Convergence of Heterogeneous Network and IT ... - IEEE Xplore

4 downloads 10150 Views 409KB Size Report
to support Cloud and mobile Cloud computing services. The proposed ... support the Infrastructure as a Service paradigm, the proposed architecture adopts the.
Future Network & MobileSummit 2013 Conference Proceedings Paul Cunningham and Miriam Cunningham (Eds) IIMC International Information Management Corporation, 2013 ISBN: 978-1-905824-37-3

Convergence of Heterogeneous Network and IT infrastructures in Support of Fixed and Mobile Cloud Services Markos P. ANASTASOPOULOS1, Anna TZANAKAKI1, Georgios S. ZERVAS2, Bijan R. ROFOEE2, Reza NEJABATI2, Dimitra SIMEONIDOU2, Giada LANDI3, Nicola CIULLI3, Jordi F. RIERA4, Joan A. GARCÍA-ESPÍN4 1 Athens Information Technology, Markopoulou Avenue, Athens, 19002, Greece Tel: +302106682766, Fax: +302106682703, [manast, atza]@ait.gr 2 Merchant Venturers School of Engineering, University of Bristol, UK 3 Nextworks, via Livornese 1027, Pisa, 56122, Italy 4 Fundació i2CAT, C./ Gran Capità 2-4, 08034, Barcelona, Catalunya, Spain Abstract: This paper proposes a next generation ubiquitous converged infrastructure to support Cloud and mobile Cloud computing services. The proposed infrastructure facilitates interconnection of fixed and mobile end users with IT resources through a heterogeneous network integrating optical metro and wireless access networks. To support the Infrastructure as a Service paradigm, the proposed architecture adopts the concept of virtualization across the technology domains involved. To optimally plan the proposed virtual infrastructures a holistic approach considering jointly the presence of all network technology domains and the IT resources is applied. Our modelling results identify trends and trade-offs relating to end-to-end service delay, resource requirements and energy consumption levels of the infrastructure across the various technology domains. Keywords: Optical Networks, Virtualization, Wireless Networks, IT resources, Cloud Computing, Mobile Cloud Computing.

1. Introduction As the availability of high-speed Internet access is increasing at a rapid pace and new demanding applications are emerging, distributed computing systems are gaining increased popularity. Over the past decade, large-scale computer networks supporting both communication and computation were extensively employed in accordance to the cloud computing paradigm. Cloud computing facilitates access to computing resources on an ondemand basis, enabling customers to access remote computing resources that they do not have to own. This introduces a new business model and facilitates new opportunities for a variety of business sectors. At the same time it increases sustainability and efficiency in the utilization of the available resources reducing the associated capital and operational expenditures as well as the overall energy consumption and CO2 footprint. Recently the concept of mobile cloud computing, where computing power and data storage are moving away from the mobile devices to remote computing resources [1] is also gaining increased attention. It is predicted that cloud computing services are emerging as one of the fastest growing business opportunities for Internet service providers and telecom operators [2]-[3]. In addition, mobile internet users are expected to exceed in number the desktop internet users after year 2013, introducing a huge increase in mobile data, a big part of which will come from Cloud computing applications [2].

Copyright © 2013 The authors

www.FutureNetworkSummit.eu/2013

Page 1 of 9

Existing mobile cloud computing solutions allow mobile devices to access the required resources by accessing a nearby resource-rich cloudlet, rather than relying on a distant “cloud,” [4]. In order to satisfy the low-latency requirements of several content-rich mobile cloud computing services such as high definition video streaming, online gaming and real time language translation [3], one-hop, high-bandwidth wireless access to the cloudlet is required. In the case where a cloudlet is not nearby available, traffic is offloaded to a distant cloud such as Amazon’s Private Cloud, GoGrid [5] or Flexigrid [6]. However, the lack of service differentiation mechanisms for mobile and fixed cloud traffic across the various network segments involved, the varying degrees of latency at each technology domain and the lack of global optimization tools in the infrastructure management and service provisioning make the current solutions inefficient. To effectively enable this emerging business opportunity, there is a need for a converged infrastructure supporting integrated wireless and wired high capacity optical networks interconnecting IT resources that allows seamless orchestrated on-demand service provisioning across the heterogeneous technology domains. Such a converged infrastructure will reduce the capital and operational expenditures, increase efficiency and network performance, migrate risks, support guaranteed QoS and meet the quality of experience (QoE) requirements of Cloud and mobile Cloud services. To address this need, this paper is focusing on a next generation ubiquitous converged network infrastructure. The infrastructure model proposed is based on the Infrastructure as a Service (IaaS) paradigm and aims at providing a technology platform interconnecting geographically distributed computational resources that can support a variety of Cloud and mobile Cloud services. The proposed architecture addresses the diverse bandwidth requirements of future cloud services by integrating advanced optical network technologies offering fine (sub-wavelength) switching granularity with state-of-the-art wireless Long Term Evolution (LTE) access network technology, supporting end user mobility, through wireless backhauling. To enable sharing of the physical resources and support the IaaS paradigm as well as the diverse and deterministic QoS needs of future Cloud and mobile Cloud services, the concept of virtualization across the technology domains is adopted. To optimally plan the proposed virtual infrastructures (VIs), in terms of specific objectives such as energy consumption and resource requirements, a holistic approach considering jointly the presence of all network technology domains and the IT resources is applied. Our modelling results identify trends and trade-offs relating to end-to-end service delays as well as resource requirements and energy consumption levels of the infrastructure across the various technology domains.

