; however, it also leads to some disadvantages, such as energy consumption overhead due to the rate overhead of layered video coding and the flexible length of phases. The subset-based relay assignment algorithm performs fine-grained UE mode control in units of a timeslot and can improve the energyefficiency aspect compared to the full search method; this is the case because it controls the subset unit according to the state of the UEs. The algorithm performs relay assignment based on the transmissionavailable video layer prediction to reduce the traffic overhead.

BS Interference Communication

D2D UE

We consider a cellular channel allocation scenario for live video transmission over D2D-underlaid 5G cellular networks, as shown in Fig. 1 [13]. Fig. 2 illustrates the conceptual topology of the D2D relaybased video relaying system with SHVC. Fig. 3 shows the actions of the UEs in the proposed system for layered video relaying in a unit communication period in the same situation as the one shown in Fig. 2. VOLUME x, 20xx

Cellular UE

D2D UE D2D UE

FIGURE 1. System scenario of relay-assisted D2D communications over cellular networks.

A. SYSTEM SCENARIO: ENTITIES

In our scenario, the cellular UEs (CUEs) receive video traffic directly from the multicasting BS using their cellular channels. When in an idle state, each CUE can play the role of a D2D destination UE (DUE) or relaying UE (RUE). The D2D pair (RUE-DUE) can reuse the cellular channel for D2D video communication, which avoids interference with the CUE and enhances network/energy resource efficiency. The DUE receives content indirectly from a peer RUE via the D2D link. B. SYSTEM SCENARIO: PHYSICAL LAYER (ORTHOGONAL SPECTRUM RESOURCE ALLOCATION)

In our scenario in the D2D-underlaid cellular networks, the potential DUEs can operate in two modes: D2D and cellular modes. In the D2D mode, DUEs in proximity with one another can directly communicate, instead of communicating through the BS, by sharing the orthogonal network resources of the CUEs. We consider an SHVC content-sharing scenario in which all UEs try to receive SHVC content over D2D-underlaid cellular networks in a single cell. We assume that each CUE has an allocated orthogonal network resource, which ensure that there is no interference between cellular channels. However, cellular channels may encounter interference when reusing the channel of D2D links. Therefore, channel allocation should consider the SINR to allocate a channel to the D2D link. As explained in [5], the SINR of the CUE in this system is affected by the interference from all D2D pairs reusing the channel of the CUE. (1) is the equation used to calculate the SINR of CUE i with interference from D2D pair j. Here, j is the D2D pair where RUE j helps DUE k. SIN Ric (t) =

IV. SYSTEM SCENARIO

D2D UE

Cellular UE

N0 +

Pi GiB (t) , j∈R Pj Gjk,i (t)θij (t)

P

(1)

where θij indicates whether the D2D link for j reuses the channel of i (when the channel is reused θij is equal to 1, and it is 0 when it is not reused); j 6= k; Pi is the transmitting power of CUE i; Pj is the transmitting power of the D2D pair; GiB (t) is the link gain between the BS and i; Gjk,i (t) is the gain of the D2D links which act as interference source for the cellular link of i; R is the set to which all j’s belong, j ∈ R; and D is the set to which all k’s belong, k ∈ R. 5



CUE

EL#2

CUE RUE

EL#1 Control (Optimized relay set)

BL

Streaming

DUE

Video relaying

Live streaming EL#2

Tp

EL#1

List of the helpers

Tp

Relay assignment

BL

List of the D2D requesters

Optimizing for relay assignment

Changing of the communication mode on UEs in Tp

BS

FIGURE 2. Proposed relay-assisted cellular-D2D communications for live video relaying scenario: Layered live video relaying.

Cellular mode D2D mode TP BL

EL#1

EL#2

UE1

given to interference caused by reusing cellular channels. The interfering factors in the SINR of the D2D pair are the CUE and other D2D pairs that simultaneously use a channel. The SINR of D2D pair j can be calculated as d SIN Rjk (t) =

N0 + Pi Gi,jk (t) + D2D link setup time

UE 2

UE 3

P G (t) P j jk

jx ∈R,jx 6=j

Pjx Gjx ,jk (t)θijx (t)

, (2)

where j and jx are both included in the relay set R (we use jx to represent j 6= jx ); j 6= k; Gi,jk (t) is the gain of the cellular link which acts as an interference source for D2D pair j; Pjx is the transmitting power of D2D pair jx ; and Gjx ,jk (t) is the gain of D2D pair jx which reuses channel i similarly to to pair j. The channel gain of the link is calculated [26] as

UE 4

G = 10−P LdB /10 . FIGURE 3. Proposed relay-assisted cellular-D2D communications for live video relaying scenario: Actions in the unit period.

Because the subject of this paper is based on D2Dunderlaid cellular networks, careful consideration should be 6

(3)

The total path loss is calculated using (4) [26], where Xu is the lognormal shadow fading path loss of the UE and α is the antenna gain. P LdB = ωdB (R) + log(Xu ) − αdB .

(4) VOLUME x, 20xx



The channel model used for the simulation is a WINNER II C1 non-line of sight (NLOS) path loss model (Suburban scenario) [27], [28]. The path loss is calculated as ωdB (R) = [44.9 − 6.5 log(hBS )] log(R) + 23.46 f [GHz] ), + 5.83 log(hBS ) + 20 log( 2

(5)

where hBS is the height of the BS, R is the distance between the BS and UE, and f is the carrier frequency. After obtaining the SINR of all available link candidates by the CUEs and DUEs in a cell, adaptive modulation and coding (AMC) should be decided, as introduced in [29], [30]. We determine d in that the bitrates of CUE i and D2D pair j are ric and rjk each specific duration Tc,i and Td,jk , respectively. Then, the available bitrate becomes time-varying, and the video bitrate should adapt accordingly. Since we consider a centrally controlled D2D system, the DUEs require communication for signaling with the BS. Therefore, the SINR according to the communication target of the DUEs should be considered. Asadi et al. [31] introduced various issues and subjects related to D2D communication in their research. Among the studies introduced in [31], the authors of [32], [33] conducted a study on centralized signaling message control for D2D communication in OFDMA systems for network-assisted device discovery. There are various ways to use assumptions based on the topic of each study in the scenario, where they reuse the control channels of CUEs (e.g., physical uplink control channels (PUCC) and sounding reference signals (SRS)) for signaling. Naslcheraghi et al. [32] did not include critical assumptions when conducting their studies in order to avoid collisions of the messages between CUEs and multiple DUEs. Additionally, Nguyen et al. [33] dealt with a field similar to the subject of [32] but employed the assumption that the messages for D2D communication were always delivered successfully without being influenced by the CUEs. Furthermore, in [6], where the subject matter is similar to that of this paper regarding 5G cellular resource allocation for D2D communication, it was assumed that the BS knows all of the information needed for the proposed resource allocation algorithm. This includes the information of the distributed video to all UEs, all possible channel information between the BS and UEs, all possible channel information between the RUE and DUE, and the gain of all links. Since both this paper and [6] studied resource allocation in D2D systems, we replace the part of the signaling process with the assumption made in [6]. C. QOS/QOE PARAMETERS FOR REAL-TIME SHVC TRAFFIC

