1

Cooperative Decentralized Resource Allocation inHeterogeneous Wireless Access Medium

Muhammad Ismail, Student Member, IEEE, Atef Abdrabou, Member, IEEE, and Weihua Zhuang, Fellow, IEEE

Abstract—In this paper, radio resource allocation in a hetero-geneous wireless access medium is investigated. Mobile terminals(MTs) are equipped with multiple radio interfaces and areassumed to have multi-homing capabilities. A novel algorithm,namely prediction based resource allocation algorithm, is pro-posed for the resource allocation. Unlike the existing solutionsin literature, the proposed algorithm does not require a centralresource manager to perform the radio resource allocation. TheMT plays an active role in the resource allocation operationby requesting a bandwidth share from each available networkbased on the available resources at the network, such that thetotal allocated bandwidth from different networks satisfies theMT service requirement. The proposed algorithm is suitable forimplementation in a dynamic environment with call arrivals anddepartures, and relies on network cooperation to perform thedecentralized radio resource allocation in an efficient manner.Simulation results are presented to investigate the performancetradeoffs of the proposed algorithm.

Index Terms—Heterogeneous wireless networks, resource allo-cation, multi-homing.

I. INTRODUCTION

Currently, the wireless communication network is a hetero-geneous environment, with overlapped coverage from differentnetworks [1]. These wireless access networks include cellularnetworks, wireless metropolitan area networks (WMANs),wireless local area networks (WLANs), and so on. Suchnetworks have complimentary service capabilities in termsof bandwidth, coverage area, and cost. Hence, in this het-erogeneous wireless access medium, the integration of thesedifferent networks will lead to better service quality to mobileusers and enhanced performance for the networks [2]. A veryimportant component of the integrated architecture is radioresource management mechanisms for bandwidth allocationand call admission control. These mechanisms are essentialin order to satisfy the required bandwidth by mobile termi-nals (MTs) via different available wireless networks and tomake efficient utilization of the available resources from thesenetworks.

In literature, various works have studied the problem ofradio resource allocation in a heterogeneous wireless accessmedium. Two types of radio resource allocation mechanismscan be distinguished in such an environment. The first type,referred to as single-network resource allocation, includes the

M. Ismail and W. Zhuang are with the Department of Electricaland Computer Engineering, University of Waterloo, Waterloo, Canada, e-mail:{m6ismail, wzhuang}@uwaterloo.ca.

A. Abdrabou is with the Department of Electrical Engineering, UAEUniversity, Al-Ain, Abu Dhabi, UAE (e-mail: atef.abdrabou@uaeu.ac.ae.).

This work was supported by a research grant from the Natural Science andEngineering Research Council (NSERC) of Canada.

solutions that utilize a single radio interface of an MT. Hence,each call obtains its required bandwidth from a single accessnetwork at any time instant. The second type, referred to asmulti-homing resource allocation, includes the solutions wheremultiple radio interfaces of an MT are used simultaneously tosatisfy the user’s requirement. In this case, the MT obtains itsrequired bandwidth from all wireless access networks availableat its location.

In this paper, the multi-homing resource allocation problemin a heterogeneous wireless access medium is investigated.A novel resource allocation algorithm, namely the predictionbased resource allocation (PBRA) algorithm, is proposed tosolve this problem. While the existing solutions in literaturerequire a central resource manager to perform the allocation,the newly developed algorithm is decentralized. In such adecentralized architecture, the MT plays an active role in theresource allocation operation and relies on network coopera-tion in order to satisfy its required bandwidth. The MT asksfor a bandwidth share from each available network, basedon the available capacity at each network, such that the totalallocated bandwidth from all the networks satisfies its servicerequirement. As a result, each network base station (BS) oraccess point (AP) can perform its own resource allocationwithout the need for a central resource manager over thedifferent wireless networks. The proposed algorithm is suitablefor implementation in a dynamic environment as it accountsfor the stochastic user mobility and call traffic models in orderto perform the radio resource allocation in an efficient manner.

The rest of this paper is organized as follows: In the nextSection, the related work is reviewed. The system model isdescribed in Section III. For completeness, in Section IV, thedecentralized optimal resource allocation (DORA) algorithm[3] for a static system is reviewed. The proposed coopera-tive resource allocation algorithm for a dynamic system ispresented in Section V. Complexity analysis of the proposedalgorithm is presented in Section VI. Simulation results anddiscussions are given in Section VII. Finally, Section VIIIdraws some conclusion. The important symbols used in thispaper are summarized in Table I.

II. RELATED WORK

The problem of radio resource allocation in a heterogeneouswireless access medium is addressed in [3] - [6]. The existingsolutions can be classified in two categories, namely single-network and multi-homing resource allocation mechanisms.This classification is based on whether a single radio interfaceor multiple radio interfaces of an MT are used simultaneouslyfor the same application.

2

TABLE ISUMMARY OF IMPORTANT SYMBOLS

Symbol DefinitionBm Required bandwidth by MT mBmin

m Minimum required bandwidth of MT mBmax

m Maximum required bandwidth of MT mbnms Allocated bandwidth from network n to MT m through BS/AP sCns Transmission capacity of network n BSs/APs s

C̃lk Maximum number of calls of service class l which can be supported in service area kI Number of iterations required for the DORA algorithm to reach an optimal solutionK Set of service areas in the geographical regionL Set of service classesM Set of MTs in the geographical region

Mns Set of MTs in the coverage area of network n BS/AP s

M⃗j+1lk Vector of predicted number of calls of service class l in service area k during period Tj+1

Mlk Number of existing calls of service class l in service area k

M̂lk Target value of number of calls of service class l in service area k in the CPRAMlk(t

ja) Number of existing calls of service class l in service area k at time instant tja

M̃lk(tja + τ) The predicted number of calls of service class l in service area k at time instant tja + τ using Mlk(t

ja)

M̃lk(Tj+1) The maximum predicted number of calls of service class l in service area k during period Tj+1

N Set of available wireless access networks in the geographical regionNk Set of wireless access networks available in service area kSn Set of BSs/APs of network n in the geographical regionSnk Set of BSs/APs of network n covering service area k

T⃗ jlk Time vector of arrival events for calls of service class l in service area k during period Tj

T lc Time duration of video call that belongs to service class l

Tkr User residence time in service area k

T lkh Channel holding time for a video call of service class l in service area k

tja Time instant for an arrival event during period Tj

υlk Arrival rate of both new and handoff video calls of service class l in service area kλns Link access price of network n BS/AP s

µ(1),(2)m MT m coordination parametersαo Fixed step size, o ∈ {1, 2, 3}ϵlk Upper bound on call blocking probability for service class l in service area kτ Prediction duration

The single-network resource allocation mechanisms arestudied in [4] and [5]. In [4], a utility function based resourceallocation mechanism is developed for a code division multipleaccess (CDMA) based cellular network and a WLAN. In[5], two resource management mechanisms are introducedfor bandwidth allocation and call admission control in aheterogeneous wireless access medium. The single-networkresource allocation mechanisms suffer from a limitation that anincoming call is blocked if no network in its service area canindividually satisfy the required bandwidth of the call. Hence,these mechanisms do not fully exploit the available resourcesfrom different networks.