Figure 1: Virtualization over heterogeneous network infrastructures

2. Architecture and Network Technologies The infrastructure model proposed aims at providing a technology platform interconnecting geographically distributed computational resources that can support a variety of Cloud and mobile Cloud services. The proposed architecture comprises an advanced heterogeneous Copyright © 2013 The authors

www.FutureNetworkSummit.eu/2013

Page 2 of 9

multi-technology network infrastructure integrating optical metro and wireless access network domains interconnecting data centres (DCs) and adopts the concept of physical resource virtualization across the technology domains involved Figure 1. In contrast to the existing solutions that use small DCs in the wireless access and large DCs in the core to support mobile and fixed cloud traffic, respectively, the proposed solution relies on a common DC infrastructure fully converged with the broadband wireless access and the metro optical network. The technology domains comprising the converged network infrastructure are described below. 2.1

Wireless Access and Backhauling Solutions

In this study, the wireless domain of the physical infrastructure (PI) assumes a heterogeneous topology comprising a cellular LTE system [7] for the wireless access part and a collection of wireless micro wave links for the interconnection of the LTE-enabled based stations and the edge nodes of the optical metro network solution. LTE is among the prime wireless access cellular technologies in the fourth generation (4G) standard. LTE is anticipated to offer a theoretical net bit-rate capacity of up to 100 Mbps per sector in the downlink and 50 Mbps per sector in the uplink if a 20 MHz channel is used. These data rates can be further increased if multiple-input multiple-output (MIMO) technology is adopted. At the same time LTE marks the transition from a circuit switched to all-IP network architecture enhanced with greatly improved QoS characteristics such as low packet transmission delays (smaller than 5 ms for small packets), fast and seamless handover from one cell to another supporting up to 350Km/h for high speed vehicular communications scenarios, operation with different bandwidth allocations (scalable bandwidth up to and including 40 MHz while operation in wider bandwidths e.g. up to 100MHz is also possible). Furthermore, LTE can support a wide range of services and performance requirements (e.g. real and non-real time streaming, conversational and interactive services with low or high delay as well as background) in a wide range of environments such as indoor, urban and rural. 2.2

Optical Metro Network Solution

To support the metro network segment, we propose the use of a wavelength division multiplexed (WDM) optical network technology referred to as the Time Shared Optical Network (TSON) [8], [9]. TSON is designed and implemented as a novel frame-based, time multiplexing metro network solution, offering dynamic connectivity with fine bandwidth granularity. TSON offering sub-wavelength granularity is able to support short-lived connections and facilitate fast service delivery (as low as 300 ȝs), low end-to-end delay and multiple levels of guaranteed QoS. In this context, it is well suited to support Cloud and mobile Cloud services. In the proposed scenario where TSON is integrated with the wireless access LTE network supporting mobile users, mobility is accommodated by TSON through reallocation of services to different DCs depending on the relevant location of the end users applying the concept of virtual machine (VM) migration. Furthermore, fixed and mobile cloud traffic differentiation is achieved through prioritization/sorting of the Ethernet frames. TSON edge nodes provide the interfaces between the wireless and the optical domains as well as optical and DC domains. The ingress TSON edge nodes are responsible for traffic aggregation and mapping, while the egress edge nodes have the reverse functionality. The TSON core nodes switch transparently the optical frames to the appropriate output port utilizing fast optical switching bas7ed on PLZT technology. The switching time of the core TSON node is 10 ns.