SHVC is a newly standardized layered video codec for the H.265/HEVC standard [24]. In this study, we use an SHVCencoded video stream for bitrate adaptation of video content [23], [25]. A video with three spatial layers shows a dynamic bitrate range from 3 to 15 Mbps. The characteristics of VOLUME x, 20xx

layered video coding are independent of the decoding of cumulative streams. The upper layer of the streams is encoded by referring to the lower layers. The lowest layer can be decoded by itself. However, all upper layers can be decoded only when their lower layers are available. The number of layers to be transmitted is determined dynamically according to the available bitrate, and then, the quality of the decoded video is represented by the PSNR, as shown in [34]. However, even if intermittent playback failure occurs, the PSNR, which is not significantly affected by the failure, is not enough to evaluate the QoE. Recently, live and real-time video streaming has been studied as a use case in various environments, according to the recent development of the network environment. Although not a general metric, the failure of video frame playback is being used in a variety of ways to evaluate their research. The number of playback failures and the duration of the failures can be important evaluation methods depending on the target scenario or killer application. In [35], Cho et al. did not consider D2D communication but evaluated the performance of SVC-encoded video live streaming using MPEG media transport (MMT) and dynamic adaptive streaming over HTTP (DASH), i.e., the next-generation multimedia transmission standard in wireless networks. They referred to an event in which a video frame playback failed as a freezing event and evaluated the suspended time due to this type of event. Alternatively, to assess the performance of a system that transmits SVC-encoded traffic over unstable D2D communications, Zhang et al. [36] referred to the failure of video playback as a halt, and evaluated the number of halts and halt durations. In another example, Wu et al. [37] evaluated their algorithms using the PSNR and the frame success rate when sending live video for cloud gaming; this use case is very sensitive to video playback failure. They defined frame success as concurrently successful video frame receiving and playback. They also claimed that this is useful for assessing the performance of live video services. As described, freezing events occur when the on-time transmission fails until the playback deadline of the video frame. Therefore, we use two parameters, i.e., the MTTF and PSNR, to evaluate QoE-driven freezing-less live video relaying. In Section V, we explain the problem formulation used to maximize the QoS/QoE of the live video relaying over D2Dunderlaid cellular networks. V. PROBLEM FORMULATION

This paper considers a scenario in which the BS multicasts the same content to all UEs within a cell of a 5G cellular network. Based on the assumption in Section IV, the BS knows all the information related to resource and relay allocation for the UEs in the system. Thus, the BS can predict the SINR and the modulation and coding scheme-applied bitrate [29], [30] of all available links for every specific period by using the equations given in Section IV. For this prediction, we assume that the BS knows the predefined maximum Tx power of 7



TABLE 2. Subsets for sets C, R, and D.

TABLE 1. Notation used for problem formulation.

Symbol

Definition

Sets U

The set of all UEs participating in the system

C1

The subset of CUEs whose bitrate is enough to receive the BL during a unit period

C2

The subset of CUEs whose bitrate is enough to receive the BL and EL#1 during a unit period

ρ

The number of channels

Tp

Duration of the unit communication period

t

Current timeslot

χ

Timeslot duration

C

The set of all UEs in cellular mode

D

The set of all UEs in D2D mode

C3

The subset of CUEs whose bitrate is enough to receive all three layers during a unit period

Tc,i (t)

The duration available for the CUE i to receive selected video layers from the BS over the cellular network

R1

The subset of RUEs that receive the BL and are ready to send the BL, R1 ⊆ C1

T c,i (t)

The duration that the CUE i needs to receive scheduled video traffic from the BS over the cellular network

R2

The subset of RUEs that receive the BL and EL#1 and are ready to send them, R2 ⊆ C2

Td,jk (t)

The duration available for the D2D pair j (j helps k by relaying)

D0

T d,jk (t)

The duration that the DUE k needs to receive scheduled video traffic from the RUE j using D2D communication

The subset of DUEs that are predicted not to be able to receive the BL directly from the BS, but to receive it from any RUE

TsigC

The delay caused by signaling required for the video streaming service via cellular communication

D1

The subset of DUEs that receive the BL and are predicted not to be able to receive the EL#1 directly from the BS, but to receive the EL#1 from any RUE in R2

TsigD

The delay caused by signaling required for the video streaming service via D2D communication

RR1

The subset of RUEs acting as the relay node to relay the BL, RR1 ⊆ R1

TL

The D2D link-setup delays to make a connection between the RUE and the DUE

RR2

The subset of RUEs acting as the relay node to relay the EL#1, RR2 ⊆ R2

Tpro

The delay caused by the additional processing being performed in the RUE

DR0

The subset of DUEs receiving the BL from the RUE, DR0 ⊆ D0

DR1

The subset of DUEs receiving the EL#1 from the RUE, DR1 ⊆ D1

UE and performs power control through periodic signaling with UEs. This signaling is introduced in Section VI. Table 1 contains the notation for the problem formulation. We assume that the traffic rates of the BL, enhancement 1 layer (EL#1), and enhancement 2 layer (EL#2) in a unit period are r0 , r1 , and r2 , respectively. Then, the required rates for decoding three layers, c0 , c1 , and c2 , become c0 = r0 , c1 = c0 + r1 , and c2 = c1 + r2 , respectively. Using (6), at the initial time in every unit period, the system determines the maximum video layer that the CUEs can receive in a unit period without any interference from D2D pairs. At this point, the system performs categorization (using the subsets of C and D in Table 2) according to the maximum available video layer of each UE and performs channel allocation to receive available CUEs. Next, in a unit period, the ci (t) works as the bitrate threshold for the CUEs in the two-step optimization process for relay allocation in our algorithm. In the optimization process, the system allows for regulation of the transmit power of the CUEs for channel reusing by the D2D pairs. However, in all processes in the algorithm, the bitrate threshold should be guaranteed. The specific constraints are described in the two-step optimization process. li (t) = arg max [cl (t) ≤ ric (Tc,i (t), Pi )] , ∀i ∈ c,

(6)

l∈[0,...,L−1]

where Pi indicates the transmitting power for CUE i, and l denotes the layer of the video. li (t) is the selected video layer that can be transmitted without any flaws. ci (t) is the bitrate 8

Definition

of the video chunk that should be transmitted on-time in the unit period. After categorizing the UEs using (6), the algorithm performs the two-step optimization process to determine the optimal D2D pairs. In the algorithm, the priority of D0 is higher than that of D1, which means that all available channels are first allocated to D0 according to (7) and then to D1 according to (8). Delay constraints for the BL and EL#1 are considered in (7) and (8), respectively, so that decodable GOP data of a specified layer are received in the unit period during the slack time Td,jk (t). 0 wi,jk (t) = d {i, j, k | c0 (t) ≤ rjk (Td,jk (t), Pj , Pi Gi,jk , Pjx Gjx ,jk )},

i ∈ C, j ∈ R1, k ∈ D0, ∀jx ∈ R, (7) 1 wi,jk (t) = d {i, j, k | c1 (t) ≤ rjk (Td,jk (t), Pj , Pi Gi,jk , Pjx Gjx ,jk )},

i ∈ C, j ∈ R2, k ∈ D1, ∀jx ∈ R, (8) where Pj is the transmission power of the RUE, Pi Gi,jk is the interference from i, and Pjx Gjx ,jk is the interference from jx , as described in (2). 0 In (7) and (8), wi,jk has a value of 1 or 0 and denotes whether or not RUE j can relay the BL to DUE k, respectively. Additionally, W 0 is a set of the optimal D2D pairs for 1 all UEs j and k in R1 and D0. Likewise, wi,jk has a value of 1 or 0 and denotes whether or not RUE j can relay EL#1 to VOLUME x, 20xx



ric (Tc,i (t), Pi , Pj0 Gj0 k0 ,i ) ≥ ci (t), ∀i ∈ C, c Pi ≤ Pmax , ∀i ∈ C, d Pj0 ≤ Pmax , ∀j0 ∈ R1, X 0 wi,j (t) ≤ ρ, ∀k0 0 k0 j0 ∈R1

(10) (11) (12) ∈ D0,

W1

D1

D0 D (a)

R1

Already assigned for an element of D0 in step 1

D0

R2

R2

D1

(b)

(c)

FIGURE 4. Two-step optimization process for finding relay candidates. (a) Subset relations in the two-step optimization. (b) Step 1: Find the optimal D2D pair set W 0 . (c) Step 2: Find the optimal D2D pair set W 1 .

the live relay scenario, it does not consider a single channel assignment for multiple D2D links at the same time. 1 minimize{w1 (t) | wi,j (t) = 1}, 1 k1 Pi ,Pj1

(13)

where ci (t) is the bitrate of the video chunk that should be communicated on CUE i; ci (t) also works as the bitrate threshold of i; in (10). Additionally, Pj0 Gj0 k0 ,i is the interference caused by D2D pair j0 in (10) and ric (Tc,i (t), Pi , Pj0 Gj0 k0 ,i ) is the interference-considered bitrate of i. Constraint (10) prevents the bitrate degradation of the cellular link due to the D2D link. (11) and (12) are constraints limiting the transmitting power of the CUEs and RUEs in the system. Constraint (13) ensures that the number of channels used for D2D communication does not exceed the number of available cellular channels. Since this paper focuses on VOLUME x, 20xx

W0

...

subject to

R2

...