The multi-homing resource allocation mechanisms are stud-ied in [6] and [7]. In [6], radio bandwidth is allocated todifferent call traffic types based on a utility fairness concept.In [7], the problem of bandwidth allocation is formulatedusing game theory. The mechanisms of [6] and [7] supportmulti-homing MTs. Hence, each call obtains its requiredbandwidth for a specific application from all wireless accessnetworks available at its location. This has the followingadvantages: 1) The available resources from different wirelessaccess networks can be aggregated to support applicationswith high required bandwidth using multiple threads at theapplication layer; 2) The multi-homing concept can reduce thecall blocking rate and improve the overall system capacity.

The existing resource allocation mechanisms that supportMTs with multi-homing capabilities in a heterogeneous wire-

less access medium require a central resource manager toperform the resource allocation and admission control. Acentral resource manager is needed in these cases as theallocated bandwidth from each network BS/AP to a givenconnection should sum up to the bandwidth required by theconnection. As a result, a global view of the BS/AP resourceavailability of every network is required in order to performcoordination among the allocations from different networksto satisfy the total required bandwidth for the connection.This global view is provided by the central resource manager.However, this is not practical in a case that these networksare operated by different service providers. A central resourcemanager which controls the operation of different networks’BSs/APs in such a case raises some issues [3] related to:Firstly, the question of which network will be in charge of theoperation and maintenance of the central manager, consideringthe fact that this network will control the resources of othernetworks; Secondly, the changes required in different networkstructures and operations in order to account for this centralresource manager; Finally, the fact that the central controlleris a single point of failure. Hence, if it breaks down, the wholemulti-homing service fails and this may extend to the operationof different networks. As a result, in such a networkingenvironment, it is desirable to have a decentralized solutionthat enables each network BS/AP to perform its own resourceallocation and admission control while at the same time tocooperate with other available networks’ BSs/APs to support

3

MTs with multi-homing capabilities. In [3], a decentralizedoptimal resource allocation (DORA) algorithm is proposed tosupport MTs with multi-homing capabilities in heterogeneouswireless access medium. Using the algorithm, each networkBS/AP solves its own utility maximization problem to allocateits resources to multi-homing MTs. The MTs coordinate theallocation from different networks in order to satisfy its totalrequired bandwidth. The DORA algorithm is an iterative one,that relies on signaling exchange between an MT and differentnetworks to find the optimal allocation from each network tosatisfy the total required bandwidth of the MT. The algorithmis limited to a static system model, without new arrival anddeparture of calls in different service areas, with the objectiveof identifying the role of each entity in the heterogeneouswireless access medium in such a decentralized architecture.In a dynamic system, with MTs stochastic mobility and calltraffic models, applying the DORA algorithm can be tooexpensive. This is due to the fact that, with every arrivalor departure of a call in any service area, reallocations ofresources from different networks to all the existing calls arerequired in order to reach the optimal resource allocation.As a result, excessive signaling is needed for informationexchange between the existing MTs and the BSs/APs ofdifferent networks. This signaling overhead is a function of thecall arrival and departure rates, the numbers of calls in differentservice areas, and the number of iterations required for thealgorithm to converge to the optimal solution. This approachcan lead to high handoff latency which is not desirable forseamless service provision.

In this work, a decentralized algorithm is proposed forresource allocation in a heterogeneous wireless access mediumfor MTs with multi-homing capabilities. The new algorithmaccounts for the system dynamics, in terms of call arrivalsand departures and their service requests, in order to performan efficient resource allocation. By efficient resource allocationwe mean a resource allocation that can: 1) significantly reducethe required resource reallocations to existing calls and theassociated signaling overhead over the air interface in thedecentralized network architecture, with MTs arrivals to anddepartures from different service areas, and 2) achieve anacceptable call blocking probability and a sufficient amountof allocated resources per call.

III. SYSTEM MODEL

A. Wireless Access Networks

Consider a geographical region with a set, N , of availablewireless access networks using different technologies, N ={1, 2, . . . , N}. Each network, n ∈ N , is operated by a uniqueservice provider. Network n ∈ N has a set, Sn, of BSs/APs inthe geographical region, Sn = {1, 2, . . . , Sn}. The BSs/APs ofeach network have different coverage areas from those of othernetworks. With overlapped coverages from different networksin some areas, the geographical region can be described by aset, K, of service areas, K = {1, 2, . . . ,K}. A unique subsetof BSs/APs from all the networks cover each service area,k ∈ K, as shown in Figure 1. The set of networks available atservice area k is given by Nk, and the set of BSs/APs from

Fig. 1. The network coverage areas

network n covering service area k is given by Snk. Networkswith overlapped coverage are assumed to operate in differentfrequency bands, hence interference issues among differentnetworks are not considered. The downlink transmission ca-pacity of each network, n ∈ N , BS/AP, s ∈ Sn, is Cns Mbps.Each BS/AP broadcasts an identification (ID) beacon which isused in the MT attachment procedure [8]. It is assumed thatdifferent networks are already connected through a backboneto exchange their roaming signaling information. We rely onthe roaming signaling backbone in order to exchange thesignaling information required by our proposed algorithms.

B. Transmission Model

Consider a downlink scenario, where an MT, m, can get itsrequired bandwidth, Bm, on the downlink from all wirelessaccess networks available at its location using its multi-homingcapability. The set of MTs available in the geographical regionis denoted by M. The set of MTs which lie in the coveragearea of the sth BS/AP of the nth network is denoted byMns.The allocated bandwidth in the downlink from network n to anMT m through BS/AP s is denoted by bnms, where n ∈ N ,m ∈ Mns and s ∈ Sn. Let B = [bnms] be a matrix ofallocated bandwidth from network n through BS/AP s to MTm ∈ M, where bnms = 0 if MT m /∈ Mns. Although theproposed algorithm studies radio resource allocation on thedownlink, it can be employed for radio resource allocation onthe uplink.

C. Service Traffic Models

As multi-homing resource allocation is employed to supportapplications with a high required transmission rate, we con-sider video service applications such as on-demand streaming.A video call of MT m is considered to be a variable bitrate (VBR) service that is allocated a bandwidth Bm in therange [Bmin

m , Bmaxm ] [9], where Bmin

m guarantees a minimumquality-of-service (QoS) requirement for the video call. Themore allocated bandwidth to a video call, the higher theperceived video quality experienced on the MT. However,

4

there is a maximum bandwidth Bmaxm that can be allocated to a

video call, which is enforced to incorporate the MTs technicallimitations [9]. With sufficient resources in the service area,a VBR call is allocated its maximum required bandwidthBmax

m . However, when all BSs/APs in the service area reachtheir capacity limitation, the allocated bandwidth for the VBRvideo call is reduced towards the minimum required bandwidthBmin

m in order to accommodate more calls. There exists a set,L = {1, 2, . . . , L}, of service classes. Each service class, l,has unique Bmin

l and Bmaxl values. The allocated bandwidth

to MT m with a VBR video call from service class l is Bl.The number of existing calls of service class l in service areak is denoted by Mlk. It is assumed that there exists sufficientcapacity through the available networks in the geographicalregion to satisfy a target call blocking probability for eachservice class l in each service area k. The maximum numberof calls of each service class l which can be supported in eachservice area k, given the transmission capacities of availableBSs/APs, is denoted by C̃lk. This maximum number of callscan be determined following a capacity analysis similar tothe one in [10]. A call admission control procedure is inplace, which ensures that Mlk ≤ C̃lk, so that feasible resourceallocation solutions exist.