Copyright © 2013 The authors

www.FutureNetworkSummit.eu/2013

Page 3 of 9

2.3

Converged Infrastructure

Mobile Users

A critical function in the converged infrastructure proposed is that of the interfaces between the different technology domains, which have to take care of the mapping and aggregation/de-aggregation of the traffic from one domain to the other. Figure 2 illustrates a general representation of the converged infrastructure indicating the interfaces between the different technology domains i.e. wireless to optical network domain and optical network to DC domains, as functional blocks. FIFO #1

Fixed Traffic

FIFO #1 ….

Rx FIFO 1

Tx FIFO Ȝ1

Rx FIFO 1

FIFO #1

Rx FIFO 2

Tx FIFO Ȝ2

Rx FIFO 2

FIFO #1

Tx FIFO Ȝ3

Rx FIFO 3

Tx FIFO Ȝ4

Rx FIFO 4

Buffer

Buffer Rx FIFO 3

Aggregator FIFO #N

Rx FIFO 4

Wireless

Buffer

De Aggregation

TSON

Buffer

….

Scheduler

FIFO #M

DC

Figure 2: Multi-Queuing model for the converged Wireless-Optical Network and DC Infrastructure

More specifically, the edge TSON nodes receive Ethernet frames carrying traffic generated by fixed and mobile users and arrange them to different buffers that are part of the node. These Ethernet frames are aggregated into TSON frames, which are then assigned to a suitable time-slot and wavelength for further transmission in the network on a First In First Out (FIFO) basis [8]-[9]. When these frames reach the interface between the optical and the DC domains the reverse function takes place to allow scheduling of demands to the suitable computing resources.

3. Planning the Optimal Virtual Infrastructure A key enabler of the proposed solution is the concept of cross-domain and cross-technology virtualization for the creation of infrastructure slices. This involves the abstraction of the physical resources into logical resources that can then be assigned as independent entities to different VIs. The virtualization process involves planning of the VIs i.e. identification of the optimal VIs that can support the required services in terms of both topology and resources and mapping of the virtual resources to the physical resources. Figure 1 illustrates the general architecture proposed, including a virtual layer formed over the physical layer. 3.1

Problem Formulation

In this paper we propose to plan VIs considering jointly the presence of all network technology domains and the IT resources incorporated in the PI with the aim to offer optimized VIs in terms of specific objectives. In highly dynamic heterogeneous environments the problem of optimal VI planning is complex since information regarding the position and the application requirements of the mobile devices, the available resources in the DCs and the optical network as well as the performance of the wireless network domain is uncertain. In order to assess the performance and requirements of this type of VIs we have developed, an optimization scheme suitable for VI planning taking into account both the dynamicity and mobility of end users over an integrated IT and heterogeneous network infrastructure suitable for Cloud and mobile Cloud services. In addition, the end-to-end delay in service delivery is quantified considering the varying degrees of latency introduced by the different technology domains. The PI under consideration is described in detail in section 2. Traffic demands corresponding to traditional Cloud applications are generated at randomly selected nodes TSON nodes in the wired domain and need to be served by a set of IT servers. Mobile traffic on the other hand is generated at the wireless access domain and in some cases needs Copyright © 2013 The authors

www.FutureNetworkSummit.eu/2013

Page 4 of 9

to traverse a hybrid multi-hop wireless access/backhaul solution before it reaches the IT resources through the optical metro network. The granularity of optical network demands is a portion of wavelength (e.g. Ȝ/100), while the IT locations at which the services will be handled, are not specified and are of no importance to the services themselves. Therefore, identification of the suitable IT resources that will support the services is part of the optimization output. In the general case, the VI planning problem should be solved taking into account a set of constraints that guarantee the efficient and stable operation of the resulting infrastructures. A main assumption is that every demand has to be processed at a single IT server. This allocation policy reduces the complexity of implementation. To formulate this requirement is introduced to indicate whether demand is assigned to server the binary variable or not. This assumption is expressed through the following equation:

