A Framework for Joint Wireless Network Virtualization and ... · implemented or in contention with 5G include densi˝cation ... cognitive radio, beam division multiple access(BDMA),cloud-basedradioaccessnetworks(CRAN)

Received June 20, 2017, accepted August 18, 2017, date of publication August 30, 2017, date of current version October 25, 2017.

Digital Object Identifier 10.1109/ACCESS.2017.2746666

A Framework for Joint Wireless NetworkVirtualization and Cloud Radio AccessNetworks for Next GenerationWireless NetworksMOHAMAD KALIL1, ARAFAT AL-DWEIK2,3, (Senior Member, IEEE),MOHAMED F. ABU SHARKH4, (Member, IEEE),ABDALLAH SHAMI2, (Senior Member, IEEE), ANDAHMED REFAEY2,5, (Senior Member, IEEE)1IBM Analytics, IBM, Toronto, ON L3R 9Z7, Canada2Western University, London, ON N6A 5B9, Canada3Khalifa University, Abu Dhabi AUH 127788, United Arab Emirates4Ferris State University, Grand Rapids, MI 49307, USA5Manhattan College Parkway, Riverdale, NY 10471, USA

Corresponding author: Mohamad Kalil ([email protected])

This work was supported by the Telecommunications Regulatory Authority, the ICT Fund, Abu Dhabi, United Arab Emirates, under Grant11/15/TRAICTFund/KU.

ABSTRACT Wireless network virtualization (WNV) and cloud radio access networks (CRANs) arepromising technologies with the potential to be game changing for the fifth generation (5G) wirelessnetworks. In particular, these technologies may have significant impact on the capital expenditure, qualityof service provisioning, as well as spectral efficiency in 5G networks. These two technologies are mostlyconsidered separately in previous works. This paper, however, investigates both the gains and requirementsof integrating WNV with CRAN. In this paper, we propose WNV schemes for CRAN, where the objectiveis to maximize the overall system throughput and minimize delay. The proposed schemes are designed tomaintain a high level of isolation between mobile network operators (MNOs), which allows the deploymentof different scheduling polices by different MNOs, and managing intercell interference, which may leadto significant throughput gain. Overall, the results presented in this paper reveal that a joint CRAN-WNVarchitecture can be highly efficient when MNOs have unbalanced loads, because MNOs with high loads canseamlessly access the underutilized resources of underloaded MNOs. The throughput gain in unbalancedloads can be as much as 50% using optimal sharing schemes when compared with static sharing, andabout 18% when compared with the WNV without CRAN. The resource allocation problem in the jointCRAN-WNV is formulated, and both optimal and low complexity suboptimal solutions are derived. Theobtained results show that integrating the two technologies in a joint architecture can significantly improvethe network performance. However, reducing the complexity by adopting efficient sharing techniques mayhave tangible impact on the throughput when compared with optimal sharing.

INDEX TERMS Wireless resource virtualization, CRAN, ICI management, resource allocation.

I. INTRODUCTIONIn recent years, mobile data traffic has experienced substan-tial growth due to a combination of the increased availabilityof new devices along with the individual user data demandsurge caused by the plethora of data hungry applications.In 2016, the global mobile data traffic grew by 63% withalmost half a billion mobile devices and connections newly

added compared to 2015 [1]. Moreover, mobile data trafficgrowth rates are expected to be monotonically increasinguntil 2021, where it is anticipated to increase by sevenfold.As a result, mobile network operators (MNOs) are in needfor cost-cutting scalable solutions that can offer enhancedutilization and coverage of their wireless networks to copewith the mobile data market growth. A possible solution to

208142169-3536 2017 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.

VOLUME 5, 2017

M. Kalil et al.: Framework for Joint WNV and CRANs for Next Generation Wireless Networks

these challenges is revamping the network technology stackas part of the introduction of fifth generation (5G) networks.With such demand increase, the planned improvement intro-duced with 5G in terms of capacity, speed and efficiency willbe vastly absorbed. Examples for key enabling technologiesimplemented or in contention with 5G include densificationof existing cellular networks using extra-small cells, provi-sion of peer-to-peer (P2P) communications, multi-tier het-erogeneous networks, full duplex communication, massivemultiple-input multiple-output (massive MIMO), millimeter-wave technologies, cognitive radio, beam division multipleaccess (BDMA), cloud-based radio access networks (CRAN)and wireless networks virtualization (WNV) [2].

The latter two technologies are, at least in principle,game changing. Starting with WNV, the effectiveness of theconcept of virtualization is already a reality in terms ofinfrastructure as a service (IaaS). In less than 5 years, cloudcomputing IaaS offerings havemoved from a fledgling idea tothe dominant technology that 90% of large corporations useto operate their infrastructure. The success of the cloudmodelstems from its three defining features where virtualizationis the first and foremost, second is the dynamic scalability,and finally, the economic flexibility of providers offeringpay-as-you go portfolios.

The massive penetration and client satisfaction inspiredthe introduction of virtualization proposals of other criticalresources. Software defined networks (SDN) and networkfunction virtualization (NFV) are now materializing as themode of operation for networks intertwined within clouddata centers. However, in the wireless service operators’realm, things are less straightforward. Explicitly, virtualizingwireless resources carries a set of challenges that are yetto be addressed by wireless networks designers even as theintroduction of 5G is approaching. Some of these challengesoriginate from the lack of consensus between MNOs sharingthe virtualized resources either due to business reasons or dueto the difference in implementation vision. Some are techni-cal challenges that are yet to find a satisfactory solution thatwould facilitate immediate deployment.

The challenges arise with the definition of wirelessresource virtualization; (what is to be virtualized and shared?)and go on to span fairness of usage and end with imple-mentation nuts and bolts such as computational challengeswithin a base station. The main attraction of WNV to theC-level executives at MNOs is the direct savings on capitalexpenditure (CAPEX). WNV enables MNOs to share variousresources such as network infrastructure, backhaul, licensedspectrum, core and radio access network, as well as electricalpower, yielding better efficiency in terms of energy consump-tion and resource utilization. Slashing the carbon footprintis a priority for MNOs both to satisfy imposed governmen-tal and environmental requirements and to bring expensesunder control. This is no minor issue as energy consump-tion by communication components was about 50 gigawattin 2012 [3]. In addition, theoretically, the spectrumutilization should benefit from direct positive impacts.

Spectrum utilization improvement is becoming a red hotpriority for MNOs. This is because despite the increasingdemand for wireless spectrum, stark underutilization of spec-trum is a reality that shows in spectrum utilization measure-ments [4]. A Nokia experiment [5] reports that only 20% ofthe radio access networks maximum capacity is utilized mostof the time while 80% is idle except for peak time.

Implementing WNV has the potential to enhance resourceutilization, which can be seen in the following two scenarios.First, with WNV, an MNO with underutilized resources canseamlessly handover resources to overloaded MNOs. Evenin the scenario where all MNOs are overloaded, poolingmore wireless resources and then sharing them among a largenumber of users would increase the multiplexing gain. Takethe case of Rayleigh fading channels, the aggregated capacityof a cell can increase by ln(K ), whereK is the number of usersin the cell. In total, WNV has the potential to save billions inoperating expenditure (OPEX) and CAPEX [6]–[8]. Savingson CAPEX and OPEX in terms of maintenance and skilledpersonnel are self evident.