Video call arrivals are modeled as a Poisson process, whichis a widely adopted assumption [10]. In particular, the arrivalprocess of both new and handoff video calls from serviceclass l to service area k is modeled by a Poisson processwith parameter υlk. According to statistics of on-demand videostreaming [11], [12], the video call duration is very likely tobe heavy-tailed. A very important feature of heavy-taildnessis the ’mice-elephants’ phenomenon [13]. With respect tothe video call duration, it implies that most video callshave a quite short duration while a small fraction of videocalls have an extremely large duration. However, performanceanalysis is extremely difficult with heavy-tailed distributions.For effective and tractable analysis, it is proposed in [14]to fit a large class of heavy-tailed distributions with hyper-exponential distributions. For tractability, we use a two-stagehyper-exponential distribution to model the video call duration.For a video call of MT m that belongs to class l, the probabilitydensity function (PDF) of the call duration, T l

c , with mean T̄ lc ,

is given by [10]

fT lc(t) =

alal + 1

· alT̄ lc

· e−alT̄ lct+

1

al + 1· 1

alT̄ lc

· e−1

alT̄lct,

al ≥ 1, t ≥ 0. (1)

In (1), the parameter al can characterize the mice-elephantfeature. A large fraction of users, al

al+1 , have a call duration

with mean time T̄ lc

al, while the other fraction of users, 1

al+1 ,have a call duration with mean time alT̄ l

c .

D. Mobility Models and Channel Holding Time

User residence time is used to characterize the user mobilitywithin a given service area k ∈ K, which is assumed to followan exponential distribution. For service area k ∈ K, the PDF

of the user residence time T kr , with mean T̄ k

r , is given by

fTkr(t) =

1

T̄ kr

e− t

T̄kr , t ≥ 0. (2)

The channel holding time in a given service area is givenby T lk

h = min(T lc , T

kr ), where T l

c and T kr are independent of

each other. Hence,

Pr{min(T lc , T

kr ) > t} = Pr{T l

c > t, T kr > t}

= Pr{T lc > t} · Pr{T k

r > t}.(3)

As a result, the PDF of the channel holding time is given by

fT lkh(t) = fT l

c(t)[1− FTk

r(t)] + fTk

r(t)[1− FT l

c(t)], t ≥ 0

(4)where FT l

c(t) and FTk

r(t) are the cumulative distribution

functions (CDFs) for the call duration and user residence timerespectively. Using (1) and (2), we have

fT lkh(t) =

alal + 1

· ( 1

T̄ kr

+alT̄ lc

) · e−( 1

T̄kr+

alT̄ lc)t

+1

al + 1· ( 1

T̄ kr

+1

alT̄ lc

) · e−( 1

T̄kr+ 1

alT̄lc)t,

t ≥ 0. (5)

IV. DECENTRALIZED OPTIMAL RESOURCE ALLOCATION(DORA)

The resource allocation problem for MTs with multi-homingcapabilities in a heterogeneous wireless access medium isexpressed by the following convex optimization problem [3]

maxB

N∑n=1

Sn∑s=1

∑m∈Mns

ln(1 + ηbnms)

s.t.∑

m∈Mns

bnms ≤ Cns, ∀n ∈ N , s ∈ Sn

Bminm ≤

N∑n=1

Sn∑s=1

bnms ≤ Bmaxm , ∀m ∈M

(6)

where η is used for scalability of bnms [15] and [Bminm , Bmax

m ]is defined for service l of MT m among the L availableservice classes. We refer to problem (6) as optimal resourceallocation problem (ORAP). The resource allocation objec-tive of the ORAP is to find the optimal allocation bnms,∀n ∈ N ,m ∈ Mns and s ∈ Sn, that maximizes thetotal utility in the region. The first constraint in (6) satisfiesthe BS/AP capacity limitation, while the second constraintsatisfies the call required bandwidth of MT m from all wirelessaccess networks available at its location. The heterogenity inthe network settings of problem (6) appears in two aspects.The first aspect is the heterogenity in networks’ capacities,which is introduced by the term Cns. The second aspectis the heterogenity in coverage areas, which is introducedthrough the term Mns. The transmission technologies fordifferent networks are handled through the MT’s differentradio interfaces.

A decentralized optimal resource allocation (DORA) al-gorithm for the ORAP is proposed in [3] using full dual

5

decomposition [16], [17]. The DORA is an iterative algorithm.It performs an optimal bandwidth allocation to a static set Mof calls at the BSs/APs of all networks based on the updateof three parameters, namely link access price λns of networkn BS/AP s and coordination parameters µ(1)

m and µ(2)m of MT

m, over a number of iterations, until an optimal solution isreached. Each network BS/AP starts with an initial feasiblevalue for its link access price λns. Similarly, each MT startswith an initial feasible value for its coordination parametersµ(1)m and µ(2)

m . The MTs broadcast their coordination parame-ters to all BSs/APs available at their locations. The BSs/APsperform their bandwidth allocation to the MTs based on theirlink access price values and the coordination parameters fromthe MTs. Each BS/AP updates its link access price value basedon its capacity limitation and the total traffic load carried inits coverage area. Also, each MT updates its coordinationparameters based on the allocated bandwidth from differentBSs/APs and its required bandwidth. The MTs broadcasttheir updated coordination parameters to the BSs/APs and theprocess continues until the algorithm converges to the optimallink access price value λ∗ns, ∀n ∈ N , s ∈ Sn, coordinationparameters µ(1)∗

m and µ(2)∗m , ∀m ∈ M, and hence bandwidth

allocation matrix B∗. The DORA is described in AlgorithmIV.1, where i is an iteration index, αo with o ∈ {1, 2, 3} is afixed step size, [·]+ is a projection on the positive orthant, ψis a small tolerance, and I is the number of iterations to reachthe optimal solution. For more details and discussions on theDORA, we refer the reader to the work of [3].

