Adaptive Resource Allocation for Multimedia QoS Management in Wireless Networks

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 2, MARCH 2004 547

Adaptive Resource Allocation for Multimedia QoSManagement in Wireless Networks

Lei Huang, Member, IEEE, Sunil Kumar, Member, IEEE, and C.-C. Jay Kuo, Fellow, IEEE

Abstract—Adaptive resource allocation for multimedia qualityof service (QoS) support in broadband wireless networks isexamined in this work. A service model consisting of three serviceclasses with different handoff-dropping requirements is presented.Appropriate call-admission control and resource-reservationschemes are developed to allocate resources adaptively to thereal-time service classes with a stringent delay bound. Moreover,we propose an effective and efficient measurement-based dynamicresource allocation scheme to meet the target handoff-droppingprobability. The nonreal-time applications, serviced by thebest-effort model, are supported. The system accommodatesadaptive multimedia applications to further reduce the blockingand dropping probabilities of real-time applications. Based on amultidimensional model analysis, simulations are conducted toevaluate the system performance. The simulation results show thatthe proposed system can satisfy the desired QoS of multimediaapplications under different traffic loads, while achieving highutilization.

Index Terms—Admission control, cellular network, handoff,quality of service (QoS), resource reservation, service model,wireless multimedia network.

I. INTRODUCTION

FUTURE broad-band wireless networks, such as thegeneral packet radio system (GPRS) and the universal

mobile telecommunications system (UMTS), will extendcurrent second generation (2G) voice-based wireless servicesto broad-band multimedia services through packet-switchedtechnology. Compared with wired networks, wireless networksprovide more freedom to communications at the cost of alower bandwidth, higher latency, and a higher burst error rate.Providing multimedia services with a quality of service (QoS)guarantee in such an environment presents more challengesdue to the limited bandwidth resource, the highly variableenvironment, and user’s mobility.

To address this complex problem, QoS in wireless net-works is considered at two levels, i.e. the application leveland the connection level. Application-level QoS is related toperceived quality at the user end and is commonly considered

Manuscript received March 2, 2002; revised July 10, 2003 and October 17,2003.

L. Huang is with Department of Electrical Engineering and Computer Sci-ence, Loyola Marymount University, Los Angeles, CA 90045 USA (e-mail:[email protected]).

S. Kumar is with Department of Electrical and Computer Engi-neering—Systems, Clarkson University, Potsdam, NY 13699 USA (e-mail:[email protected]).

C.-C. J. Kuo is with Integrated Media Systems Center and the Departmentof Electrical Engineering Systems, University of Southern California, Los An-geles, CA 90089-2564 USA (e-mail: [email protected]).

Digital Object Identifier 10.1109/TVT.2003.823290

in packet-switched networks. A set of parameters, such asdelay/delay jitter, error/loss and throughput, etc., are used todescribe application-level QoS. Since packet-switched net-works take advantage of a higher degree of multiplexing amongservices, packets for a certain service flow may experiencevarying delay, delay jitter, and loss. Efficient packet-accessprotocols and packet-scheduling schemes play key roles insolving these QoS problems. Connection-level QoS is relatedto connection establishment and management. It measures theconnectivity and continuity of service in a wireless network,mostly by two parameters: the new-call-blocking probability,which measures service connectivity, and the handoff-droppingprobability, which measures service continuity during handoff.For a mobile user, dropping an ongoing call is generally moreunacceptable than blocking a new call request. Therefore,minimizing the handoff-dropping probability is usually a mainobjective in the wireless system design. On the other hand, thegoal of a network service provider is to maximize the revenueby improving network resource utilization, which is usuallyassociated with minimizing the new-call-blocking probabilitywhile keeping the handoff dropping below a certain threshold.

In recent years, there has been increasing research interest insupporting connection-level, as well as application-level, QoSfor multimedia applications in wireless networks. Differentcall-admission control and resource-reservation schemes havebeen proposed to reduce the handoff-dropping probabilityand/or the new-call-blocking probability. One of the firstbandwidth-eservation schemes for handoff was introduced inmid 1980s [1]. In this scheme, a set of channels are perma-nently reserved—exclusively for handoff calls—to keep thehandoff-dropping probability lower than the new-call-blockingprobability. It was shown that this reservation scheme wasoptimal to minimize a linear objective function of these twoprobabilities under certain assumptions [2].

However, static reservation is not efficient for varyingtraffic conditions found in wireless networks. Lately, severaldistributed call-admission-control schemes have been proposedto dynamically calculate the required bandwidth in order tomaintain a low cell-overload probability [3], [4]. However,the statistical models used in calculation were not realistic.Moreover, these schemes were designed based on traditionalmobile networks with only voice traffic. Thus, they cannoteffectively handle a variety of connection bandwidths, trafficloads, and user’s mobility. In [5], the concept of shadow clusterwas introduced for resource reservation and admission controlto reduce the call-dropping probability by predictive resourceallocation. In this scheme, a shadow cluster represents a setof cells around an active mobile. However, how to determine

0018-9545/04$20.00 © 2004 IEEE

548 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 2, MARCH 2004

the shadow cluster was not explained clearly. Moreover, thisscheme required each base station in the shadow cluster topredict future resource demands according to the informationabout active mobile users’ bandwidth requirement, position,movement pattern, and time. Consequently, it is computation-ally too expensive to be practical.