CRAN is another technology that is to be incorporated inwireless networks separately [2]. CRAN architecture is basedon decoupling the baseband processing unit (BPU) and aradio access unit (RAU) in every base-station (BS) as shownin Fig. 1. Therefore, a cell site in CRANs consists only of alow cost RAU, denoted as remote radio head (RRH) whileall BPUs are pooled in a cloud at remote data centers andare shared by all BSs. CRAN can achieve the same perfor-mance utilizing a number of BPUs lower than that neededby RAN [3]. This design can achieve power consumptionreduction of about 71% [9].

Unlike the traditional RAN, CRAN architecture facilitatesexchanging of traffic and channel information over the entirenetwork, and hence, can be used to effectively apply network-wide WNV to optimize the allocation of system resourcesacross multiple cell sites and MNOs.

One of the main issues affecting the poor efficiency ofcurrent wireless networks is the high variance of traffic loads.To provide high quality of service (QoS), the network shouldbe designed to accommodate peak traffic loads in orderto achieve low blocking probability. This, in turn, resultsin resource over-provisioning and low network efficiency.Presently, a straightforward resolution does not exist underthe rigid structure of current wireless networks. Scalable andflexible resource allocation techniques must be consideredto handle a broad variety of traffic loads at different times.Therefore, the design of high capacity networks with efficientutilization of the infrastructure in such scenarios is mostlyunattainable.

In fact, the maximum exploitation of WNV concept isachieved when the sharing MNOs have different load pro-files. More specifically, this is ideal when there is a mixedbandwidth demand profiles. In such cases, MNOs with highdemands can seamlessly access the excess resources of otherMNOs, which would maximize utilization of the sharedresource pool since it is constructed in a way similar to

VOLUME 5, 2017 20815


FIGURE 1. CRAN LTE network architecture.

FIGURE 2. Virtualized CRAN shared between two MNOs.

how a cloud data center is. This opens dynamic scalabilityvenues for MNOs with different sizes. WNV is not onlyeffective when various loads are at the same location. It canbe effective at mitigating coverage and QoS issues in caseswhere demand is high for everyone. Consider cases wherethis high demand is faced by certain geographic areas at acertain time. An example would be a sport stadium or a socialevent. A CRAN-WNV would be primarily effective here aswireless resources should be dynamically assigned acrossMNOs as well as across cells. Cloud resources can easilybe re-deployed to focus on the high demand area and hence,impacting coverage and QoS.

Despite the apparent advantages of employing WNV tech-niques, current network architectures such as long term evo-lution (LTE) supports coordination only between adjacentbase stations (BSs), denoted as evolved Node B (eNB), in apeer-to-peer fashion over X2 interface, which makes WNVdeployment and resource allocation across the entire networkprohibitively complex [10]. Therefore, WNV for traditional

LTE networks can only be applied, in its current form, on asmall scale over a few adjacent eNBs.

We hypothesize that CRAN, once it is integrated withWNV architecture, can be very effective in mitigating boththe dynamic traffic requirement challenge and the lack ofinformation exchange between cells. However, a compre-hensive scheme that facilitates that integration is not yetavailable. The aim of this paper is to put forth a solution thatvirtualizes the wireless resources of cloud-based RANs suchthat they can be shared between multiple MNOs as shownin Fig. 2. The allocation of wireless resources is determinedby the sharing contract between MNOs, the traffic load ateach RRH, and the interference between different RRHs.

The rest of the paper is organized as follows: Relatedresearch work is presented in Section II followed by thecontributions of this work. The proposed system and sharingmodels are presented in Section III. the optimal solutionis formulated in Section IV. Low complexity solutions arepresented in Section V. Experimental results and analysis

20816 VOLUME 5, 2017


are depicted in Section VI, and Section VII concludes thepaper.

II. RELATED WORK AND CONTRIBUTIONSIn the recent literature, it can be clearly noticed that thereis a growing interest in virtualizing wireless resources in asingle base station [8]. For example, a WNV scheme foreNB is investigated in [7]. The scheme allows MNOsto implement different scheduling policies. However,the scheme does not consider network-wide virtualization norcoordination between interfering cell zones to prevent inter-cell interference (ICI). In addition, the scheme suffers fromhigh complexity because it requires solving two optimizationproblems to maintain isolation between MNOs. In the firstoptimization process, the resources of each MNO are allo-cated to their users. The allocation results of problem oneare then fed to the second problem as constraints such thatthe throughput each user obtains is equal or greater than thethroughput achieved when sharing is not considered.

An LTE air interface virtualization scheme is proposedin [11], where a hypervisor is added on top of the physicalresources. The hypervisor is responsible for virtualizing theeNB into a number of virtual eNBs that can be used by dif-ferent MNOs. It is shown that more capacity can be achievedby sharing spectrum resources between different MNOs.However, the scheme does not provide optimal solutions normanage ICI. Furthermore, the instantaneous channel qualityof users is not considered in the scheduling decisions, whichlimits multiuser diversity gain.

More practical scenarios that consider load balancing arestudied in [12], where the hypervisor manages the sharingprocess of multiple eNBs among multiple MNOs. Neverthe-less, only fixed resource allocation across BSs is considered.The load is balanced between multiple BSs by moving usersfrom high-traffic cells to low-traffic cells. However, transfer-ring users across cells increases handover overhead, and maydegrade the system capacity since users may be transferredto BSs further away, which would reduce the quality of thewireless link.

An efficient low-complexity scheme that is able to vir-tualize the wireless resource blocks (RBs) at an eNB andshare them between users of multiple MNOs is proposedin [13]. The scheme aims at maximizing the throughputof MNOs users while maintaining access proportional fair-ness among users as well as MNOs. However, the pro-posed scheme does not consider network-wide virtualization.A utility-based resource provisioning scheme for WNV withmassive MIMO is investigated in [14]. A single BS equippedwith a large number of antennas serves users of differentservice providers. The problem is formulated as a combinato-rial optimization problem of high computational complexity.Consequently, a low-complexity solution for the combinato-rial problem is derived by linear programming relaxation.

Cross-layer resource optimization in CRAN has receivedincreasing attention in the last years, most of which inves-tigates the network utility maximization including spectrum

utilization, energy efficiency and proportional fairness [3].With CRANarchitecture, a larger number of RRH can be con-sidered in coordinated multipoint CoMP transmission, whichimproves the transmission performance at the expense of sig-naling overhead. The tradeoff between system capacity andsignaling overhead has been studied in [15]. User clusteringfor downlink CRAN is studied in [16]. Cooperative clustersfor every user is formed such that the network capacity ismaximized under limited-capacity backhaul. The impact ofsingle and N-nearest RRH association strategies on CRANergodic capacity is investigated in [17], where it is shownthat the ergodic capacity gain is not linearly proportional tothe RRH density. Therefore, each user should not associatewith more than 4 RRHs to balance the performance gain andimplementation cost.

A resource sharing scheme for a CRANwith capacity con-strained fronthaul links is proposed in [18], where a networkoperator lends radio resources to several service providers aswell as controls user admission and association. A threshold-based policy is introduced to maintain isolation among dif-ferent service providers, which controls the interferenceamong RRHs.

An auction mechanism is designed in [19] to sched-ule the radio and fiber links resources available in CRANwhile maximizing social welfare. The auctioneer owns theCRAN infrastructure, where bidders are mobile operators.Consequently, the auctioneer sells radio resources at BSsand fiber links at fronthaul network. However, the authorsassumed that the spectrum resources at RHH are predefinedand they do not scale.

Minimizing network power consumption of CRAN isinvestigated in [20], where the power consumption of thetransport network and RRHs is considered. The authorsassume that transport links and RRHs can support sleepmode. The problem is formulated as a joint RRH selectionand power minimization beamforming problem. The networkpower consumption is reduced by minimizing the numberof active RRHs and reducing their transmit power subject toQoS constraints. Through simulations, the authors show thatthe network power consumption can be notably reduced.