Algorithm IV.1 DORA [3]

Input: Cns ∀n ∈ N , ∀s ∈ Sn, Bm ∀m ∈M;Initialization: i←− 1; λns(1) ≥ 0; µ(1)

m (1) ≥ 0; µ(2)m (1) ≥

0, bnms(0) = {}, j = 0;while j = 0 do

Bandwith Allocationfor n ∈ N do

for m ∈M dofor s ∈ Sn do

if m ∈Mns thenbnms(i) = [( η

λns(i)+(µ(1)m (i)−µ

(2)m (i))

− 1)/η]+;end if

end forend for

end forif |bnms(i)− bnms(i− 1)| > ψ then

Update of Link Access Pricefor n ∈ N do

for s ∈ Sn doλns(i + 1) = [λns(i) − α1(Cns −∑

m∈Mnsbnms(i))]

+;end for

end forUpdate of Coordination Parametersfor m ∈M doµ(1)m (i + 1) = [µ

(1)m (i) − α2(B

maxm −∑N

n=1

∑Sn

s=1 bnms(i))]+;

µ(2)m (i+ 1) = [µ

(2)m (i)− α3(

∑Nn=1

∑Sn

s=1 bnms(i)−Bmin

m )]+;end fori←− i+ 1

elsej = 1;

end ifend whileI = i;Output: B∗, I .

The algorithm is originally proposed for a static systemmodel, without arrivals of new calls or departures of existingones. The study in [3] identifies the role of each entity inthe heterogeneous wireless access medium in a decentralizedarchitecture.

V. DECENTRALIZED RESOURCE ALLOCATION IN ADYNAMIC ENVIRONMENT

In the DORA algorithm, when the optimal call traffic load(∑

m∈Mnsb∗nms) at network n BS/AP s is less than its

capacity limitation (Cns), its optimal link access price valueλ∗ns = 0 and the calls under its jurisdiction are allocated theirmaximum required bandwidths. As the optimal call traffic loadreaches the capacity limitation, λ∗ns > 0 and the allocatedbandwidth to each of the calls in service is reduced towardsthe minimum required bandwidth in order to support new calls.

In a dynamic system, the call traffic load at each BS/APfluctuates over time with call arrivals to and departures fromits coverage area. This results in a fluctuating (time-varying)optimal value for the link access price λ∗ns and bandwidth al-location matrix B∗ with every call arrival or departure. Hence,bandwidth reallocations to the existing calls are triggered.In order to reach the optimal bandwidth allocation in sucha decentralized architecture, information exchanges betweenMTs and BSs/APs for coordination parameter updates arerequired for the I iterations. This should take place with everycall arrival to or departure from any service area k. Hence,applying the DORA in a dynamic system incurs high signalingoverhead. In addition, it is possible that an arrival or departureevent occurs during the DORA I iterations, thus it may notconverge to an optimal solution. As a result, the DORA is notpractical to implement in a dynamic scenario. A complexityanalysis for the DORA implementation in a dynamic environ-ment is presented in Section VI. In this section, we discusshow to address the aforementioned implementation challengesand propose a sub-optimal decentralized algorithm for efficientresource allocation in a dynamic system.

A. Constant Price Resource Allocation (CPRA)

In order to perform an efficient decentralized radio resourceallocation in a dynamic network environment, one strategyis to avoid solving the ORAP (6) for every call arrival toor departure from any service area k. Meanwhile, our mainobjective is to satisfy the required resource allocation percall for a certain call blocking probability. These objectivesare achieved by employing fixed link access price values forresource allocation at different networks BSs/APs independent

6

of call arrivals and departures. With time-invariant BS/AP linkaccess price values, the corresponding resource allocation isreferred to as constant price resource allocation (CPRA). TheCPRA has two phases, namely setup phase and operationphase. The setup phase is executed only once at the initialoperation time of the networks, while the operation phase isexecuted every time a new MT joins the networks.

The setup phase: This phase is to determine the fixed BS/APlink access price values based on steady-state statistics ofcall traffic and user mobility in order to achieve satisfactoryperformance in terms of allocated resources per call and callblocking probability in the operation phase. Consider thegeographical region of Figure 1. In the setup phase, set thenumber of calls of each service class l in each service area k,Mlk, to a target value M̂lk. The corresponding optimal linkaccess price value for each BS/AP in the geographical regioncan be determined by solving (6) using the DORA algorithm.If we employ these BS/AP link access price values for resourceallocation in the operation phase, the radio resources of allnetworks will be distributed exactly over M̂lk calls, ∀l ∈ L,k ∈ K. Hence, in the operation phase, when Mlk = M̂lk,any incoming call of class l to service area k will be blocked.That is, the choice of the target value M̂lk ∀l ∈ L and k ∈ Kand in turn the corresponding BS/AP link access price λnsvalue ∀n ∈ N and s ∈ Sn in the setup phase determine theoverall performance of the geographical region in terms of theallocated resources per call and the call blocking probability.As a result, the value of M̂lk should be properly chosen toachieve satisfactory performance in the resource allocation.In the dynamic system, Mlk is a random variable. Given theprobability distribution of Mlk, alternatively we can representM̂lk by a design parameter ϵlk such that

Pr(Mlk > M̂lk) ≤ ϵlk, ∀l ∈ L, k ∈ K (7)

where ϵlk ∈ [0, 1]. It is clear that the value of M̂lk dependson ϵlk and the distribution of Mlk. Indeed, from (7), ϵlk isthe upper bound of the call blocking probability of serviceclass l in service area k given that M̂lk ≤ C̃lk. Otherwise,M̂lk = C̃lk, and both the optimal solution of ORAP andthe CPRA achieve the same call blocking performance. Wecan choose the M̂lk value based on the requirement on callblocking probability.

Since call arrivals of service class l to service area k followa Poisson process, the channel holding time follows a generaldistribution, and all calls are served simultaneously withoutqueueing, an M/G/∞ model [18] can be used to determineM̂lk ∀l ∈ L, k ∈ K in the setup phase, using the steady-statecall traffic and user mobility statistics. The number of callsof service class l that are simultaneously present in servicearea k, Mlk, follows the Poisson distribution with mean rlk =υlk ·E[T lk

h ] [18], where E[T lkh ] is the average channel holding

time of service class l in service area k and can be calculatedfrom (5) as

E[T lkh ] =

alal + 1

· 11T̄kr+ al

T̄ lc

+1

al + 1· 1

1T̄kr+ 1

alT̄ lc

,

∀l ∈ L, k ∈ K. (8)

Using (7), M̂lk is the minimum integer which satisfies thefollowing relation [18]

M̂lk∑i=0

rilke−rlk

i!≥ (1− ϵ), ∀l ∈ L, k ∈ K. (9)

For a given ϵlk, using M̂lk ∀l ∈ L, k ∈ K, the ORAP issolved for the corresponding optimal link access price valuesλ̂ns ∀n ∈ N , s ∈ Sn.

The operation phase: In this phase, the bandwidth allocationprocess is performed for each user joining the networks basedon the following four steps.

Step 1: The link access price value for each network BS/APin the geographical region is fixed to the value calculated in thesetup phase, λ̂ns, independent of call arrivals and departures.This fixed link access price value is broadcasted by eachBS/AP via its ID beacon.

Step 2: An incoming MT uses its multiple radio interfaces tolisten to the link access price values of the BSs/APs availableat its location.

Step 3: The MT then uses the link access price values tosolve for the bandwidth share from each BS/AP such thatthe total amount of resources allocated from all the BSs/APssatisfies its required bandwidth. This can be done at MT, m,of service class l in service area k, by using Algorithm V.2(which is based on the DORA algorithm [3]).