For each demand , it’s demand volume is realized by means of a number of paths in the VI. Assuming that the vector is used to indicated the candidate path list in the VI required to support demand at server and vector the non-negative number of lightpaths allocated to path , the following demand constraints should be satisfied: Summing up the paths through each link link capacity of link

of the VI we can determine the required

where is a binary variable taking value equal to 1 if link of VI belongs to path realizing demand at server ; 0 otherwise. Using the same rationale, the capacity of each link in the VI is allocated by identifying the required paths in the PI. The resulting PI paths determine the load of each link of the PI. Since the PI consists of a heterogeneous network integrating optical metro and wireless access domains, it is assumed . For example, the wireless that each link has a different modular capacity backhaul links are treated as a collection of wireless micro wave links of 100 Mbps [13] while the TSON solution may offer a minimum bandwidth granularity of 100Mbps. Finally, the number of modules of type on link is denoted by A major issue to be taken into account is the distinction between traffic that arises from fixed and mobile devices. The accurate estimation of the resources that should be reserved in the VI to ensure seamless end-to-end service provisioning is based on the mobility model that is adopted, the size of the LTE cells as well as the traffic model used. In the ideal case, a seamless handoff for a mobile device can be 100% guaranteed if the equivalent amount of resources is reserved at all its neighbouring cells. However, a more efficient approach would be to relate the reserved resources in the neighbouring LTE cells with the handoff probability, . be the PI’s For the static scenario where handovers are not considered, let candidate path list realizing virtual link capacity . The following VI capacity constraint should be satisfied: where the summation is taken over all paths on the routing list of link and is the capacity of path realizing virtual link . However, in case of mobility an additional amount of resources should be leased across the various technology domains to ensure Copyright © 2013 The authors

www.FutureNetworkSummit.eu/2013

Page 5 of 9

seamless handoffs. To achieve this, the virtual capacity should be realized by a potential set of paths with capacity that will be used after the hand over with probability . Hence, introducing the link-path incidence coefficients for the PI taking values equal to 1 if link of PI belongs to path realizing link , 0 otherwise, the general formula specifying the PI capacity constraint can be stated as: At the same time, the planned VI must have adequate IT server resources such as CPU, memory, disk storage to support all requested services. A final consideration that should be taken into account is that, depending on the type of service that has to be provided, the end-to-end delay across all technology domains should be below a predefined threshold. Details regarding the extraction of this constraint are provided in Subsection 3.2. The objective of our formulation is to minimize the total cost during the planned time frame of the resulting network configuration that consists of the following components: a) the cost for operating capacity of the PI link of module type , and b) the total cost of the capacity of resource of IT server for processing the volume of demand . The costs considered in our modelling are related to the energy consumption of the infrastructure which can be directly associated with the operational expenditure (OpEx) of the planned VI. For the data centres, the power consumption model presented in [10], [11] has been adopted where for a service rate of Mbps, the corresponding power consumption is defined as , where subscript indicates the server, is the baseline consumption in the idle state and and is the slope of the load-dependent consumption. In the optical network domain, the cost for each optical link comprise the energy consumed by each lightpath due to transmission and reception of the optical signal, optical amplification at each fiber span and switching [12]. The switching power consumption of the TSON solution is based on actual lab measurements. Furthermore, for the wireless backhaul, the power consumption model presented in [13] has been adopted where wireless backhaul links are treated as a collection of wireless micro wave links of 100 Mbps capacity and a power dissipation of 50 W each. Thus for a given average backhaul requirement per base Mbps, the backhaul power consumption is 0.5W/Mbps. Finally, in the wireless station, access domain, the power consumption model of the LTE-enabled base station presented in [14] has been adopted. 3.2

Modelling Latency

So far, the MILP problem formulation guarantees that the capacities of the virtual and physical links will be adequate to support the transmission of the cloud computing services over the network segment. However, the total delay introduced by the heterogeneous . technology domains should be limited below a specific acceptable threshold, namely To achieve this, initially, a closed form approximation for the end-to-end delay is extracted after applying the Jackson’s Theorem to the multi-queuing model of the converged infrastructure presented in Figure 2 [15]. All queuing systems have been modelled as M/M/c queues with infinite buffer depth. Therefore, assuming that the conditions of the BCMP theorem are satisfied, the mean end-to-end cloud delay can be described through: In (7), , and are the expected delays in the wireless access, TSON network and the DC infrastructures assuming that , and physical resources have been allocated to the VI. Copyright © 2013 The authors

www.FutureNetworkSummit.eu/2013

Page 6 of 9

Figure 3 illustrates the end-to-end delay for mobile cloud services and the cloudlet approach, when adopting the proposed architecture. This delay does not take into consideration the propagation delay in TSON as given the dimensions of the optical rings (