Energy efficiency optimization for heterogeneous CRANsis investigated in [21], where RB assignment and powerallocation are optimized subject to particular QoS, inter-tier interference and maximum transmit power conditions.Using Lagrange dual decomposition method, a closed-formexpression is derived for the optimization problem. A cross-layer resource allocation scheme for CRAN that minimizespower consumption in BPU pool, fiber links, and the RRHs isreported in [22]. The proposed scheme minimizes the powerby optimizing RRH selection, joint beamforming, and elasticservice scaling according to users’ QoS requirement such asthe system expected delay being less than a certain threshold.

In this work, we consider a network architecture whereCRAN and WNV are jointly applied to maximize the net-work spectrum utilization. The main contributions of thispaper are:

VOLUME 5, 2017 20817


FIGURE 3. Virtualized CRAN shared between two MNOs.

• An optimal scheme that enables spectrum sharingbetween multiple MNOs and RRHs is proposed and for-mulated. The scheme eliminates ICI and considers fairdistribution of spectrum resources between RRHs basedon their traffic loads. The scheme allows differentMNOsto apply different customized resource scheduling poli-cies, and offers efficient resource utilization across thenetwork. Moreover, the scheme maintains high levelof isolation between MNOs such that every MNO isguaranteed to be satisfied or access its share of theresources.

• A suboptimal scheme that solves the optimization prob-lem with lower complexity is derived. The suboptimalscheme is obtained by dividing the wireless resourceallocation problem into sub-problems. The objective ofeach sub-problem is to allocate one RB to a set of non-interfering RRHs. The allocation per single RB is formu-lated as a maximum weighted independent set problem,which is solved using binary integer programming (BIP)solvers.

• To further reduce the complexity, a low-complexityheuristic algorithm that solves the BIP problem is pro-posed. The algorithm is greedy and finds the set ofnon-interfering RRHs for each RB iteratively. The time

complexity of the heuristic algorithm is considerablylower than the BIP scheme while its throughput anddelay performance are comparable.

It is worth mentioning that this work focuses on virtual-izing and sharing the spectrum resource between differentMNOs, virtualizing other resources such as BPUs and fron-thaul network resources can be found in [9].

III. SYSTEM AND SHARING MODELSConsider the downlink of a cloud-based RAN architectureshared between M MNOs, where N RRHs are distributedto cover a certain geographical area as seen in Fig. 3. Thehypervisor, which resides in the cloud, is responsible forsharing the available spectrum bandwidth between MNOs’users and sites.

The total number of users in the system is K , and usersare labeled by a unique index k ∈ [1, 2, · · · ,K ]. As LTE isthe most prominent wireless communication standard to bedeployed as CRAN [9], LTE physical layer model is assumedin this work. Therefore, orthogonal frequency-division mul-tiple access (OFDMA) is used for the downlink transmission.Using LTE terminology, the overall bandwidth of the net-work is divided into R frequency slots, denoted as RBs, eachof which consists of 12 subscribers and occupies 180 kHz

20818 VOLUME 5, 2017


TABLE 1. Summary of the most significant notation.

bandwidth. To facilitate readability, Table 1 summarizes thenotations frequently used throughout the paper.

Each RRH is assumed to be capable of transmitting overany RB. The RRHs are connected to a virtualized pool ofBPUs in remote data centers via transport networks such asoptical transport networks. The Virtualized pool of BPUsis shared by different MNOs. Wireless channel state andtraffic rate information are available at the BPU pool, whichenables efficient scaling of the computation capacity [22].Since BPUs for all RRHs are co-located in one BPU pool,information on RRHs can be easily shared across the datacenter. Therefore, it is assumed that the wireless link qualityfor each user and RRHs traffic loads are known at the hyper-visor, which is responsible for resource allocation decisions.

Generally speaking, each UE can be served by more thanone RRH. However, in order to reduce the computationalcomplexity of the proposed algorithms, we assume that eachUE is connected to one RRH.Without loss of generality, eachUE is assumed to connect to the nearest RRH.

A. ICI COORDINATIONICI Coordination Schemes for CRAN LTE mobile networkare classified into two main groups [9]. The first groupeliminates ICI using frequency reuse schemes by allocatingdifferent frequencies to interfering zones. Such schemes arerelatively simple, as they do not require synchronizationbetween cells. The second group utilizes interference pathsconstructively, such as CoMP and Joint Transmission (JT)[23]. Although it is proven that using CoMP can improve

the system throughput, it requires dynamic coordination oftransmission across multiple cells, which requires tightersynchronization and coordination. In this work, ICI is elimi-nated by dynamically allocating distinct RB sets to interferingcells. Furthermore, RRHs can serve their users independently,which reduces the complexity of the scheduling problem.

It is worth mentioning that such assumption is valid in Het-erogeneous CRAN, where interference between low powernodes is minimum [21]. In addition, the assumption thatMacrocell users are allocated different RBs in CRAN iscommonly used in the literature [24]–[26].

Thus, the signal-to-noise ratio (SNR) of the link betweenUE k and RRH n over RB r can be expressed as

γr,k =Pr,nHr,n,kσk2

(1)

where Hr,n,k is the channel gain of the wireless link betweenRRH n and UE k over RB r , Pr,n is the transmit power ofRRH n over RB r , and σ 2

k is the additive white Gaussiannoise (AWGN) variance.

Although it is proven that power allocation schemes suchas the water-filling algorithm can improve the transmissionefficiency, integrating such techniques in mobile systems ischallenging because they require tight tracking of the rapidchannel variations [26]. Moreover, it is shown that whenthe transmission power is high, equal power allocation is asefficient as the optimal power allocation [27]. In this work,we assume equal power allocation similar to the 3GPP-LTEstandard [28], [29] such that Pu,n = Pv,n = Pn,∀{u, v} ∈ Rn,where Rn is the set of RBs assigned to RRH n. In fact,

VOLUME 5, 2017 20819


equal power allocation is a common assumption in resourceallocation problems [30]–[32].

LTE supports various modulation and codingschemes (MCSs) [33]. The selection of a particular MCS isdetermined by SNR and the required block error rate (BLER).A simple MCS selection scheme is performed using a lookuptable which maps the received SNR to a MCS for a certainBLER [33]. The maximum number of bits that can be trans-mitted to UE k over RB r can be calculated as

ur,k =⌊ξ (γr,k )Tsym

⌋(2)

where ξ (γr,k ) is the spectrum efficiency of the selectedMCS,Tsym is the total number of symbols in a single RB, andbxc refers to the floor function.