∀i ∈ C, j0 ∈ R1, k0 ∈ D0, (9)

R1

...

Pi ,Pj0

C3

...

0 (t) = 1}, minimize{w0 (t) | wi,j 0 k0

C

...

DUE k, while W 1 is a set of optimal D2D pairs for all j and k in R2 and D1. In our algorithm, (7) and (8) are required for determining the availability of the D2D pair candidates for the BL and EL#1 relaying. These determining functions are used when RUEs exist in subsets R1 or R2. In the following two-step optimization process, (7) is used to determine whether k can fully receive the BL when j and k, which can perform D2D communication during Td,jk (t), reuse the channel that i is using. (8) is used for determination for the EL#1 in the two-step optimization process. Fig. 4(a) describes the construction process of sets W 0 and W 1 . The two-step optimization process is repeated for a number of D2D pair candidates to determine the set of optimal relay pairs and optimal reusable channels. For distribution of the BL with the highest priority, W 0 is constructed by utilizing all available relay candidates. Fig. 4(b) illustrates the investigation for optimized D2D pairs that can be joined in W 0 , while W 1 is constructed by utilizing the candidates that can relay EL#1 among the remaining UEs, as illustrated in Fig. 4(c). In the two-step optimization process, j0 and j1 are used to distinguish the RUEs that relay the BL or EL#1, and k0 and k1 are used to distinguish the DUEs that receive the BL or EL#1 from the RUEs. To determine the optimal pair and the optimal reusable channel for relaying the BL, (9)-(13) are first performed. By using (9), the system finds the optimal channel that satisfies the minimization of the transmitting power of i and j0 among ∀i. This allows the channel of i to be a reuse candidate for pair j0 , with constraints (10)-(13).

∀i ∈ C, j1 ∈ R2, k1 ∈ D1, (14) subject to

ric (Tc,i (t), Pi , Pj1 Gj1 k1 ,i ) ≥ ci (t), ∀i ∈ C, (15) c Pi ≤ Pmax , ∀i ∈ C, (16) d Pj1 ≤ Pmax , ∀j1 ∈ R2,

X j0 ∈R1

0 wi,j (t) + 0 k0

X

(17)

1 wi,j (t) ≤ ρ, 1 k1

j1 ∈R2

∀k0 ∈ D0, ∀k1 ∈ D1, (18) where ci (t) works as the bitrate threshold of i in (15), Pj1 Gj1 k1 ,i is the interference caused by D2D pair j0 in (15), 9



and ric (Tc,i (t), Pi , Pj1 Gj1 k1 ,i ) is the interference-considered bitrate of i. Since the relaying of the BL has a higher priority than that of EL#1, the second step of the two-step process used to determine the optimal pair and optimal reusable channel for relaying EL#1 (i.e., (14)-(18)) is performed on the remaining relay candidates after the process for the BL has been performed. By using (14), the system determines the optimal channel that satisfies the minimization of the transmitting power of i and j1 among ∀i. Then, the channel of i becomes a reuse candidate for pair j1 with constraints (15)-(18). (15) prevents the bitrate degradation of the cellular link due to the D2D link. (16) and (17) are constraints limiting the transmitting power of the CUEs and RUEs. Constraint (18) ensures that the number of D2D channels used for relaying the BL and EL#1 does not exceed the number of available cellular channels. During the optimization process with constraints (11) and (12), as well as (16) and (17), the transmitting power of the CUEs and RUEs can be controlled to satisfy the bitrate thresholds, i.e., (10) and (15), of all i. This mirrors the power control scheme proposed in [38]. This two-step optimization process is used in the primary function of our proposed algorithm in Section VI. In Section VI, the provided methods are introduced for channel allocation to D2D pairs with maximized QoS/QoE in the target service scenario. VI. PROPOSAL

This paper proposes relay-based hybrid transmissions in which the links for video relaying in a cell are adaptively selected. In the hybrid transmission method, a UE receives content mainly from the BS; additionally, if there are UEs that do not receive even the lowest layer (i.e., the BL), the other UEs relay the minimum data for as many UEs to enhance the continuity of the video service. The BS controls all links in the cell to increase the QoS and the QoE in terms of the average PSNR and MTTF for all UEs. In this scenario, the CUEs do not help other UEs as relays unless their reception is completed as much as the receivable video layers of the CUEs determined in (6). However, the cellular channels can be reused by D2D pairs, unless service QoS degradation occurs. Due to the co-channel interference issue discussed in [6], which deals with video relaying using D2D communication, the channel of the CUE is shared if the interference-applied SINR is larger than the SINR threshold in the transmission/reception during their service. Similarly, we use the bitrate threshold in the constraints, as shown in (10) and (15), to guarantee successful transmission of the scheduled video traffic for the CUEs. A. SENDING QUEUE ORDER FOR PRIORITY-BASED TRANSMISSION

A study on resource allocation methods for real-time D2D communication [39] did not consider the additional delay caused by the complexity of the proposed system. The au10

thors defined a one-way, end-to-end delay and used it to account for transmission delay. In contrast, in a study on realtime video streaming in cognitive radio networks (CRNs) [40], the time budget of the UEs was calculated using the buffer-fullness and playback rate of the video content. If the time budget is higher than the time required for the transmission of the video chunk, resource allocation is performed. However, the difference between the real-time video and the live video is that the generated video frame in the BS must be transmitted and played back immediately. In the live video service, even if the buffered video is sufficient when the time for replaying the frame expires, it is recognized as a failure and the frame is discarded as illustrated in Fig. 5. However, the live video streaming service also allows for an initial delay. Therefore, we consider the video frame to be successfully communicated if transmission is completed within a GOP duration. Generally, a video stream consists of GOPs transmitted sequentially, as shown in Fig. 5(a). If a UE receives a video stream in the usual sequential manner, as shown in the figure, the point at which the UE can relay the video layer is delayed to the point when the layer has been completely received. Alternatively, the simulation results of [6] argue that more relay candidates provide a higher QoS over D2D cellular networks. Therefore, in this study, GOP data are shuffled according to priority. The highest-priority BL frames are placed first, as shown in Fig. 5(b), to obtain a better chance of being relayed. In the figure, FN denotes the frame number. The proposed order of the transmission increases the probability of replaying the BL seamlessly in the DUEs, while the probability of replaying the lower-priority ELs [41] reduces as the trade-off increases. The proposed method enhances the continuity, which is represented by the MTTF; in most UEs, at least the BL data are delivered on time. The delay burden for the BL is minimized according to Td,jk (t) in (7) and (8). When RUE j relays data to DUE k via the D2D link, Td,jk (t) is calculated using (20). We assume that the link setup delay TL is 200 ms, and we use the discontinuous reception cycle value, as discussed in [42]. Tc,i (t) = Tp − {T¯c,i (t) + TsigC }.