Algorithm V.2 Calculation of bandwith share from each avail-able network BS/AP at MT m

Input: λ̂ns ∀n ∈ Nk, s ∈ Snk, Bm;Initialization: µ(1)

m (1) ≥ 0; µ(2)m (1) ≥ 0;

for i = 1 : I dofor n ∈ Nk do

for s ∈ Snk dobnms(i) = [( η

λ̂ns+(µ(1)m (i)−µ

(2)m (i))

− 1)/η]+;end for

end forµ(1)m (i + 1) = [µ

(1)m (i) − α2(B

maxm −∑N

n=1

∑Sn

s=1 bnms(i))]+;

µ(2)m (i + 1) = [µ

(2)m (i) − α3(

∑Nn=1

∑Sn

s=1 bnms(i) −Bmin

m )]+;end forOutput: The required bnms ∀n ∈ Nk, s ∈ Snk.

Step 4: MT m asks for the bandwidth share bnms fromBS/AP s of network n, ∀n ∈ Nk, s ∈ Snk, which performs theallocation if it has sufficient resources. The MT call is blockedif the total allocated resources do not satisfy its requiredbandwidth.

As the BS/AP link access price values are independent ofcall arrivals to and departures from different service areas, noresource reallocations to existing calls are required. Further-more, the I iterations required to reach the desired resourceallocations from all the BSs/APs to satisfy the total requiredbandwidth is solved locally at each MT without informationexchange between the MT and the BSs/APs for every iterationas in DORA. Hence, the CPRA approach requires almost nosignaling overhead in order to reach the required bandwidth

7

from each BS/AP1. The convergence of the CPRA algorithmfollows the convergence of the DORA algorithm which isgiven in [3]. However, the CPRA algorithm provides a sub-optimal solution to the ORAP of (6) as the link access pricevalue is not updated with every call arrival and departure.

In summary, in the CPRA approach, a small value of ϵlkresults in a low call blocking probability in the operationphase. However, this corresponds to a large M̂lk value andhence large BS/AP link access price values, which leads toa low amount of allocated resources per call in the operationphase. On the other hand, a large value of ϵlk means a high callblocking probability and a large amount of allocated resourcesper call in the operation phase. Hence, ϵlk should be chosen tobalance the tradeoff existing between the allocated resourcesper call and the call blocking probability.

With an appropriate choice of ϵlk, the CPRA, with its setupand operation phases, can allocate resources for a target callblocking probability in the decentralized network architecturewith dynamic call arrivals and departures.

B. Prediction Based Resource Allocation (PBRA)

The CPRA performs the resource allocation based on M̂lk

calculated according to the steady-state (long-term) call trafficand user mobility statistics. In a dynamic environment, Mlk

can deviate from M̂lk for some time. However, the resourceallocation in the operation phase does not adapt to short-termdynamics in the call traffic load. A call can be allocatedonly its minimum required bandwidth even if there existsufficient resources in the BSs/APs at its current locationthat can be used to provide better service quality. In CPRA,these extra resources which are not utilized (at a low trafficload) are actually reserved for possible incoming calls, so thatthe target call blocking probability can be achieved. For abetter service quality compromise between the existing calls(in terms of the amount of allocated resources to each call)and the potential incoming calls (in terms of the call blockingprobability), resource allocation adaptive to a short-term calltraffic load (via resource re-allocation to the calls in service)can help. To do so, in the following, we propose to update M̂lk

∀l ∈ L, k ∈ K in the operation phase periodically with periodτ , and hence update the corresponding BS/AP link accessprice values, based on the instantaneous Mlk value, denotedby Mlk(t) at time t. We refer to the corresponding resourceallocation as prediction based resource allocation (PBRA).

Let the time be partitioned into a set of periods T ={T1, T2, . . . , Tj , . . .} of constant duration τ . The beginning ofeach period Tj is denoted by tj . Let T⃗ j

lk denote a time vectorof arrival events for calls of service class l in service area kduring period Tj . The PBRA algorithm can be carried out inthe following six steps.

Step 1: With a new call arrival at tja ∈ T⃗ jlk, a =

{1, 2, . . . ,∣∣∣T⃗ j

lk

∣∣∣}, in period Tj , the number of calls of serviceclass l in service area k at the time instant, Mlk(t

ja), is used

by the BSs/APs in this service area to probabilistically predict

1This is apart from the overhead required to broadcast the fixed link accessprice value λ̂ns by every BS/AP on its ID beacon. However, the contributionof broadcasting this value to the overhead is negligible.

the number of calls at time instant tja + τ in the next timeperiod Tj+1. Hence, we refer to τ as the prediction duration.The predicted number is denoted by M̃lk(t

ja + τ) and should

satisfy

Pr(Mlk(tja + τ) > M̃lk(t

ja + τ)|Mlk(t

ja)) ≤ ϵlk,

∀l ∈ L, k ∈ K. (10)

In order to determine M̃lk(tja + τ), we calculate the condi-

tional probability mass function (PMF) of Mlk(tja + τ) given

Mlk(tja), PMlk(t

ja+τ)|Mlk(t

ja)(i), which can be found using

the transient distribution of the M/G/∞ model [19]. First,we make the following definitions under the assumption ofstationary call arrival and departure processes:

• plkτ - The probability that a call of service class l whichis in service area k at time tja is still present in the sameservice area at time tja + τ ;

• qlkτ - The probability that a call from service class l thatarrives in service area k during (tja, t

ja+τ ] is still present

at the same service area at time tja + τ ;• XB(κ, α) - A binomial random variable with parametersκ and α;

• XP (α) - A Poisson random variable with mean α.

Given Mlk(tja), we have [19]

Mlk(tja + τ) =d XB(Mlk(t

ja), p

lkτ ) +XP (υlkτq

lkτ ) (11)

where =d denotes equality in distribution. The probabilitiesplkτ and qlkτ are given by [19]

plkτ =1

E[T lkh ]

∫ ∞

τ

Pr(T lkh > s)ds

=1

E[T lkh ]

∫ ∞

τ

(1− FT lkh(s))ds. (12)

qlkτ =

∫ τ

0

1

τPr(T lk

h > s)ds

=

∫ τ

0

1

τ(1− FT lk

h(s))ds

=E[T lk

h ]

τ(1− plkτ ) (13)

where FT lkh(s) =

∫ s

0fT lk

h(t)dt is the CDF of T lk

h . Using (11)- (13), PMlk(t

ja+τ)|Mlk(t

ja)(i) can be found, from which we

can calculate M̃lk(tja + τ) using (10) as the minimum integer

satisfying

M̃lk(tja+τ)∑

i=0

PMlk(tja+τ)|Mlk(t

ja)(i) ≥ (1−ϵlk), ∀l ∈ L, k ∈ K.

(14)Step 2: The predicted values of M̃lk(t

ja + τ), ∀l ∈ L, k ∈

K for a = {1, 2, . . . ,∣∣∣T⃗ j

lk

∣∣∣}, are recorded at each BS/AP in

service area k in a vector M⃗j+1lk .