B. RESOURCE BLOCKS SHARING MODELRBs are assumed to be shared between MNOs based on acontract signed between them. RBs are either shared stat-ically, where each MNO accesses only its share of theresources, or dynamically, where MNOs can access the entireset of RBs. It is worth mentioning that, the static sharingmodel is used in the simulation as a benchmark to evaluatethe performance of the proposed dynamic sharing solutions.As wireless resource virtualization is still in its infancy stage,no well-defined sharing models exist yet [8]. Therefore,a general sharing model is assumed based on the followingconditions:

1) In the case of static sharing of RBs, MNOs are assumedto distribute their resources among cells such that thefrequency reuse factor is maximized while maintaininga proportional fairness criterion such that

max∑n

Sm,n (3a)

subject to

Sm,n +∑c∈Cn

Sm,c ≤ Rm, ∀n (3b)

Sm,1 : · · · : Sm,N = (1± α)(Lm,1 : · · · : Lm,N ), ∀m

(3c)

where Sm,n is the number of RBs allocated to MNO mat RRH n, Lm,n is the load of MNO m at RRH n,α is a small constant that relaxes the fairness constraintsin (3c) to ensure feasible solutions for the optimizationproblem, Rm is the number of RBs granted to MNO msuch that

∑mRm = R, and Cn is the set of RRHs that

interfere with RRH n.Constraint (3b) ensures that the number of RBs thatare assigned to MNO m at RRH n and its neighboringRRHs is less than MNO m share of the resources Sm,n.The Constraints in (3c) ensure that RBs are fairlyallocated to RRHs based on their load. The load canbe considered as the number of users or a number ofpackets queued in buffers for the users. As the fluc-tuation rate of the load in RRHs is slow compared

with the transmission time interval (TTI), which is 1msin LTE systems for the finest scheduling granularity,the optimization problem in (3) can be solved at acoarser granularity than TTI. Other sharing modelscan be applied here, however, maximizing the fre-quency reuse factor while considering a fairness cri-terion is an intuitive target that MNOs are looking toachieve.

2) In the case of dynamic sharing of the RBs, the servicestatus of MNO m at RRH n should be higher thana certain threshold or, if it is not the case, MNO mshould access at least Sm,n RBs. This condition ensuresisolation between MNOs such that all MNOs are eithersatisfied, or can access at least the same number ofRBs that they would access in the case of static shar-ing. The service status of an MNO can be related toaspects such as queue length of users’ buffers, spectralefficiency, or energy efficiency.

IV. PROBLEM FORMULATIONIn this work, eachMNO aims at maximizing its sumweighteddata rates, which is a very common optimization problem inwireless systems [30], [34]–[36]. The weights are selected byMNOs according to their scheduling policies. To simplify thenotations, assume that user k is connected to RRH n and everyRRH serves at least one user of each MNO. The schedulingproblem can be formulated as

maxM∑m=1

N∑n=1

∑k∈Km,n

R∑r=1

wkur,k βr,k

(4a)

subject to∑c∈Cn

∑k∈Kc

βr,k +∑

k∈Knβr,k ≤ 1, ∀n, r (4b)∑

r∈Rk

Tr,k ≤ qk , ∀k (4c)

(8m,n>8thm) or (Rm,n≥Sm,n) must hold, ∀(m, n) (4d)

where wk is the normalized weight for UE k ,Kn =⋃mKm,n is

the set of UEs connected to RRH n, Km,n is the set of UEssubscribed to MNO m and connect to RRH n, Rm,n is thenumber of RBs accessed by MNO m at RRH n, Rk is the setof RBs assigned to UE k , 8m,n and 8th

m are the service statusand service status threshold of MNO m at RRH n, and βr,k isa binary number indicator defined as

βr,k =

{1, if RB r is assigned to UE k0, otherwise.

Constraint (4b) represents the exclusive constraint whichensures that (i) each RB is assigned to one UE (at most) ateach RRH, and (ii) orthogonal sets of RBs are allocated toRRHs that may interfere with each other. It is assumed thatthe interference is avoided if interfering RRHs are grantedorthogonal sets of RBs. Constraint (4c) ensures that thetransport block size for every UE is less than its unserved

20820 VOLUME 5, 2017


data size, where Rk is the RB set that is assigned to user k .Constraint (4d) specifies whether the service status of MNOm at RRH n is higher than a certain threshold or, if thatis not the case, MNO m should access at least Sm,n RBs.This constraint ensures isolation between MNOs such thatMNOs are either satisfied, or can access at least the samenumber of RBs in case of static sharing. It is noteworthy thatconstraint (4d) can be split into two constraints by introducinga binary variable ym,n and a sufficiently large upper bound Bmso that

8m,n > 8thm − Bmym,n (5a)

Rm,n ≥ Sm,n − Bm(1− ym,n). (5b)

When ym,n = 0, constraint (5a) holds, whereas con-straint (5b) becomes Rm,n ≥ Sm,n − Bm, which is alwayssatisfied if Bm is large enough. Note that the constraintRm,n ≥ Sm,n may still be satisfied. When ym,n = 1, onlyconstraint (5b) holds. Consequently, one constraint holds, andthe other one may be satisfied.

The formulation in (4) allows MNOs to apply differ-ent scheduling policies by weighting their UEs differently.In addition, it guarantees that MNOs use their share of RBsat the overloaded RRH. However, if an MNO is underloadedat a specific RRH, its share of RBs can be granted to otherMNOs that are overloaded.

Various scheduling policies are proposed forLTE networks [37], including channel-aware policies, such asmaximum throughput (MT), proportional fair (PF), and gen-eralized PF (GPF); channel-aware and QoS-aware policies,such as modified largest weighted delay First (M-LWDF) andLOG rule; and energy-aware policies [38].

A. COMPLEXITY OF THE OPTIMAL SOLUTIONThe optimal solution can be formulated as a binary integerprogramming (BIP) optimization problem, which is compu-tationally infeasible once the problem exceeds trivial sizes.The problem’s complexity increases exponentially with thenumber of users, MNOs, and RBs. In order to give a glimpseof the BIP problem complexity, the complexity of the sub-problem of allocating RBs to RRHs is demonstrated. Let’stake a simple case where 3 RRHs are shared by 2 MNOs andinterference is possible. Let’s assume a network of 60 RBs.As mentioned earlier, orthogonal sets of RBs should beassigned to RRHs. If each RRH receives 20 RBs, the numberof combinations of 20 RBs chosen from 60 RBs for oneMNO is 60!

20!(60−20)! = 4.1918 × 1015. Even if we choosethe finest scheduling granularity (one subframe equaling 1msfor LTE), then the problem should yield a solution in less than1ms. This is practically infeasible. This necessitates develop-ing an algorithm of lower complexity and feasible runningtime.

This issue, like the case in many networking problems,exposes the architecture to vulnerability. In other words,the whole architecture can be perform as well low complexityscheduler.

B. SPECIAL CASE: BACKLOGGED TRAFFIC MODELAs discussed in Section IV-A, the complexity of the schedul-ing problem in (4) is considerably high, which prevents com-paring the optimal solution under dynamic traffic modelswith other schemes. However, the complexity of the optimalsolution can be significantly reduced under the backloggedtraffic model assumption, where users always have data totransmit. Backlogged traffic model assumption relaxes theconstraints (4c) because the data in users’ buffers are assumedto be larger than the transmitted data. Therefore, if RB r isassigned to MNO m at RRH n, the RB should be assignedto the user at RRH n who maximizes the weighted sumdata rates and belongs to MNO m. Although the backloggedtraffic model might not occur all the time in the network,it reduces the prohibitive complexity of the optimal solution.Hence, the backlogged traffic model is assumed for optimalsolution comparisons in Section VI. The scheduling problemfor backlogged traffic model can be formulated as

maxM∑m=1

N∑n=1

R∑r=1

umaxr,m,n βr,m,n (6a)

subject to ∑c∈Cn βr,m,c + βr,m,n ≤ 1, ∀r, n,m (6b)

(8m,n > 8thm) or (Rm,n ≥ Sm,n) must hold, ∀(m, n) (6c)

where umaxr,m,n = max∑

k∈Km,n

wkur,k and βr,m,c is a binary

number indicator defined as

βr,m,n =

{1, if RB r is assigned to MNO m at RRH n0, otherwise.