(19)

Td,jk (t) = Tp −{T¯c,i (t)+TsigC +TsigD +TL +Tpro }, (20) where T¯c,i (t)+TsigC is the time it takes the CUE to complete its cellular receiving and TsigD +TL +Tpro is the time it takes the RUE to start playing its role as a relay. In the centralized D2D control system, the primary signaling components that should be periodically transmitted for video streaming in the system are listed as follows. 1) The CUEs send a message, such as CSI and their current Tx power, to the BS. 2) The BS sends a message to the RUEs and DUEs containing the link setup information and information related to the content to be forwarded. 3) The RUEs and DUEs send a report message, such as CSI and their current Tx power, to the BS. VOLUME x, 20xx



TP= GOP duration BL FN1

EL#1 FN1

EL#2 BL EL#1 FN1 FN2 FN2 Receiving BL, EL#1, and EL#2

EL#2 FN2

BL FN3

EL#1 FN3

EL#2 FN3

EL#1 relaying available Receive next GOP BL relaying available (a)

BL BL BL EL#1 FN1 FN1 FN2 FN3 Receiving BL BL relaying available

EL#1 FN2

EL#1 FN3

EL#2 FN1

EL#1 relaying available

EL#2 FN2

EL#2 FN3 Receive next GOP

(b) FIGURE 5. Proposed sending queue order in the BS. (a) Sequential transmission for a GOP. (b) Serial transmission for a GOP.

timeslot

Tp = GOP duration = N timeslots (a)

Timeslots for cellular = Timeslots for D2D

(b)

Timeslots for cellular Timeslots for D2D

(c) FIGURE 6. Unequal mode ratio of flexible hybrid transmission. (a) The unit communication period. (b) Equal phase duration. (c) Unequal phase duration.

4) The BS sends a message to the RUEs and DUEs containing the link setup information and the optimized Tx power of them. 5) All UEs report the communication results to the BS. It can be assumed that more or fewer signaling messages may occur, depending on the purpose of the system. However, studies investigating the resource and rate allocation algorithms in centralized D2D control systems tend to address the signaling process briefly [6]. In this paper, to evaluate on-time video delivery while considering delays, we use TsigC and TsigD to represent the time consumed while signaling is performed for D2D and cellular communication, respectively. However, the assumption that the BS has infinite energy VOLUME x, 20xx

[6] may lead to the assumption that there is no additional delay when optimization is performed at the BS, if the BS has large processing capability; this is the case because the complexity of the proposed algorithms is linear. Similarly, the scenario investigated in in [6], [40] are the resource allocation of the real-time video streaming using D2D communication and CRN, which are controlled by the BS. Both assumed centralized control but did not account for any additional delays caused by the optimized processing in the BS. However, if any process is performed in mobile UEs, which have limited energy capacity, this must be considered explicitly. In this paper, Tpro represents the time consumed during the additional process of the RUE. B. UNEQUAL/INDEPENDENT FLEXIBLE MODE DURATION

Because live video traffic is ordered and delivered in a unit period (Fig. 6(a)) at a constant rate, the ratio between multicasting and relay duration should be 5:5 for HEVC (Fig. 6(b)), which is a universal video codec. In contrast, for SHVC, full-layer reception can be attempted using multicasting, and layers that are not fully cumulated can be transmitted during the relay duration for the UEs that have a much lower available bitrate. Relay-based D2D communications show a high sharing rate [6] when there are several relay users and a few D2D users in the cluster. Therefore, the proposed unequal and independent mode transition control (Fig. 6(c)) in the BS with even energy overhead may occur by singling in each timeslot. To reduce the average energy consumption of the UEs, the BS allows D2D communication only when on-time delivery of the video is possible, taking into account the rate threshold and available time when performing resource allocation decisions in each time slot t. C. SUBSET-BASED RELAY ASSIGNMENT ALGORITHM

Finally, we design a QoS/QoE-aware relay allocation algorithm: Algorithm 3. This algorithm is used in every unit 11



Algorithm 1: D2D link setup algorithm: Initializing 1 2 3 4

U, t, ric (Tc,i (t), Pi ), ∀i

Input ∈U Output C1, C2, C3, D0 for i = 1 : M do li (t) = argmax [cl (t) ≤ ric (Tc,i (t), Pi )] l∈[0,1,2]

5 6 7

switch li (t) do case 0 do C1 ← C1 + {i} after T c,i (t) = rc (Tc,ic0(t),pi ) i

8 9

case 1 do C2 ← C2 + {i} after T c,i (t) = rc (Tc,ic1(t),pi ) i

10 11

case 2 do C3 ← C3 + {i} after T c,i (t) = rc (Tc,ic2(t),pi )

Algorithm 2: Update state of elements in relaying subset 1 Input RR1 , RR2 , RD0 , RD1 , C1, C2, D0, D1 2 Output C1, C2, D0, D1 3 for j = 1 : MRR1 do 4 j=k  RR1 ← RR1 − {j}    RD0 ← RD0 − {k} 5 after T d,jk (t) D0 ← D0 − {k}     C1 ← C1 + {k} 6 j + +; 7 8

9

i

12 13 14

otherwise do D0 ← D0 + {i}

10

for j = 1 : MRR2 do j=k  RR2 ← RR2 − {j}    RD1 ← RD1 − {k} after T d,jk (t)  D1 ← D1 − {k}    C2 ← C2 + {k} j + +;

i + +; Algorithm 2 Algorithm 1 This algorithm uses the bitrate information according to the UEs Rx power Pi and the time Tc,i that can be allocated to the cellular mode. At this time, the maximum value of Tc,i is equal to the total length of the unit communication period (line 1). Algorithm 1 determines the layer of the highest layered video stream li (t) that the UE can completely receive through the cellular mode in unit communication period using the UEspecific rate information ric (line 4). Each UE is added to the subset C1, C2, and C3 according to the hierarchy that has succeeded in receiving after only receiving through the cellular mode up to the expected value of the time required to obtain the highest possible receiving layer. The UEs that has not received any layer is immediately added to subset D0 (line 6-13). If there are M UEs in the set U , then the computational complexity for Algorithm 1 is O(M ).

communication period to perform the optimized D2D relay assignment during the duration of the unit period. The algorithm performs the allocation process by calling subprocesses separated into Algorithm 1 and Algorithm 2. In the entire algorithm, (6) is used for the channel allocation and categorization of the CUEs and (9) and (14) are used for the channel allocation and mode control for the D2D pairs. Algorithm 1 is performed for cellular-D2D mode initialization, and is called at the beginning of every unit communication period. It is performed for all elements M in set U , which is a set of all UEs in a cell participating in the cellular mode at the current time t. Algorithm 2 is called at every timeslot to determine whether the UEs communicating in D2D mode have completed the communication. It also reflects the state change in subsets RR1 and RR2 . In the algorithm, MRR1 and MRR2 denote the number of elements in the subsets. Algorithm 3 updates the state of the UEs by calling Algorithm 1 and Algorithm 2. It also determines the optimized 12

Algorithm 2, which is called when the beginning of every timeslot, manages the mode transition of UEs belonging to RR1 , RR2 , RD0 , and RD1 which are the subsets of UEs entering D2D mode (line 1). In addition, Algorithm 2 adds the elements that have completed D2D communication to the standby subsets C1, C2, D0, and D1 of the system (line 2). The elements j and k of RR1 and RD0 constituting a D2D link pair are moved to the idle mode subsets after T d,jk (t) duration in which the BL sharing D2D communication is completed (line 4-5). The elements j and k of RR2 and RD1 constituting a D2D link pair are moved to the idle mode subsets after T d,jk (t) duration in which the EL#1 sharing D2D communication is completed (line 8-9). If there are MRR1 and MRR2 UEs in the subsets RR1 and RR2 , then the computational complexity for Algorithm 2 is O(max(MRR1 , MRR2 )).