Step 3: At tj+1, the maximum predicted number of callsof each service class l in each service area k during Tj+1,M̃lk(Tj+1), can be found from M⃗j+1

lk . That is, M̃lk(Tj+1) =

8

Fig. 2. Illustration of PBRA events

max(M⃗j+1lk ) if it is less than or equal C̃lk, otherwise

M̃lk(Tj+1) = C̃lk. This guarantees that for M̃lk(Tj+1) ≤ C̃lk

Pr(Mlk(tj+1a ) > M̃lk(Tj+1)) ≤ ϵlk,

∀l ∈ L, k ∈ K, a ∈ {1, 2, . . . ,∣∣∣T⃗ j+1

lk

∣∣∣}. (15)

Step 4: The cooperating BSs/APs in the geographical regionexchange their information of M̃lk(Tj+1) ∀l ∈ L, k ∈ K.The ORAP can be solved at each BS/AP to update its linkaccess price value which is fixed over Tj+1, independent ofcall arrivals to and depatures from different service areas, andis broadcasted on the BS/AP ID beacon.

The call arrival times, the actual and predicted numbers ofcalls of service class l in service area k associated with thesteps 1-4 are illustrated in Figure 2.

Step 5: Each MT in the geographical region during Tj+1,including both incoming and already existing ones, uses thebroadcasted BS/AP link access price values received at its lo-cation during this period to determine and ask for a bandwidthshare from each available BS/AP, following steps 2-4 in theCPRA2.

Step 6: Each MT reports to the BSs/APs available at itslocation its service class and a list of the BS/AP IDs thatthe MT can receive. This information is used by BSs/APs forthe prediction of M̃lk(Tj+2), ∀l ∈ L, k ∈ K, during the nextperiod Tj+2 in order to update their link access price valuesat time tj+2.

While the CPRA uses the target M̂lk value from the setupphase based on steady-state statistics to perform the resourceallocation in the operation phase, the PBRA updates the targetvalue by M̃lk(Tj) every period Tj , j = {1, 2, . . .}, using thecurrent number of calls in service. With this extra information,the PBRA can make a better prediction of the call traffic loadin a short-term, and hence an improved resource allocation isexpected over the CPRA. The PBRA algorithm provides animproved sub-optimal solution to the ORAP as compared tothe CPRA algorithm. The convergence of the PBRA algorithmto this sub-optimal solution follows the convergence of theDORA algorithm which is given in [3]. Since the BS/AP link

2In Algorithm V.2, λ̂ns is replaced by the updated link access price valuethat the MT receives during Tj+1.

access price values during period Tj are based on M̃lk(Tj),the BSs/APs allocate their available resources exactly amongM̃lk(Tj) calls during period Tj . Hence, from (10) and (15),and using the same argument of CPRA, ϵlk serves as an upperbound on the call blocking probability for M̃lk(Tj) ≤ C̃lk.

VI. COMPLEXITY ANALYSIS

In this section, we present a complexity analysis for theDORA implementation in a dynamic system, the CPRA andPBRA algorithms.A. Signaling Overhead

As introduced in Section IV, in order to implement theDORA algorithm in a dynamic system, information signalingneeds to be exchanged between all existing MTs and networkBSs/APs with every call arrival to and departure from anyservice area in order to reach the optimal resource allocation.This signaling overhead is a function of the arrival and depar-ture rate of the calls, the number of existing calls in differentservice areas, and the number of iterations I required for theDORA algorithm to converge to the optimal allocation. Let theaverage number of call arrivals and departures over a period beχa and χd respectively. Then, the signaling overhead on the airinterface for the DORA implementation in a dynamic systemscales as O(χa + χd) over the period. Hence, for high callarrival/departure rates, a high signaling overhead is expected.On the other hand, for the CPRA and PBRA, the BS/APlink access price values are independent of call arrivals anddepartures. Hence, their signaling overhead on the air interfacein order to reach the required resource allocation scales asO(1). As a result, the CPRA and the PBRA signaling overheadscales well with call arrival and departure rates, as comparedwith the DORA implementation in a dynamic system.B. Processing Time

For the DORA, MTs and BSs/APs exchange signaling infor-mation for I iterations to reach an optimal allocation. Assumethat the signaling exchange for the I iterations requires atotal amount of time, σ, for completion. It is expected that σincreases with the call arrival rates in the case that a networkwith contention based medium access control protocol existsamong the available wireless networks, since more MTs willbe involved in the signaling procedure. The signaling exchangefor I iterations should take place with every call arrival ordeparture. Hence, the time duration between two successiveexecution of the I-iteration signaling exchange is given asδ = min(call inter-arrival time, call departure time). Since thecall arrival is a Poisson process with parameter υlk, the callinter-arrival time follows an exponential distribution with PDFfT lk

a(t). The call departure time is given by the channel

holding time, which follows a hyper-exponential distributionwith PDF fT lk

h(t). Using the same analysis as given in (3) -

(4), the PDF of δ, fδ(t), is given by

fδ(t) =al

al + 1· ( 1

T̄ kr

+alT̄ lc

+ υlk) · e−( 1

T̄kr+

alT̄ lc+υlk)t

+1

al + 1· ( 1

T̄ kr

+1

alT̄ lc

+ υlk) · e−( 1

T̄kr+ 1

alT̄lc+υlk)t

,

t ≥ 0. (16)

9

From (16), the average of δ is given by

δ =al

al + 1· 1

1T̄kr+ al

T̄ lc+ υlk

+1

al + 1· 1

1T̄kr+ 1

alT̄ lc+ υlk

. (17)

It is apparent from (17) that the DORA processing time doesnot scale with arrival and departure rates, as δ is inverselyproportional to them. As δ decreases with increasing arrivaland/or departure rates while σ increases with increasing arrivalrates, δ can be smaller than σ. Thus, the DORA algorithm doesnot converge to an optimal allocation whenever δ is smallerthan σ. On the other hand, for the CPRA and the PBRA, theI iterations are solved locally at the MTs and no signalinginformation is exchanged for each iteration. Hence, both theCPRA and the PBRA reach the required bandwidth allocationindependent of the arrival/departure rates.

The CPRA and the PBRA require that the BS/AP linkaccess price values to be broadcasted on the BS/AP ID beacon.Moreover, the PBRA requires an exchange of the predicted calltraffic load among BSs/APs with overlapped coverage every τ .However, this signaling exchange does not take place on theair interface as in the DORA, but is executed over the signalingbackbone connecting different networks. In order to reduce thesignaling information required to exchange the predicted calltraffic load among different BSs/APs over the backbone, theprediction duration τ can be made greater than δ.