V. LOW-COMPLEXITY SOLUTIONSThe optimal method basically spans all possible RBs alloca-tions to users for all RRHs searching for the maximum utility.As shown earlier, this can be tricky. A possible suboptimaltechnique to mitigate complexity is to allocate RBs sequen-tially. The basic concept is that, at each iteration, the tech-nique aims to satisfy the MNO with the lowest satisfactionrate (LSM) allocating the RB to a user belonging to thatMNOwhich, in turn, maximizes the sum-utility.

The LSM at RRH n is defined as

sn=

{argmax

m(Sm,n−Rm,n),∀m∈Ln, if Ln 6= ∅

argmaxm

(Sm,n−Rm,n),∀m=1, 2, · · · ,M , otherwise.

(7)

where Ln is the set of MNOs that have 8m,n < 8thm . As the

objective is to maximize weighted sum utility, the assignedRB to the RRH is granted to the user who can maximizethe sum-utility. The utility and the index of the user whosubscribes to MNO sn and maximizes the sum utility can befound, respectively, as

zn,r = maxk∈Ksn,n

wkuk,r (8)

jn,r = arg maxk∈Ksn,n

wkuk,r (9)

VOLUME 5, 2017 20821


The optimization problem per RB can be treated as themaximum weighted independent set (MWIS) of a graphpath. Each RRH represents a vertex; an edge (line) is drawnbetween two vertices if they interfere with each other. A graphcan be described by the pair G = (V ,E), where the set Vis the vertices of G, and the set E is the edges of G. TheMWIS is the subset of vertices that has maximum weightedsum such that no two vertices are connected with an edge.For each RB, the MWIS can be formulated as a BIP problem.For example consider 10 cells and 2 MNOs. Assume thetotal number of users is 200, where each cell has 10 userssubscribed to each MNO, and the total number of RBs is 100.In the optimal solution, the number of decision variables is10 × 200 × 100 = 20,000. For the suboptimal scheme,the number of users that maximizes the sum-utility and sub-scribed to the LSM is 10. Therefore, the number of decisionvariables for each BIP problem is 10. As the number of RBsis 100, the suboptimal solution solves 100 BIP problems eachof which has 10 decision variables. Fig. 4 shows an exampleof a graph and its MWIS.

FIGURE 4. Example of interference graph G = (V , E) of five weightedvertices (RRHs), where V = {1, 2, 3, 4, 5}, and E = {(1, 4), (1, 5),(2, 4), (3, 5)}. The independent sets are = {1}, {2}, {3}, {4}, {5}, {1, 2}, {1, 3},{1, 2, 3}, {2, 3}, {2, 5}, {2, 3, 4}, {2, 4, 5}, {3, 4}, {4, 5}. The MWIS is {1, 2, 3}.

For each RB, the MWIS IS can be found by solving thefollowing optimization problem

IS = maxN∑n=1

vn βn (10a)

subject to ∑c∈Cn

βc ≤ 1, ∀n (10b)

where βn is a binary variable, equal to one if n ∈ IS and zerootherwise, and vn is the weight of vertex (RRH) n. In order to

bias the scheduler towards allocating RBs in favour of highlyloaded RRHs, the weights are chosen such that

vn =

{zn,r (Sm,n − Rm,n), if 8sn,n < 8th

snzn,r , otherwise.

(11)

Table 2 shows the pseudo code of an iterative low-complexity algorithm that solves (10) by using a BIP solver.At each iteration, one RB is assigned to the MWIS of RRHsthat maximizes the sum-weighted utility. Line 4 finds theLSM (sn) at every RRH, the index (jsn,r ) and the utilityvalue (zsn,r ) of the user who subscribes to MNO sn andmaximizes the sum utility, and the weight (vsn ) of the RRH sn.In line 6, the algorithm solves the MWIS optimization prob-lem (10) and finds the subset IS . The RB is assigned to users inline 7. The number of RBs assigned increases for each RRHthat belongs to IS in line 8.

TABLE 2. Per RB optimal allocation algorithm.

Although the algorithm solves the BIP problem R timeseach TTI, the complexity of the algorithm is relatively lowas compared to (4) because the size of the BIP optimizationproblem is significantly smaller than the problem in (4).In particular, the BIP hasN decision variables. However, for alarge number of RRHs, it might be computationally expensiveto solve the BIP problem shown in (10) R times every TTI.Therefore, a low-complexity heuristic algorithm that solvesthe BIP is presented in Table 3. The heuristic algorithm isgreedy in the sense that it assigns an RB to the RRH that hasthe maximum weight vn, then excludes its interfering RRHsfrom the allocation process. The first 6 lines in the heuristicalgorithm are similar to those in Table 2, where the LSMs andtheir users who maximize the sum utility are found. The RRHindex that has the maximum weighted utility n∗ is found inline 9 and is added to the subset IS in line 10. The RRH n∗

and its interfering RRHs indices are deleted from the potentialset of RRHs Sind . Consequently, interfering RRHs are notassigned the same RBs, thereby eliminating interference. AnRB is assigned to jn,r ,∀n ∈ IS in line 13, and Rsn,n is updatedin lines 14.

A. THE COMPLEXITY OF THE HEURISTIC ALGORITHMFor every TTI, the heuristic algorithm runs R major itera-tions (lines 3-21). Each major iteration finds the LSM anda candidate user jn,r for each RRH. Finding the LSM at RRHrequiresM operations, whereas finding the user (jn,r ) requires

20822 VOLUME 5, 2017


TABLE 3. Heuristic algorithm.

Ksn,n operations. The number of UEs is usually much largerthan the number of MNOs, which makes finding jn,r thedominating operation. Assigning each RB to the subset ISrequires at most N operations, assuming that no RRHs inter-fere with each other. Therefore, the worst-case complexity isO(R× (N +Kmax)), where Kmax is the maximum number ofusers that connect to an RRH and subscribe to one MNO.

Although the suboptimal schemes solve a BIP problem foreach RB, it has lower complexity than the optimal solutionbecause the size of its search space is significantly smallerthan that of the optimal solution. However, as the numberof RRHs increase, it will gradually struggle to solve the RBIP problems in feasible times. Therefore, the demand arisesfor a heuristic algorithm to solve each BIP independently intrivial time. We present a greedy algorithm to perform thatstep. In this algorithm, RB is assigned to the RRH with thehighest weight. All Interfering RRHs are excluded.

VI. SIMULATION SCENARIOS AND EXPLORATORYRESULTSIn this work, we consider a layout that comprises of 22 hexag-onal cells. The number of users for each MNO at each cellis assumed to be a uniform random variable with mean K .UEs are assumed to be uniformly distributed across each celland have an average SINR between 5 and 10 dB. The RBs areassumed to be independent and frequency-flat with Rayleighfading. Each RB is assumed to be fixed during one subframein the time domain and over one RB in the frequency domain,but changes independently over different subframes, differ-ent RBs, and different users which corresponds to a quasi-static channel.

For benchmarking purposes, the proposed schemes arecompared to the static sharing and the heuristic single-cellWNV (SCWNV) scheme presented in [7]. In the static shar-ing, each MNO at each RRH receives its share of RBs andallocates them to its users according to the MNO schedul-ing policy. In the SCWNV scheme, the spectrum resourcesassigned to all MNOs at a single RRH are virtualized andshared between all UEs connected to the RRH.

In the following two subsections, numerical results are pre-sented for two different scenarios. The first scenario considersthe optimal solution, and thus only 6 RRHs are consideredand the backlogged traffic model is adopted. In the secondscenario, the entire layout of the 22 cells is considered anda dynamic traffic model is assumed. Therefore, the optimalsolution is excluded due to its high complexity while theproposed suboptimal and heuristic schemes are consideredand compared to the static sharing and SCWNV schemepresented in [7].