D2D link set using the two-step optimization process and commands the start of D2D communication. The optimization for BL relaying is performed for the DUEs in D0 and R1. In the case of EL#1 relaying, it is performed for the DUEs in D1 and R2. In the algorithm, MD0 , MR1 , MD1 , and MR2 denote the number of elements in the corresponding subsets. In the next section, we perform a system-level simulation to evaluate the performance of the QoS/QoE-aware relay allocation algorithm and analyze the results. VII. SYSTEM-LEVEL SIMULATION

In the first simulation, we conduct a performance evaluation of the proposed algorithm by comparing it with the algorithm given in [13], based on the number of CUEs and DUEs and the radius of the D2D cluster. In the simulation, the CUEs and DUEs are uniformly distributed in a cluster, and streaming video content has three video layers: BL, EL#1, and EL#2. Table 3 shows the parameters used in the simulation. In the table, the cell shape and the default cluster radius are given in [43]–[45]. Also, we set the unit period to 1 s, i.e., VOLUME x, 20xx



the generally used chunk size of the video. In this simulation, we assume that failure to play a video frame fails because a

Algorithm 3: QoS/QoE-driven D2D link setup algorithm 1 Input C, C1, C2, D0, D1 0 1 2 Output W , W , RR1 , RR2 , RD0 , RD1 3 Initialize t = 0 4 Set C1, C2, D0 {the output of Algorithm 1} 5 while t < Tp do 6 update C1, C2, D0, D1 {the output of Algorithm 2}; 7 R1 ← ∀c1 ; 8 R1 ← ∀c2 ; 9 if R1 6= φ then 10 for k = 1 : MD0 do 11 for j = 1 : MR1 do 12 compute Td,jk (t), ∀i ∈ C; 13 find relay pair candidates reusing i’s 0 channel minPi ,Pj {w0 (t) | wi,jk (t) = 1}, ∀i, j ∈ R1, k ∈ D0; 14 set D2D pair(j, k) to W 0 ; 15 R1 ← R1 − {j} ; 16 D0 ← D0 − {k}; 17 RR1 ← RR1 + {j}; 18 RD0 ← RD0 + {k}; 19 j + +;

22 23 24 25 26 27

28 29 30 31 32 33

D1 ← ∀c1 ; R2 ← ∀c2 ; if φ ∈ D0 && R2 6= φ then for k = 1 : MD1 do for j = 1 : MR2 do compute Td,jk (t), ∀i ∈ C; find relay pair candidates reusing i’s 1 channel minPi ,Pj {w1 (t) | wi,jk (t) = 1}, ∀i, j ∈ R2, k ∈ D1; set D2D pair (j, k) to W 1 ; R2 ← R2 − {j} ; D1 ← D1 − {k}; RR2 ← RR2 + {j}; RD1 ← RD1 + {k}; j + +; k + +;

34 35 36

37 38

39

The D2D link setup algorithm, which is performed in units of the time slot duration χ within a unit communication period equal to the generation period of the live video stream, obtains initialized subsets C1, C2, C3, and D0 using Algorithm 1 at the beginning of the period line (line 3-4). For real-time streamed video sharing, C1, C2, C3, and D0 participate in Algorithm 3. The reason of C3 is excluded here is because most of the unit communication period is consumed in the cellular mode for the full receiving of the highest video layer. After subsets initialization using Algorithm 1, each D2D communication completion time of UEs participating in D2D communication is checked for each χ, and Algorithm 2 is called to perform the transition between subsets according to the state of each UE (line 6). The subsets C1 and C2 that have received the BL or EL#1 through cellular or D2D communication are added to R1, which is a subset of the BL relay candidate (line 7-8). The Algorithm 3 determines the D2D pair with the lowest energy consumption when it can deliver the BL through D2D communication among all the combinations between the elements of R1 and the elements of D0 that have not even received the BL (line 10-13). The relay and destination UEs of the determined pair are moved to the subsets which indicate activate D2D communication (line 14-18). After completing the link setup process for the propagation of the BL, elements of C1 having the BL become destination UEs for the propagation of EL#1 (line 21). Elements of C2 with EL#2 become the relay candidate UE (line 22). If there are no UEs that have received no BL (line 23), and there are UEs capable of transmitting EL#1, the D2D link setup process starting. The Algorithm 3 determines the D2D pair with the lowest energy consumption when it can transfer the EL#1 through D2D communication among all the combinations between the elements of R2 and the elements of D1 that have not received the EL#1 (line 24-27). The relay and destination UEs of the determined pair are moved to the subsets which indicate activate D2D communication (line 28-32). Initiate communication between the D2D pair for the propagation of the BL at the current time t (line 35-36). Initiate communication between the D2D pair for the propagation of the EL#1 at the current time t (line 37-38). Then, the algorithm moves to the next timeslot in the unit communication period. The computational complexity of Algorithm 3 can be expressed as O(max(MD0 MR1 , MD1 MR2 )).

k + +;

20 21

Algorithm 3

if W 0 6= φ then D2D communication begin for all elements of Wj0 if W 1 6= φ then D2D communication begin for all elements of Wj1 t←t+χ

VOLUME x, 20xx

GOP playback failure requires separate criteria to determine the playback failure event and the duration of the failure. Depending on the type of video streaming service, there may be a designated unit of failure, but this paper deals with live video streaming scenarios and therefore performs frameby-frame measurements to provide detailed analysis. This method uses the video frame rate in the table to obtain the duration of the failure. We compared the average PSNR performance with the average energy consumption of the proposed algorithm (PM) and the conventional algorithm in [13] (CM1). TABLE 3. Parameters used for the system-level simulation.

Parameter

Value

Cell layout Cell radius Cluster radius Bandwidth Operating frequency Video frame rate Average bitrate of SHVC content (BL, EL#1, EL#2) Average bitrate of HEVC-encoded content Maximum transmitting power of UE Noise power density Slot duration

Single cell 500 m 50 m 5 MHz 1920-1924.32 MHz 30 frames per second 3, 7, 15 [Mbps] (l0, l1, l2) 13.5 Mbps 24 dbm -114 dbm/Hz 100 ms

13



4 0

3 0 0

C M 1 P M

3 8

2 5 0

3 6

2 0 0

3 4

1 5 0

M T T F

P S N R [d B ]

C M 1 P M

3 2

1 0 0

3 0

5 0

2 8 0 1

2

3

4

5

6

7

8

9

1 0

1 1

1 2

N u m b e r o f C U E s / N u m b e r o f D U E s

1

2

3

4

5

6

7

8

9

1 0

1 1

1 2


FIGURE 7. Comparisons of the PSNR; the cluster radius is 50m.

FIGURE 9. Comparisons of the MTTF; where the cluster radius is 50m.

0 .2 4

PM CM1

E n e r g y c o n s u m p tio n [J o u le s /s ]

C M 1 P M 0 .2 0

0 .1 6

0 .1 2

0 .0 8 1

2

3

4

5

6

7

8

9

1 0

1 1

1 2


FIGURE 8. Comparisons of the energy consumption; the cluster radius is 50m.

Energy consumption [Joules/s] FIGURE 10. Comparisons of the PSNR with varying cluster radii.