VII. SIMULATION RESULTS AND DISCUSSION

This section presents simulation results for the resourceallocation in heterogeneous wireless access medium for MTswith multi-homing capabilities, using the PBRA algorithm incomparison with the ORAP solution and the CPRA. Considera geographical region that is entirely covered by an IEEE802.16e WMAN BS and partially covered by a 4G cellularnetwork BS and an IEEE 802.11b WLAN AP [3]. Hence,N = {1, 2, 3}, with the WMAN, cellular network, and WLANindexed as 1, 2 and 3 respectively. As a result, three serviceareas can be distinguished. One service area is covered by allthree networks, another is covered by both the WMAN andcellular network BSs, and the last one is covered only by thecellular network BS. We consider a single VBR service class(l = 1) and study the performance of the PBRA algorithmin the service area (k = 1) that is covered by all threenetworks, in terms of the allocated resources per call andthe call blocking probability. For simplicity, in the following,we drop the l and k notations. The allocated capacity fromnetwork n BS/AP to the service area under consideration isgiven by C11 = 4 Mbps, C21 = 0.656 Mbps, C31 = 2 Mbps.The Cns values can support a total of 26 VBR calls withrequired bandwidth allocation Bm ∈ [0.256, 0.512] Mbps forMTs with multi-homing capabilities, that is C̃ = 26. Thearrival process of new and handoff video calls is modeled by aPoisson process with parameter υ (call/minute). The video callduration is modeled by a hyper-exponential distribution withthe PDF given in (1) and a1 = 1. The average call duration isT̄c = 20 minutes. The user residence time in the service areaunder consideration follows an exponential distribution withthe PDF given in (2) and an average time T̄r = 15 minutes[10]. The parameter η is set to 1 [15].

0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.50.25

0.3

0.35

0.4

0.45

0.5

0.55

0.6

Call Arrival Rate υ (Call/minute)

Res

ourc

e A

lloca

tion

per

Cal

l (M

bps)

ORAPPBRA, τ = 0.25 minutePBRA, τ = 0.5 minutePBRA, τ = 1 minuteCPRA

(a)

0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.510

−4

10−3

10−2

10−1

Call Arrival Rate υ (Call/minute)

Cal

l Blo

ckin

g P

roba

bilit

y

ORAP

PBRA, τ = 0.25 minute

PBRA, τ = 0.5 minute

PBRA, τ = 1 minute

CPRA

(b)

Fig. 3. Performance comparison: (a) Resource allocation per call; (b) Callblocking probability.

A. Performance Comparison

In the following, the performance of the PBRA algorithmis compared to the optimal solution of problem (6) (ORAP)in terms of resource allocation per call and the call blockingprobability. The optimal solution of ORAP represents a cen-tralized resource allocation. Although it is not appropriate forpractical implementation when different networks are operatedby different service providers, we use the solution of ORAP toserve as an upper bound for the system performance in termsof the allocated resources per call and a lower bound for thesystem performance in terms of the call blocking probability.Also, an CPRA is considered, where no update of the linkaccess price values takes place.

Figure 3 shows performance comparison among the CPRA,PBRA and ORAP solutions versus the call arrival rate υ,with ϵ = 1% and τ = 0.25, 0.5 and 1 minute. At a lowarrival rate, the predicted number of simultaneously presentcalls is low, hence the estimated link access price value is lowand the allocated resource amounts per call using the PBRAalgorithm for the different τ values are high. At a high arrivalrate, the predicted number of simultaneously present calls in

10

0.05 0.1 0.15 0.20.36

0.37

0.38

0.39

0.4

0.41

0.42

Upper bound on Call Blocking Probability ε

Res

ourc

e A

lloca

tion

per

Cal

l (M

bps)

(a)

0.05 0.1 0.15 0.210

−2

10−1

Upper bound on Call Blocking Probability ε

Cal

l Blo

ckin

g P

roba

bilit

y

(b)

Fig. 4. The PBRA algorithm performance versus ϵ: (a) Resource allocationper call; (b) Call blocking probability.

the system is high. For a larger values of τ , less resourcesare allocated per call as explained in the next sub-section.The CPRA provides a lower bound of the performance interms of resource allocation, as it does not update the BS/APlink access price values. For the ORAP solution, there isno call blocking probability for a call arrival rate υ < 1.5call/minute. All the algorithms achieve the desired upperbound for call blocking probability, ϵ, for υ ≤ 1.9 call/minute.For υ > 1.9 call/minute, the predicted number of callssimultaneously present in the system is larger than C̃. Hence,according to the CPRA and the PBRA, the predicted numberis made equal to C̃, and the algorithms achieve the samecall blocking probability as the ORAP solution. Overall, thePBRA performance lies between CPRA and ORAP solutionperformance, as expected. By properly chosing the τ value, adesired compromise between performance and implementationcomplexity can be achieved by the PBRA algorithm.

B. Performance of The PBRA Algorithm

In the following, we study the performance of the PBRAalgorithm versus its two parameters, namely the upper bound

on the call blocking probability ϵ and the prediction durationτ .

Figure 4 shows the performance of the PBRA algorithm interms of the amount of allocated resources per call and callblocking probability versus ϵ, with the call arrival rate υ = 1.7call/minute and the prediction duration τ = 1 minute. As ϵincreases, the PBRA accounts for the simultaneous presenceof less calls in the next τ in its calculation of the link accessprice value. As a result, the call blocking probability increaseswith ϵ. From Figure 4b, the call blocking probability does notexceed its upper bound ϵ. However, the allocated resourcesper call is improved with ϵ, as less resources are reserved forincoming calls which will more likely be blocked. Hence, atradeoff exists between these two performance metrics.

Figure 5 shows the performance of the PBRA in terms ofthe amount of allocated resources per call and call blockingprobability versus the prediction duration τ , with the callarrival rate υ = 1.7 call/minute and ϵ = 1%. With a largerprediction duration, the PBRA updates the BS/AP link accessprice less frequently and a larger number of simultaneouslypresent calls is predicted. Hence, the resource allocation per

1 2 3 4 5

0.29

0.3

0.31

0.32

0.33

0.34

Prediction Duration τ (minute)

Res

ourc

e A

lloca

tion

per

Cal

l (M

bps)

(a)

1 2 3 4 510

−3

10−2

Prediction Duration τ (minute)

Cal

l Blo

ckin

g P

roba

bilit

y

(b)

Fig. 5. The PBRA algorithm performance versus τ : (a) Resource allocationper call; (b) Call blocking probability.

11

call is reduced. Again, the call blocking probability does notexceed its upper bound ϵ with the τ values.

VIII. CONCLUSION

In this paper, a decentralized resource allocation algorithmis proposed for a heterogeneous wireless access medium tosupport MTs with multi-homing capabilities. The proposedPBRA algorithm aims to perform an efficient resource allo-cation in a dynamic system, in order to reduce the signalingoverhead required over the air interface for resource allocationin a decentralized architecture and achieve an acceptable callblocking probability and a sufficient allocated resources percall. The PBRA algorithm relies on short-term call trafficprediction and network cooperation to achieve the objectives.The two parameters ϵlk and τ can be properly chosen to strikea balance between the desired performance in terms of theallocated resources per call and the call blocking probability,and between the performance and the implementation com-plexity. Each MT plays an active role in the resource allocationoperation by requesting a bandwidth share from each availablenetwork based on the available resources at the network, suchthat the total allocated bandwidth from different networkssatisfies the MT service requirement.