A. SCENARIO 1: BACKLOGGED TRAFFIC MODELIn this case, throughput of the limited-size layout is evaluatedand compared using various schemes. The MT schedulingpolicy and the backlogged traffic model are adopted forboth MNOs. The backlogged traffic model implies that eachMNO is fully loaded and operates at full capacity. There-fore, the sharing gain is minimal and it is only due to themultiplexing gain of the wireless resources. To demonstratethe benefit of sharing resources between MNOs, users ofMNO2 are forced to hibernate for a random period of timewith an average value of Tsleep.

Fig. 5 compares the average throughput per UE for a dif-ferent average number of UEs per cell per MNO. As it can benoted from the figure, the general trend for all schemes is thatas the average number of UEs increases the average numberof RBs assigned to each UE decreases, which lowers theaverage throughput per UE. As expected, the performance ofnon-optimal WNV schemes is upper bounded by the optimalscheme and lower bounded by the static sharing scheme.Moreover, the proposed suboptimal and heuristic schemesoutperform the SCWNV scheme and they are only 9%worsethan the optimal.

FIGURE 5. Average throughput per UE for Tsleep = 40%.

The throughput advantage of the proposed suboptimal andheuristic approaches is due to the fact that users can access theentire set of the RBs. On the contrary, SCWNV limits usersto access only the RBs that are assigned to their RRHs.

VOLUME 5, 2017 20823


On the other hand, the performance of SCWNV schemeoutperforms the performance of the static sharing schemeas it virtualizes the resources of both MNOs at each RRH.Static sharing offers the lowest throughput because the UEsof eachMNO only access RBs that are dedicated to theMNO,and they cannot utilize the unused RBs of others MNOs.Although MNO1 and MNO2 own the same number of RBs,the average throughput per user is lower for MNO2’s usersbecause they were forced to hibernate Tsleep of the time. Theaverage throughput per user for the proposed suboptimal andheuristic schemes are slightly lower than that achieved by theoptimal scheme. The SCWNV scheme slightly outperformsthe static sharing scheme. However, the gain achieved by theWNV schemes for MNO2’s UEs is less than that for MNO1’sUEs for the following reason.WNV offers two types of gains:1) multi-MNOmultiplexing gains as a result of increasing thenumber of users per cell, and 2) gain results form sharing RBsof underloaded MNOs with overload MNOs. As MNO1 isfully loaded, WNV only offers multi-MNO multiplexinggains toMNO2’s users. On the other hand,MNO2 is assumedto be underloaded andMNO1’s users benefit from both gains.

Fig. 6 shows the average aggregate throughput per cell.As number of UEs increases, the average aggregate through-put increases as a result of the multi-MNOmultiplexing gain.Since MNO2 users are forced to hibernate 40 % of the time,MNO1 has higher average throughput than MNO2.

FIGURE 6. Average throughput per cell for Tsleep = 40%.

The average user throughput for different values of Tsleepfor users of MNO1 and MNO2 are shown in Figs. 7 and 8,respectively. It is worth noting that Tsleep indicates the loadof MNO2. As value of Tsleep becomes longer, the load onMNO2 becomes lighter. In case of static sharing, MNO1 andMNO2 are fully isolated from each other. Therefore, light-ening the load on MNO2 has no impact on MNO1, andthe average throughput of MNO1 users is constant for anyTsleep value. On the contrary, for WNV schemes the averagethroughput of MNO1’s users builds up as the MNO2 load

FIGURE 7. Average throughput per UE for different vales of Tsleep:MNO1 users.

FIGURE 8. Average throughput per UE for different vales of Tsleep:MNO2 users.

drops. As Tsleep increases, the users of MNO2 access lessRBs, and therefore their average throughput decreases or thebenefit of MNO1’s users.

B. SCENARIO 2: DYNAMIC TRAFFIC MODELIn this scenario, performance is measured through a crit-ical metric which is the average throughput and head-of-line (HoL) packet delay. A simulation environment consistentwith the first scenariowas employed (MNOs count, andRBs).However, a more realistic system model was chosen as weconsider the entire layout (22 RRHs) and allow MNOs toemploy varying scheduling policies. Namely, MNO1 usesthe M-LWDF scheduling policy, whereas MNO2 applies theMT scheduling policy. In this experiment, UEs are consid-ered active only when they have data to be transmitted.MNO1’s UEs have a Poisson traffic distribution with a packetarrival rate λ, and fixed-size packet of 1 KB. In contrast,

20824 VOLUME 5, 2017


MNO2’s UEs - to make sure the system is under sufficientload- are assumed to use a greedy distribution which alwaysrequires transmitting data.

We start by investigating the performance of the pro-posed, SCWNV, compared to static sharing schemes whenwe gradually change the traffic load by changing the packetarrival rate λ, to users’ buffers. The similarity in profilesbetween MNOs (number of RBs, scheduling policy, numberof UEs, and average SINR) dictates the resulting averageperformance by both MNOs as shown.

The average aggregate throughput per cell for users ofMNO1 and MNO2 is shown in Figs. 9 and 10, respec-tively, for different traffic loads of MNO1. As the value ofλ increases, the average data arrivals decreases and the loadof MNO1 becomes lighter. In the static sharing scenario,MNO2 cannot access RBs that are assigned toMNO1. There-fore, average aggregate throughput of MNO2 is not affectedby the variation in the traffic load of MNO1.

FIGURE 9. Average aggregate throughput: MNO1 users.

However, WNV schemes allow MNO2 to access theentire RB pool. Consequently, reducing the MNO1’s loadwill directly yield more available resources to MNO2.Hence, the throughput of MNO2 grows as the load ofMNO1 becomes lighter for all WNV schemes. The sub-optimal scheme slightly outperforms the proposed heuristicscheme at the expense of higher computational complexity.It is noted the proposed schemes all perform visibly betterthan the SCWNV and static sharing. This is seen in thedifference in the average normalized running time ofthe schemes: 2.52× 104 for the suboptimal, 0.712× 104 forthe heuristic, 0.544×104 for the heuristic [7], and 1 for staticsharing. In terms of complexity, we note that the static sharingscheme sacrifices performance by seeking less complexity.On the contrary, despite being slower than static sharing,heuristic method gets considerably better metric values whileexecuting at acceptable running speed that is faster than thesuboptimal and optimal scheme.

FIGURE 10. Average aggregate throughput: MNO2 users.

With higher values of λ which correspond to lower loads,all the algorithms are able to satisfy the demand. As trafficload increases, the distinction starts to materialize betweenWNV schemes and static sharing scheme for MNO1. Theproposed schemes clearly outperform the SCWNV and thestatic sharing schemes.

Fig. 11 shows the average HoL packet delay for differentvalues of λ. The averageHoL delay for the suboptimal and theheuristic schemes are similar, but much less than that for theSCWNV scheme and static sharing schemes for small valueof λ. For large values of λ, the load on MNO2 is relativelylow and the average HoL delay for all schemes converge.

FIGURE 11. Average head-of-line packet delay of MNO1’s UEs.