Fig. 7 and Fig. 8 show the results of the average PSNR and the average energy consumption performance of UEs with varying UE ratios (CUEs per DUEs). From Fig. 7, we can conclude that the PM has better PSNR performance when the UE ratios are small. The enhanced PSNR performance of PM is due to the small number of relay candidates successfully delivering the BL to a large number of DUEs. However, the gap in the PSNR performance between the PM and CM1 is small. Furthermore, when Fig. 7 and Fig. 8 are compared, although PM have successfully propagated the BL through multiple relays, the energy consumption of the PM is larger than that of the CM1; thus, we cannot confirm the superiority of the overall performance of the PM. The MTTF results in 9 show a similar trend as the results in Fig. 7. However, in the case of the MTTF comparison, the performance difference between the CM1 and PM becomes more evident. In this comparison, the PM enhanced the average MTTF by 70% or more, with 20% more energy consumption in the UEs. Additionally, Fig. 10 and Fig. 11 show how the cluster radius affects the PSNR and MTTF in the system scenario. Fig. 10 indicates that the PM shows an average improvement 14

of 1 dB in the PSNR for different cluster radii, as compared to the conventional method. With a small cluster radius, the PM shows up to 30% energy consumption overhead; however, as the cluster radius increases, the energy overhead decreases rapidly and has an overhead of less than 5%. The larger the cluster radius, the lower the energy overhead of the PM. This is because the UEs that do not participate in the relay for live video relaying are searched in a wider range and are excluded as potential relay candidates. Fig. 11 shows that the PM significantly improved the MTTF performance for various cluster radii, as compared to the CM1. The MTTF became 2-4 times longer. As the cluster radius increased, the MTTF became longer, while the energy consumption overhead was almost the same as that in the CM1. This is the effect of the proposed algorithm in which live video streams are not pre-cached but are instead shared within the unit communication period duration. Fig. 12 illustrates a summary of Fig. 10 and Fig. 11. When the radius of the cluster increases, the PM reduces distortion of the QoE in terms of the MTTF. The CM1 shows a QoE distortion that is approximately twice as high. However, when VOLUME x, 20xx



PM CM1

Energy consumption [Joules/s] FIGURE 11. Comparisons of the MTTF with varying cluster radii.

8 0

C M 1 P M ( C M 1 P M (

Q o S /Q o E d is to r tio n [% ]

6 0

(P S P S N (M T M T T

N R ) R ) T F ) F )

4 0

2 0

0 2 0

4 0

6 0

8 0

1 0 0

C lu s te r r a d iu s [m ]

FIGURE 12. QoS/QoE distortion with varying cluster radii.

using the PSNR as a QoS parameter, it is not easy to compare the PSNR performance between the proposed method and the conventional method. This phenomenon results from the fact that intermittent video playback failures are not significantly revealed in the average PSNR, as shown in Fig. 12. Thus, in the second simulation, we focus on clarifying the QoS/QoE performance of the PM. We also analyze the characteristics and utility of both parameters. In the simulation, the CUEs and DUEs are randomly distributed in a cell to make the difference of the performance of the systems clear by creating a harsher environment, as compared to the first simulation. In this simulation, we compare the performance of the PM with that of the algorithm given in [6] (i.e., the CM2). In [6], a cooperative transmission scheme was proposed to enhance the performance of realtime video streaming in 5G networks using D2D communication. The proposed streaming system in [6] performs cooperative transmission by the BS and CUEs, which cache the video segments for transmission of the requested video by the DUEs. The following components constitute the system proposed in [6]. VOLUME x, 20xx

1) D2D communication component. 2) Source rate-distortion (RD) and video packetization components. 3) End-to-end reconstructed distortion component. 4) Energy consumption component. 5) Co-channel interference control component. Components 1, 4, and 5 play roles that are similar to processes in the algorithm proposed in this paper. However, there is a big difference in that the video encoding process is performed by components 2 and 3, where D2D helpers consider the allocated channel state and the RD between the RUEs and DUEs. According to components 2 and 3, the additional process performed in the RUE is video encoding. We assume that the video encoding to support more exquisite the spatial and quality scalability on relaying node leads to better adaptability for the D2D communication. That is, the RUE generates the optimal video chunk by considering the RD and changing the resolution and quality. To compare the system performance between the PM and CM2, we assume that there is no additional delay due to the optimization and decision processes performed on the BS. The assumption that the BS has infinite energy [6] may lead to the assumption that there is no additional delay caused by optimization performed at the BS if the BS has a large processing capability. Similarly, the scenario in [40] considers resource allocation of real-time video streaming in CRN, which is controlled by the BS via centralized control. Nevertheless, they did not account for any additional delays caused by optimization processing in the BS. However, if the additional process is performed in the UEs that have limited energy capacity, this must be considered explicitly. In the case of the CM2, the additional components perform video encoding in the RUEs. To emulate energy and delay overheads caused by additional processes in the RUEs, we set the emulation parameters as shown in Table 4. In the emulation, the video encoder we used is provided by [46]. The results of the emulated overheads are included in the simulation results. We simulated the average PSNR of the CM2 and PM with different CUEs to DUEs ratios in the system. As shown in Fig. 13, the PSNR tends to decrease as the ratio increases because the DUEs failed to receive the full video frames on time. In some regions, the PSNR of the CM2, which encodes the video to utilize the capacity of the D2D links fully, is slightly larger than that of the PM. However, in the middle region of the figure, the PSNR performance of the CM2 is TABLE 4. Parameters used for the emulation.

Parameter Processor (CPU) Processor (GPU) Memory Open source library for video encoding

Description Exynos9 8895 2.3GHz + ARM Cortex-A53 1.7GHz ARM Mali-G71 546MHz 6GB LPDDR4X SDRAM FFmpeg with X265 encoder

15



3 6

5

3 2

P S N R

[d B ]

3 4

3 0

2 8 0 .0

0 .2

0 .4

0 .6

0 .8

1 .0


E n e r g y c o n s u m p tio n [J o u le s /s ]

C M 2 P M

C M 2 P M 4

3

2

1 0 .2 0 .1 0 .1 0 .0 0 .0

0 5 0 5 0 0 .1

FIGURE 13. Average PSNR performance of the DUEs; the cluster radius is 50m.

0 .2

0 .3

0 .4

0 .5

0 .6

0 .7

0 .8

0 .9

N u m b e r o f C U E s / N u m b e r o f D U E s FIGURE 15. Comparisons of the average energy consumption; the cluster radius is 50m.

4 0 3 9

2 3 0

3 8 2 2 0 2 1 0

3 6 2 0 0

3 5 1 9 0

3 4

C M 2 P M

3 3 3 2 0 .0

0 .2

0 .4

0 .6

0 .8

1 .0


FIGURE 14. Average PSNR performance of the CUEs; the cluster radius is 50m.

M T T F

P S N R

[d B ]

3 7

1 8 0 1 7 0 1 6 0

C M 2 P M

1 5 0 1 4 0 0 .2

0 .4

0 .6

0 .8


FIGURE 16. Comparisons of the MTTF; the cluster radius is 50m.

much lower than that of the PM in the region where the PSNR decreases sharply. The rapidly decreasing PSNR is caused by the additional delay due to the fact that video encoding consumes a portion of the D2D communication time, which should be dedicated to transmitting lower-quality video. The average PSNR performance of the CUEs is shown in Fig 14. In the case of the PM, the target bitrate constraints are used to prevent a decrease in the QoS of the CUEs caused by interferences from D2D pairs. Alternatively, in the case of the CM2, target SINR constants are used. As illustrated in Fig. 14, both methods guarantee high PSNR performance of the streaming service regardless of the change in the ratio. To compare the average energy consumption of the UEs in the PM and CM2, performance evaluation was performed for the CUEs (RUEs included). This was done because there is no difference in the operations performed by the DUEs in the PM and CM2. With the emulation parameters in Table 4, we were able to analyze the efficiency of video encoding on the UE in live streaming services. In the experiment, the average energy consumption of the UEs in the CM2 was between four times and 44 times greater than that of the PM, as shown in Fig. 15. There are exceptions where the CM2 finds optimal RD conditions and performs video encoding with high in16

tensity, resulting in an increase in energy consumption, as in some ratios (i.e., 0.5 and 0.6) where the PSNR slightly increased in Fig 13. However, this gap generally decreases as the ratio increases because the RUEs in the CM2 sharply reduces the target video quality for encoding to adapt to the D2D link. Therefore, the energy required for encoding is dramatically reduced as the video quality decreases. In the case of the PM, the energy consumption decreases as the amount of transmitted video traffic decreases. However, due to the rapidly reduced energy for encoding, the difference between the CM2 and PM was reduced. As noted in the PSNR evaluation in Fig. 13, a slight PSNR improvement due to the video encoding was present for some ratios, but the degree of improvement was very small when considered with the energy consumption. In addition, as the number of RUEs able to help one DUE decreases, the possibility of having a helper that has a large amount of slack time decreases. Therefore, the additional delay caused by the encoding process leads to rapid PSNR degradation. Fig. 16 shows the performance evaluation of the UEs for the CM2 and PM in terms of the MTTF. In this figure, we can VOLUME x, 20xx



see that the trend of the MTTF differs from the PSNR trend in Fig. 13. This difference is due to the different characteristics of the PSNR and MTTF. Because the PSNR reflects the average quality of the video frames, which are components of the video, the PSNR value does not decrease significantly, even if the frame is intermittently lost. However, the MTTF is a useful parameter for determining the stability of the service. As we can see in the figures, there are some cases where the PSNR value of CM2 is higher than that of the PM, but the MTTF value is lower than that of the PM. VIII. CONCLUSION