REFERENCES

[1] M. Kassar, B. Kervella, and G. Pujolle, “An overview of vertical handoverstrategies in heterogeneous wireless networks,” Computer Communica-tions, vol. 31, no. 10, pp. 2607-2620, June 2008.

[2] M. Ismail and W. Zhuang, “Network cooperation for energy saving ingreen radio communications,” IEEE Wireless Communications, vol. 18,no. 5, pp. 76-81, Oct. 2011.

[3] M. Ismail and W. Zhuang, “A Distributed multi-service resource alloca-tion algorithm in heterogeneous wireless access medium,” IEEE J. onSelect. Areas Commun.,vol. 30, no. 2, pp. 425-432, Feb. 2012.

[4] X. Pei, T. Jiang, D. Qu, G. Zhu, and J. Liu, “Radio resource managementand access control mechanism based on a novel economic model inheterogeneous wireless networks,” IEEE Trans. Vehicular Technology,vol. 59, no. 6, pp. 3047-3056, July 2010.

[5] W. Shen, and Q. Zeng, “Resource management schemes for multipletraffic in integrated heterogeneous wireless and mobile networks,” Proc.17th Int. conf. ICCCN, pp. 105-110, August 2008.

[6] C. Luo, H. Ji, and Y. Li, “Utility based multi-service bandwidth allocationin the 4G heterogeneous wireless access networks,” Proc. IEEE WCNC,April 2009.

[7] C. Truong, T. Geithner, F. Sivrikaya and S. Albayrak, “Network levelcooperation for resource allocation in future wireless networks,” Proc.1st IFIP Wireless Days, 2008.

[8] S. J. Gerasenko, A. Rayaprolu, S. Ponnavaikko and D. K. Agrawal,“Beacon signals: what, why, how, and where,” Computer, vol. 34, pp.108-110, 2001.

[9] R. Litjens, H. van den Berg, and R. J. Boucherie, “Throughputsin processor sharing models for integrated stream and elastic traffic,”Performance Evaluation, vol. 65, no. 2, pp. 152-180, Feb. 2008.

[10] W. Song, Y. Cheng, and W. Zhuang, “Improving voice and data servicesin cellular/WLAN integrated network by admission control,” IEEE Trans.Wireless Communications, vol. 6, no. 11, pp. 4025-4037, Nov. 2007.

[11] M. Li, M. Claypool, R. Kinicki, and J. Nichols, “Characteristics ofstreaming media stored on the Web,” IEEE/ACM Trans. Networking, vol.5, no. 4, pp. 601626, Nov. 2005.

[12] W. Song and W. Zhuang, “Resource allocation for conversational,streaming, and interactive services in cellular/WLAN interworking,” Proc.IEEE Globecom’07, pp. 4785-4789, Nov.-Dec. 2007.

[13] N. Benameur, S. B. Fredj F. Delcoigne, S. Oueslati-Boulahia, andJ. W. Roberts,, “Integrated admission control for streaming and elastictraffic,” Proc. 2nd Intl. Workshop on Quality of Future Internet Services,pp. 69-81, Sept. 2001.

[14] A. Feldmann and W. Whitt, “Fitting mixtures of exponentials to long-tail distributions to analyze network performance models,” PerformanceEvaluation, vol. 31, no. 3-4, pp. 245-279, Jan. 1998.

[15] H. Shen and T. Basar, “Differentiated Internet pricing using a hierarchi-cal network game model,” Proc. 2004 American Control Conference, pp.2322-2327 vol.3, 2004.

[16] D. P. Bertsekas, Non-linear programming, Athena Scientific, 2003.[17] L. S. Lasdon, Optimization theory for large systems, Macmillan series

in operations research, 1970.[18] D. Gross, J. F. Shortle, J. M. Thompson and C. .M. Harris, Fundamentals

of Queueing Theory, Wiley series in probability and statistics, Aug. 2008.[19] M. R. H. Mandjes and P. Zuraniewski, “M/G/infinity transience, and its

applications to overload detection,” Performance Evaluation, vol. 68, no.6, pp. 507-527, Feb. 2011.

Muhammad Ismail (S’10) received the BSc.and MSc. in Electrical Engineering (Electronicsand Communications) from Ain Shams University,Cairo, Egypt in 2007 and 2009, respectively. He isa research assistant and currently working towardshis Ph.D. degree at the Department of Electricaland Computer Engineering, University of Waterloo,Canada. His research interests include distributedresource allocation, quality-of-service provisioning,call admission control, green wireless networks, andcooperative networking. He served as a TPC member

in the ICWMC in 2010, 2011, and 2012. He is serving in the IEEE INFOCOM2014 organizing committee as a web chair. He joined the International JournalOn Advances in Networks and Services editorial board since January 2012.He has been an editorial assistant for the IEEE Transactions on VehicularTechnology since January 2011. He has been a technical reviewer for severalconferences and journals (IEEE Communications Magazine, IEEE Transac-tions on Mobile Computing, IEEE Transactions on Wireless Communications,IEEE Communications Letters, International Journal in Sensor Networks, andIET Communications).

Atef Abdrabou (M’09) received the Ph.D. degree in2008 from University of Waterloo, Ontario, Canada,in electrical engineering. In 2010, he joined the De-partment of Electrical Engineering, UAE University,Al-Ain, Abu Dhabi, UAE, where he is an AssistantProfessor. Dr. Abdrabou is a co-recipient of a BestPaper Award of IEEE WCNC 2010. He receivedthe prestigious National Science and EngineeringResearch Council of Canada (NSERC) postdoctoralfellowship for academic excellence, research po-tential, communication, and leadership abilities in

2009. His current research interests include network resource management,QoS provisioning and information dissemination in self-organizing wirelessnetworks.

Weihua Zhuang (M93-SM01-F’08) has been withthe Department of Electrical and Computer Engi-neering, University of Waterloo, Canada, since 1993,where she is a Professor and a Tier I Canada Re-search Chair in Wireless Communication Networks.Her current research focuses on resource allocationand QoS provisioning in wireless networks. She is aco-recipient of the Best Paper Awards from the IEEEMultimedia Communications Technical Committeein 2011, IEEE Vehicular Technology Conference(VTC) Fall 2010, IEEE Wireless Communications

and Networking Conference (WCNC) 2007 and 2010, IEEE InternationalConference on Communications (ICC) 2007 and 2012, and the InternationalConference on Heterogeneous Networking for Quality, Reliability, Securityand Robustness (QShine) 2007 and 2008. She received the OutstandingPerformance Award 4 times since 2005 from the University of Waterloo,and the Premier’s Research Excellence Award in 2001 from the OntarioGovernment. Dr. Zhuang is the Editor-in-Chief of IEEE Transactions onVehicular Technology, and the Technical Program Symposia Chair of the IEEEGlobecom 2011. She is a Fellow of the Canadian Academy of Engineering(CAE), a Fellow of the IEEE, and an elected BoS member of the IEEEVehicular Technology Society.