VII. CONCLUSIONWireless Network Virtualization (WNV) is a promising solu-tion to achieve next level gains in capital investment savings,energy efficiency and spectral utilization for 5G environ-ments. In this work, we hypothesized that combining WNV

VOLUME 5, 2017 20825


with Cloud Radio Access Network (CRAN) opens up venuesfor better resource utilization in cases where unbalanced loadbetween mobile network operators (MNOs) is faced. Signif-icant performance gains are also achieved when the wholeresource pool is under high demand. In addition, the proposedenhanced architecture has the potential to mitigate dynamictraffic control issues and data sharing between system com-ponents in addition to realizing WNV’s potential in termsof capital and operating expenditure savings. Optimal andsuboptimal solutions that combine WNV and CRAN whileconsidering wireless resources were proposed. Moreover,a practical solution is introduced in the form of a heuristictechnique that works for all problem scales. Experimentalresults show that the proposed CRAN-WNV architecture out-performs static sharing in terms of both aggregate throughputand delay. This validates the joint architecture design andmotivates further work to harvest more utilization gains fromCRAN-WRV architectures.

REFERENCES[1] (2016). Cisco Visual Networking Index: Global Mobile Data Traffic

Forecast Update, 2016–2021. [Online]. Available: http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/mobile-white-paper-c11-520862.html

[2] E. Hossain and M. Hasan, ‘‘5G cellular: Key enabling technologiesand research challenges,’’ IEEE Trans. Instrum. Meas., vol. 18, no. 3,pp. 11–21, Jun. 2015.

[3] M. Peng, Y. Sun, X. Li, Z. Mao, and C. Wang, ‘‘Recent advances incloud radio access networks: System architectures, key techniques, andopen issues,’’ IEEE Commun. Surveys Tuts., vol. 18, no. 3, pp. 2282–2308,3rd Quart., 2016.

[4] M. Höyhtyä et al., ‘‘Spectrum occupancy measurements: A survey anduse of interference maps,’’ IEEE Commun. Surveys Tuts., vol. 18, no. 4,pp. 2386–2414, 4th Quart., 2016.

[5] A. Aittokallio. Nokia Unveils Radio Cloud Architecture. Accessed:Sep. 07, 2017. [Online]. Available: http://telecoms.com/399321/nokia-unveils-radio-cloud-architecture

[6] D.-E.Meddour, T. Rasheed, andY. Gourhant, ‘‘On the role of infrastructuresharing for mobile network operators in emerging markets,’’ Comput.Netw., vol. 55, no. 7, pp. 1576–1591, 2011.

[7] M. Kalil, A. Shami, and Y. Ye, ‘‘Wireless resources virtualization inLTE systems,’’ in Proc. IEEE Int. Conf. Comput. Commun. (INFOCOM),Apr./May 2014, pp. 363–368.

[8] C. Liang and F. R. Yu, ‘‘Wireless network virtualization: A survey, someresearch issues and challenges,’’ IEEE Commun. Surveys Tuts., vol. 17,no. 1, pp. 358–380, 1st Quart., 2015.

[9] A. Checko et al., ‘‘Cloud RAN for mobile networks—A technologyoverview,’’ IEEE Commun. Surveys Tuts., vol. 17, no. 1, pp. 405–426,1st Quart., 2015.

[10] M. Hasan, E. Hossain, and D. Niyato, ‘‘Random access for machine-to-machine communication in LTE-advanced networks: Issues andapproaches,’’ IEEE Commun. Mag., vol. 51, no. 6, pp. 86–93,Jun. 2013.

[11] Y. Zaki, L. Zhao, C. Goerg, and A. Timm-Giel, ‘‘LTE wireless virtualiza-tion and spectrum management,’’ in Proc. 3rd Joint IFIP Wireless MobileNetw. Conf. (WMNC), Oct. 2010, pp. 1–6.

[12] M. Li et al., ‘‘Investigation of network virtualization and load balanc-ing techniques in LTE networks,’’ in Proc. IEEE 75th Veh. Technol.Conf. (VTC), May 2012, pp. 1–5.

[13] M. Kalil, A. Moubayed, A. Shami, and A. Al-Dweik, ‘‘Efficient low-complexity scheduler for wireless resource virtualization,’’ IEEE WirelessCommun. Lett., vol. 5, no. 1, pp. 56–59, Feb. 2016.

[14] V. Jumba, S. Parsaeefard, M. Derakhshani, and T. Le-Ngoc, ‘‘Resourceprovisioning in wireless virtualized networks via massive-MIMO,’’ IEEEWireless Commun. Lett., vol. 4, no. 3, pp. 237–240, Jun. 2015.

[15] B. Dai and W. Yu, ‘‘Sparse beamforming and user-centric cluster-ing for downlink cloud radio access network,’’ IEEE Access, vol. 2,pp. 1326–1339, Oct. 2014.

[16] D. Liu, S. Han, C. Yang, and Q. Zhang, ‘‘Semi-dynamic user-specificclustering for downlink cloud radio access network,’’ IEEE Trans. Veh.Technol., vol. 65, no. 4, pp. 2063–2077, Apr. 2016.

[17] M. Peng, S. Yan, and H. V. Poor, ‘‘Ergodic capacity analysis of remoteradio head associations in cloud radio access networks,’’ IEEE WirelessCommun. Lett., vol. 3, no. 4, pp. 365–368, Aug. 2014.

[18] B. Niu, Y. Zhou, H. Shah-Mansouri, and V. W. S. Wong, ‘‘Adynamic resource sharing mechanism for cloud radio access networks,’’IEEE Trans. Wireless Commun., vol. 15, no. 12, pp. 8325–8338,Dec. 2016.

[19] S. Gu, Z. Li, C. Wu, and H. Zhang, ‘‘Virtualized resource sharing incloud radio access networks through truthful mechanisms,’’ IEEE Trans.Commun., vol. 65, no. 3, pp. 1105–1118, Mar. 2017.

[20] Y. Shi, J. Zhang, and K. B. Letaief, ‘‘Group sparse beamforming forgreen cloud-RAN,’’ IEEE Trans. Wireless Commun., vol. 13, no. 5,pp. 2809–2823, May 2014.

[21] M. Peng, K. Zhang, J. Jiang, J. Wang, and W. Wang, ‘‘Energy-efficient resource assignment and power allocation in heterogeneous cloudradio access networks,’’ IEEE Trans. Veh. Technol., vol. 64, no. 11,pp. 5275–5287, Nov. 2015.

[22] J. Tang, W. P. Tay, and T. Q. S. Quek, ‘‘Cross-layer resource allo-cation with elastic service scaling in cloud radio access network,’’IEEE Trans. Wireless Commun., vol. 14, no. 9, pp. 5068–5081,Sep. 2015.

[23] Coordinated Multi-Point Operation for LTE Physical Layer Aspects,document 3GPP TR 36.819 V11.1.0, 2011.

[24] L. Liu, S. Bi, and R. Zhang, ‘‘Joint power control and fronthaul rate allo-cation for throughput maximization in OFDMA-based cloud radio accessnetwork,’’ IEEE Trans. Wireless Commun., vol. 63, no. 11, pp. 4097–4110,Nov. 2015.

[25] A. Abdelnasser and E. Hossain, ‘‘Resource allocation for an OFDMAcloud-RAN of small cells underlaying a macrocell,’’ IEEE Trans. MobileComput., vol. 15, no. 11, pp. 2837–2850, Nov. 2016.

[26] M. S. K. Lau, W. Yue, and L. Ping, ‘‘Equal power allocation of IDMAsystems: Feasibility, optimality, and throughput,’’ IEEE Trans. WirelessCommun., vol. 8, no. 2, pp. 746–756, Feb. 2009.

[27] B. Luo, Q. Cui, H. Wang, and X. Tao, ‘‘Optimal joint water-filling forcoordinated transmission over frequency-selective fading channels,’’ IEEECommun. Lett., vol. 15, no. 2, pp. 190–192, Feb. 2011.

[28] Physical Layer Aspects Forevolved Universal Terrestrial RadioAccess (UTRA) (Release 7), document 3GPP TR 25.814 V7.1.0,2007.