In this paper, we investigated the issue of efficient resource allocation for live video streaming in cellular-D2D cooperative communications in order to provide acceptable QoE. In this context, we proposed three techniques: (1) prioritybased video transmission, (2) flexible communication mode switching of UEs, and (3) subset-based relay assignment. Through extensive simulation works, we demonstrated that the proposed scenario dramatically improves the continuity of the service by distributing the essential video layer for video playback. In particular, although the UEs in the proposed scenario operate as relays, the energy consumption is considerably more efficient as the subset-based relay assignment algorithm manages the UEs for D2D communication efficiently. The system-level simulation results show that our framework has superior performance in terms of the MTTF, which can recognize intermittent video playback failures. Notably, we can identify the MTTF as a QoE parameter for live video streaming services because the stability of live video cannot be perceived intuitively using the PSNR. However, the MTTF cannot represent the quality of the video itself as a general metric in other applications, where the benefits of the MTTF do not stand out. Although the MTTF is not an absolute QoE parameter for video services, it can be useful for live video streaming in unstable environments due to various limitations. Therefore, it is expected that the proposed methods and the QoE parameter will be highly utilized as the demand for interactive live video content increases in D2D-underlaid 5G cellular networks. REFERENCES [1] CISCO, “CISCO visual networking index: Global mobile data traffic forecast, 2016–2021,” CISCO white paper, 2017. [2] C. Bila, F. Sivrikaya, M. A. Khan, and S. Albayrak, “Vehicles of the future: A survey of research on safety issues,” IEEE Transactions on Intelligent Transportation Systems, vol. 18, no. 5, pp. 1046–1065, 2017. [3] G. I. Tsiropoulos, A. Yadav, M. Zeng, and O. A. Dobre, “Cooperation in 5G HetNets: Advanced spectrum access and D2D assisted communications,” IEEE Wireless Communications, vol. 24, no. 5, pp. 110–117, 2017. [4] P. Gandotra and R. K. Jha, “Device-to-device communication in cellular networks: A survey,” Journal of Network and Computer Applications, vol. 71, pp. 99–117, 2016. [5] P. Gandotra, R. K. Jha, and S. Jain, “A survey on device-to-device (D2D) communication: Architecture and security issues,” Journal of Network and Computer Applications, vol. 78, pp. 9–29, 2017. [6] N. S. Vo, T. Q. Duong, H. D. Tuan, and A. Kortun, “Optimal video streaming in dense 5G networks with D2D communications,” IEEE Access, vol. 6, pp. 209–223, 2018. VOLUME x, 20xx

[7] D. Liu, L. Wang, Y. Chen, M. Elkashlan, K.-K. Wong, R. Schober, and L. Hanzo, “User association in 5g networks: A survey and an outlook,” IEEE Communications Surveys & Tutorials, vol. 18, no. 2, pp. 1018–1044, 2016. [8] Y. Zeng, H. Hu, T. Xu, and B. Jia, “User pairing stability in D2D-relay networks,” IEEE Communications Letters, vol. 21, no. 10, pp. 2278–2281, 2017. [9] Z. Zhou, C. Gao, C. Xu, Y. Zhang, S. Mumtaz, and J. Rodriguez, “Social big-data-based content dissemination in internet of vehicles,” IEEE Transactions on Industrial Informatics, vol. 14, no. 2, pp. 768–777, 2018. [10] C. Campolo, A. Molinaro, G. Araniti, and A. O. Berthet, “Better platooning control toward autonomous driving: An LTE device-to-device communications strategy that meets ultralow latency requirements,” IEEE Vehicular Technology Magazine, vol. 12, no. 1, pp. 30–38, 2017. [11] J. Li, R. Bhattacharyya, S. Paul, S. Shakkottai, and V. Subramanian, “Incentivizing sharing in realtime D2D streaming networks: A mean field game perspective,” IEEE/ACM Transactions on Networking, vol. 25, no. 1, pp. 3–17, 2017. [12] H. Dai, Y. Huang, R. Zhao, J. Wang, and L. Yang, “Resource optimization for device-to-device and small cell uplink communications underlaying cellular networks,” IEEE Transactions on Vehicular Technology, vol. 67, no. 2, pp. 1187–1201, 2018. [13] Y. Li, T. Jiang, M. Sheng, and Y. Zhu, “QoS-aware admission control and resource allocation in underlay device-to-device spectrum-sharing networks,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 11, pp. 2874–2886, 2016. [14] W. Xing, F. Liu, C. Wang, M. Xiao, and P. Wang, “Multi-source networkcoded D2D cooperative content distribution systems,” Journal of Communications and Networks, vol. 20, no. 1, pp. 69–84, 2018. [15] J. Gui and J. Deng, “Multi-hop relay-aided underlay D2D communications for improving cellular coverage quality,” IEEE Access, vol. 6, pp. 14318– 14338, 2018. [16] B. Ying and A. Nayak, “A power-efficient and social-aware relay selection method for multi-hop D2D communications,” IEEE Communications Letters, 2018. [17] W. Wang, “Collaborative multimedia source-protocol coordination: A cross-layer QoE study in modern wireless networks,” IEEE Systems Journal, vol. 11, no. 4, pp. 2403–2409, 2017. [18] L. Lei, Y. Kuang, N. Cheng, X. S. Shen, Z. Zhong, and C. Lin, “Delayoptimal dynamic mode selection and resource allocation in device-todevice communications—Part I: Optimal policy,” IEEE Transactions on Vehicular Technology, vol. 65, no. 5, pp. 3474–3490, 2016. [19] L. Lei, Y. Kuang, N. Cheng, X. Shen, Z. Zhong, and C. Lin, “Delayoptimal dynamic mode selection and resource allocation in device-todevice communications—Part II: Practical algorithm,” IEEE Transactions on Vehicular Technology, vol. 65, no. 5, pp. 3491–3505, 2016. [20] Q. Wang, T. M. Steinman, and W. Wang, “Quality driven modulation rate optimization for energy efficient wireless video relays,” Computer Communications, vol. 115, pp. 2–9, 2018. [21] C. Li, Y. Yan, and B. Zhang, “Network coding aided collaborative realtime scalable video transmission in D2D communications,” IEEE Transactions on Vehicular Technology, 2018. [22] M. J. Piran, S. R. Islam, and D. Y. Suh, “Cash: Content-and networkcontext-aware streaming over 5g hetnets,” IEEE Access, vol. 6, pp. 46167– 46178, 2018. [23] A. A. Ramanand, I. Ahmad, and V. Swaminathan, “A survey of rate control in HEVC and SHVC video encoding,” in Multimedia & Expo Workshops (ICMEW), 2017 IEEE International Conference on, pp. 145–150, IEEE, 2017. [24] G. J. Sullivan, J.-R. Ohm, W.-J. Han, T. Wiegand, et al., “Overview of the high efficiency video coding(HEVC) standard,” IEEE Transactions on circuits and systems for video technology, vol. 22, no. 12, pp. 1649–1668, 2012. [25] J. M. Boyce, Y. Ye, J. Chen, and A. K. Ramasubramonian, “Overview of SHVC: Scalable extensions of the high efficiency video coding standard,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 26, no. 1, pp. 20–34, 2016. [26] M. Zulhasnine, C. Huang, and A. Srinivasan, “Efficient resource allocation for device-to-device communication underlaying LTE network,” in Wireless and Mobile Computing, Networking and Communications (WiMob), 2010 IEEE 6th International Conference on, pp. 368–375, IEEE, 2010. [27] P. Kyösti, J. Meinilä, L. Hentilä, X. Zhao, T. Jämsä, C. Schneider, M. Narandzic, M. Milojevic, A. Hong, J. Ylitalo, et al., “WINNER II 17