[29] I. C. Wong, O. Oteri, and W. Mccoy, ‘‘Optimal resource allocation inuplink SC-FDMA systems,’’ IEEE Trans. Wireless Commun., vol. 8, no. 5,pp. 2161–2165, May 2009.

[30] M. Rahman and H. Yanikomeroglu, ‘‘Enhancing cell-edge performance:A downlink dynamic interference avoidance scheme with inter-cell coor-dination,’’ IEEE Trans. Wireless Commun., vol. 9, no. 4, pp. 1414–1425,Apr. 2010.

[31] V. Chandrasekhar and J. G. Andrews, ‘‘Spectrum allocation intiered cellular networks,’’ IEEE Trans. Commun., vol. 57, no. 10,pp. 3059–3068, Oct. 2009.

[32] A. Leith, M. S. Alouini, D. I. Kim, X. Shen, and Z. Wu, ‘‘Flexibleproportional-rate scheduling for OFDMA system,’’ IEEE Trans. MobileComput., vol. 12, no. 10, pp. 1907–1919, Oct. 2013.

[33] LTE; Evolved Universal Terrestrial Radio Access (E-UTRA); PhysicalLayer Procedures, document 3GPP.TS.36.213 11.0.0, 2012.

[34] G. Song and Y. Li, ‘‘Cross-layer optimization for OFDM wirelessnetworks—Part II: Algorithm development,’’ IEEE Trans. WirelessCommun., vol. 4, no. 2, pp. 625–634, Mar. 2005.

[35] G. Song, Y. Li, and L. J. Cimini, ‘‘Joint channel- and queue-aware schedul-ing for multiuser diversity in wireless OFDMA networks,’’ IEEE Trans.Commun., vol. 57, no. 7, pp. 2109–2121, Jul. 2009.

[36] M. Katoozian, K. Navaie, and H. Yanikomeroglu, ‘‘Utility-based adap-tive radio resource allocation in OFDM wireless networks with trafficprioritization,’’ IEEE Trans. Wireless Commun., vol. 8, no. 1, pp. 66–71,Jan. 2009.

[37] F. Capozzi, G. Piro, L. A. Grieco, G. Boggia, and P. Camarda,‘‘Downlink packet scheduling in LTE cellular networks: Key design issuesand a survey,’’ IEEE Commun. Surveys Tuts., vol. 15, no. 2, pp. 678–700,2nd Quart., 2013.

[38] M. Kalil, A. Shami, and A. Al-Dweik, ‘‘QoS-aware power-efficient sched-uler for LTE uplink,’’ IEEE Trans. Mobile Comput., vol. 14, no. 8,pp. 1672–1685, Aug. 2014.

20826 VOLUME 5, 2017


MOHAMAD KALIL received the Ph.D. degree inelectrical and computer engineering from the Uni-versity of Western Ontario, London, ON, Canada,in 2015. He is currently a Staff Software Engineerwith the Analytical Department, IBM, Toronto,ON, Canada. His current research interests are inthe area of machine learning, and data analytics,cross-layer design, radio resource management,and wireless network virtualization.

ARAFAT AL-DWEIK (S’97–M’01–SM’04) receivedthe B.Sc. degree in telecommunication engineer-ing fromYarmouk University, Jordan, in 1994, andtheM.S. (summa cum laude) and the Ph.D. (magnacum laude) degrees in electrical engineering fromCleveland State University, Cleveland, OH, USA,in 1998 and 2001, respectively. From 1999 to2001, he was with Efficient Channel Coding, Inc.,Cleveland, OH, USA, as a Research and Develop-ment Engineer, where he was involved in advanced

modulation, coding, and synchronization techniques. From 2001 to 2003,he was the Head of Department of Information Technology, Arab AmericanUniversity in Palestine. From 2003 to 2012, he was with the CommunicationsEngineering Department, Khalifa University, United Arab Emirates. Hejoined University of Guelph, ON, Canada, as an Associate Professor, from2013 to 2014. He has been a Visiting Research Fellow with the Schoolof Electrical, Electronic and Computer Engineering, Newcastle University,Newcastle upon Tyne, U.K., since 2006. He is also a Research Professor anda Member of the School of Graduate Studies, Western University, London,ON, Canada. He has extensive editorial experience, where he serves asan Associate Editor of the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY

and the IET Communications. He was a TPC Member in several majorconferences such as the IEEE GLOBECOM, ICC, PIMRC, and WCNC. Hehas received several research awards and he was a recipient of the FulbrightScholarship.

MOHAMED F. ABU SHARKH (M’16) receivedthe Ph.D. degree fromWestern University in 2016,with a focus on resource management in next gen-eration cloud computing technologies, introducingefficient techniques to tackle resource allocationchallenges catering to service delay, energy effi-ciency, performance evaluation, and high avail-ability. He is currently a Software EngineeringFaculty with Ferris State University and also aVisiting Scholar with the Electrical and Computer

Engineering Department, Western Michigan University, Kalamazoo, MI,USA. He endeavors to combine the insights gained while working as soft-ware implementation consultant along with researchmodeling, optimization,and performance evaluation techniques to design solutions catering to realworld challenges. This yielded successful results for projects in the area ofcloud computing covering resource allocation, cloud systems performancemeasurement, high availability, energy efficiency, and machine learningsupported by industrial partners such as Samsung and Ericsson. His currentresearch interests span topics including cloud computing resource optimiza-tion, Natural language processing, real time bidding for cloud resources inInternet of Things environments and exploiting machine learning for betterresource utilization.

ABDALLAH SHAMI (SM’09) received the B.E.degree in electrical and computer engineer-ing from Lebanese University, Beirut, Lebanon,in 1997, and the Ph.D. degree in electrical engi-neering from the Graduate School and UniversityCenter, City University of New York, New York,NY, USA, in 2002. Since 2004, he has been withWestern University, Canada, where he is currentlya Professor with the Department of Electrical andComputer Engineering. His current research inter-

ests are in the area of network-based cloud computing and wireless/datanetworking. He is currently an Associate Editor of the IEEE COMMUNICATIONS

SURVEY AND TUTORIALS, the IET Communications Journal and the WileyJournal of Wireless Communications and Mobile Computing. He has chairedkey symposia of the IEEEGLOBECOM, the IEEE ICC, the IEEE ICNC, andICCIT. He is a former Chair of the IEEE Communications Society TechnicalCommittee on Communications Software.

AHMED REFAEY (SM’15) received the B.Sc.and M.Sc. degrees from Alexandria University,Egypt, in 2003 and 2005, respectively, and thePh.D. degree from Laval University, QC, Canada,in 2011, and the P.Eng. degree. He was a SeniorSystems Architect with the Research And Devel-opment Group, Mircom Group of Companies,from 2013 to 2016, a Post-Doctoral Fellow withthe ECE Department, The University of WesternOntario, from 2012 to 2013, and a Professional

Researcher with the LRTS Laboratory, Laval University, in the field of wire-less communications from 2007 to 2011. He was a System/Core NetworkEngineer leading a team of junior engineers and technicians in the telecomfield in the three large companies of Fujitsu, Vodafone, and Alcatel-Lucent.He is currently an Assistant Professor with the Manhattan College and alsoan Adjunct Research Professor with The University of Western Ontario.

VOLUME 5, 2017 20827

Documents

A Framework for Joint Wireless Network Virtualization and ... · implemented or in contention with 5G include densi˝cation ... cognitive radio, beam division multiple access(BDMA),cloud-basedradioaccessnetworks(CRAN)