[28]

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44]

[45]

[46]

18

channel models part I channel models,” WINNER II Public Deliverables, 2007. P. Kyösti, J. Meinilä, L. Hentilä, X. Zhao, T. Jämsä, C. Schneider, M. Narandzic, M. Milojevic, A. Hong, J. Ylitalo, et al., “WINNER II channel models part II radio channel measurement and analysis results,” WINNER II Public Deliverables, 2007. L. Lei, Y. Zhang, X. Shen, C. Lin, and Z. Zhong, “Performance analysis of device-to-device communications with dynamic interference using stochastic petri nets,” IEEE transactions on wireless communications, vol. 12, no. 12, pp. 6121–6141, 2013. B.-J. Chang and S.-H. Liou, “Adaptive cooperative communication for maximizing reliability and reward in ultra-dense small cells LTE-A toward 5G cellular networking,” Computer Networks, vol. 115, pp. 16–28, 2017. A. Asadi, Q. Wang, and V. Mancuso, “A survey on device-to-device communication in cellular networks,” IEEE Communications Surveys & Tutorials, vol. 16, no. 4, pp. 1801–1819, 2014. M. Naslcheraghi, L. Marandi, and S. A. Ghorashi, “A novel device-todevice discovery scheme for underlay cellular networks,” in Electrical Engineering (ICEE), 2017 Iranian Conference on, pp. 2106–2110, IEEE, 2017. P. Nguyen, P. Wijesinghe, R. Palipana, K. Lin, and D. Vasic, “Networkassisted device discovery for LTE-based D2D communication systems,” in Communications (ICC), 2014 IEEE International Conference on, pp. 3160–3165, IEEE, 2014. Z. Liu, M. Dong, H. Zhou, X. Wang, Y. Ji, and Y. Tanaka, “Device-todevice assisted video frame recovery for picocell edge users in heterogeneous networks,” in Communications (ICC), 2016 IEEE International Conference on, pp. 1–6, IEEE, 2016. Y. Cho, S. Park, K. Kim, and D. Suh, “Adaptive live streaming system performance of MMT and DASH over a deployed LTE network,” Electronics Letters, vol. 52, no. 13, pp. 1128–1130, 2016. X. Zhang, J. Chen, F. Tian, and Z. Yuan, “The role of video coding in D2D communications,” in Globecom Workshops (GC Wkshps), 2017 IEEE, pp. 1–6, IEEE, 2017. J. Wu, C. Yuen, N.-M. Cheung, J. Chen, and C. W. Chen, “Streaming mobile cloud gaming video over TCP with adaptive source–FEC coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 27, no. 1, pp. 32–48, 2017. P. Sun, K. G. Shin, H. Zhang, and L. He, “Transmit power control for D2Dunderlaid cellular networks based on statistical features,” IEEE Trans. Vehicular Technology, vol. 66, no. 5, pp. 4110–4119, 2017. J. Xu and C. Guo, “Resource allocation for real-time D2D communications underlaying cellular networks,” IEEE Transactions on Mobile Computing, 2018. Early Access doi:10.1109/TMC.2018.2849743. A. E. Omer, M. S. Hassan, and M. El-Tarhuni, “Window-based adaptive technique for real-time streaming of scalable video over cognitive radio networks,” IET Communications, vol. 11, no. 17, pp. 2643–2649, 2017. M. J. Piran, N. H. Tran, D. Y. Suh, J. B. Song, C. S. Hong, and Z. Han, “QoE-driven channel allocation and handoff management for seamless multimedia in cognitive 5G cellular networks,” IEEE Transactions on Vehicular Technology, vol. 66, no. 7, pp. 6569–6585, 2017. Y. Xu, R. Yin, T. Han, and G. Yu, “Dynamic resource allocation for deviceto-device communication underlaying cellular networks,” International Journal of Communication Systems, vol. 27, no. 10, pp. 2408–2425, 2014. L. Wang, H. Tang, H. Wu, and G. L. Stüber, “Resource allocation for D2D communications underlay in Rayleigh fading channels,” IEEE Transactions on Vehicular Technology, vol. 66, no. 2, pp. 1159–1170, 2017. A. Grassi, G. Piro, G. Bacci, and G. Boggia, “Uplink resource management in 5G: when a distributed and energy-efficient solution meets power and QoS constraints,” IEEE Transactions on Vehicular Technology, vol. 66, no. 6, pp. 5176–5189, 2017. E. Yaacoub, “Cooperative energy efficient D2D clustering in LTE-A with enhanced QoS,” Telecommunication Systems, vol. 67, no. 3, pp. 401–414, 2018. FFmpeg tool. Accessed: Oct. 11, 2018. Available online: https://trac.ffmpeg.org/wiki/Encode/H.265.

JIHYEOK YUN received his B.S. in Electronic Engineering from Kyung Hee University in 2011, and his M.S. in Electronics and Radio Engineering from Kyung Hee University in 2013. Since September 2013, he has been a Ph.D. student for Electronics and Radio Engineering in the College of Electronics and Information at Kyung Hee University. He has been working as a Korean delegate for the ISO/IEC Moving Picture Experts Group (MPEG) since 2011.

MD. JALIL PIRAN (SM’10-CM’16) is an Assistant Professor with the Department of Computer Science and Engineering, Sejong University, Seoul South Korea. Jalil Piran completed his PhD in Electronics and Radio Engineering from Kyung Hee University, South Korea, in 2016. Subsequently, he continued his work as a Postdoctoral Research fellow in the field of "Resource Management" and "QoE" in "5G Cellular Networks" and “Internet of Things (IoT)” in the Networking Lab, Kyung Hee University. Dr. Jalil Piran published substantial number of technical papers in well-known international journals and conferences in research fields of “Resource allocation and management in; 5G mobile and wireless communication, HetNet, IoT, Multimedia Communication, Streaming, adaptation and QoE, and Cognitive Radio Networks. He received "International Association of Advanced Materials (IAAM) Scientist Medal of the year 2017 for notable and outstanding research in the field of New Age Technology & Innovation,” in Stockholm, Sweden. Moreover, he has been recognized as the "Outstanding Emerging Researcher" by the Iranian Ministry of Science, Technology, and Research in 2017. In addition, his PhD dissertation has been select as the “Dissertation of the Year 2016” by the Iranian Academic Center for Education, Culture, and Research in the field of Electrical and Communications Engineering. In the worldwide communities, Dr. Jalil Piran is an active member of Institute of Electrical and Electronics Engineering (IEEE) since 2010, an active delegate from South Korea in MPEG since 2013, and an active member of IAAM since 2017.

DOUG YOUNG SUH (S’89-M’90) received the BSc degree in Department of Nuclear Engineering from Seoul University, South Korea, in 1980, and MSc and PhD degrees from Department of Electrical Engineering in Georgia Institute of Technology, Atlanta, Georgia, USA, in 1986 and 1990, respectively. In September 1990, he joined Korea Academy of Industry and Technology and conducted research on High-definition Television (HDTV) until 1992. Since February 1992, he is a professor in College of Electronics and Information Engineering in Kyung Hee University, South Korea. His research interests include networked video and video game. He has been working as a Korean delegate for ISO/IEC MPEG since 1996.

VOLUME x, 20xx