Oliveira et al. [6] proposed an admission-control scheme forwireless multimedia applications by considering two types oftraffic, i.e. real-time and nonreal-time. This scheme dynamicallyreserves bandwidth in cells surrounding the one in which theconnection originated, to provide QoS guarantees in high-speedmultimedia wireless networks. However, bandwidth reservationin all neighboring cells is a waste of resource, as the mobileuser hands off to only one of them. Another admission-controlscheme was proposed based on a three-class service model forintegrated service packet networks with mobile hosts [7]. In thisscheme, each mobile host that is requesting a new connectionhas to provide its accurate mobility specification, which consistsof the set of cells the mobile host is expected to visit during itslifetime. This limits the flexibility gained from mobility.

In contrast, most of the research effort on application-levelQoS focuses primarily on the wireless local area network(WLAN) environment, where the connection-level QoSproblem is often ignored [8]–[10].

In this paper, we propose a system to provide appropriate QoSaccording to service requests from end users, under the con-straint of limited and varying bandwidth resources. The mainfeatures of the proposed system are highlighted as follows.

• It is based on a comprehensive service model consistingof three service classes (i.e. handoff-guaranteed, handoff-prioritized, and best-effort).

• It deploys different resource-reservation schemesadaptively for real-time service classes (i.e., handoff-guaranteed and handoff-prioritized) to guarantee theirconnection-level QoS through a connection-orientedvirtual-circuit service.

• It uses an efficient dynamic call-admission-controlscheme to meet the target handoff-dropping probabilityof real-time services.

• It exploits the rate-adaptive feature of multimedia ap-plications to further improve the efficiency of resourceutilization.

The rest of this paper is organized as follows. The proposedsystem for QoS provisioning is described in Section II. A mea-surement-based dynamic guard-channel scheme is developed toachieve the target handoff-dropping probability in Section III.Some mathematical analysis for the proposed system is pre-sented in Section IV and simulation results are provided in Sec-tion V. Finally, concluding remarks are given in Section VI.

II. PROPOSED ADAPTIVE QOS MANAGEMENT SYSTEM

A wireless communication network typically consists of afixed network backbone and a wireless access system. The fixednetwork part, through mobile switching centers (MSC), pro-vides connections between radio-access ports, often called basestations (BS). The BS in turn provides wireless connections

Fig. 1. Block diagram of our proposed QoS management system.

to mobile terminals (MT) located in its coverage areas (calledcells). BS are distributed over the geographical area where com-munication services are covered. Continuous service coverageover a larger service area is achieved by handoff, which is theseamless transfer of a call from one BS to the other as the mo-bile unit crosses cell boundaries.

The block diagram of our proposed QoS management systemis illustrated in Fig. 1. With QoS as the kernel, the proposedsystem allows different applications to request different QoSfrom the network through a service model. Application profilesare mapped into the service model by different forms of trafficspecifications. Network resources are adaptively allocated todifferent service classes by employing adaptive resource-allo-cation schemes, including call-admission control and resourcereservation, according to the service model and QoS require-ments. The adaptation module enables the negotiation of QoSbetween applications and networks whenever it is necessary.Each component is described in detail below.

A. Service Model

An appropriate service model that describes a set of offeredservices is the foundation of QoS provisioning [11]. ExistingQoS-aware networks, such as ATM [12], InteServ [13], andDiffServ [14], designed their service models based on QoSrequirements of applications in the corresponding networkinfrastructure. Based on this concept, we have designed ourservice model for multimedia applications in wireless networksas described below.

First, multimedia applications are classified into real-timeand nonreal-time applications according to their delay re-quirements. In order to achieve desired QoS for a real-timeapplication, it is usually necessary to maintain a minimumbandwidth during its lifetime. We adopt the virtual circuitconcept to establish connection for a real-time applicationrequest. The setup of a connection requires call-admissioncontrol and resource reservation to prevent network congestionand dropping of ongoing calls. We select the handoff-droppingprobability as the primary QoS requirement and assume that it

HUANG et al.: ADAPTIVE RESOURCE ALLOCATION FOR MULTIMEDIA QoS MANAGEMENT IN WIRELESS NETWORKS 549

has a more significant impact on the overall connection-levelQoS measurement.

For nonreal-time applications, we use the best-effort serviceadopted in traditional IP networks. Data from these applica-tions can be stored at a network node, such as the BS oran MT. Whenever the network has spare resources that areunused by real-time applications, these applications will beserviced under an appropriate scheduling algorithm. To im-prove resource utilization of the entire network, we also useresources reserved for real-time applications but not yet beingin use to carry nonreal-time data. No call-admission controlor resource reservation is required here. Based on our earlierdiscussion, we categorize applications to the following threeclasses.

• Handoff-guaranteed service represents real-time appli-cations that require absolute continuity, i.e., no handoffdropping is permitted before the call is completed.

• Handoff-prioritized service represents real-time applica-tions that can tolerate a reasonably low handoff-droppingprobability.

• Best-effort service represents nonreal-time applicationsthat do not need a minimum bandwidth to set up aconnection.

B. Application Profile

The above service model covers many application-level QoSaspects, such as delay, priority, and bandwidth adaptation, aswell as pricing and mobility aspects. The network uses differentapplication profiles for different service classes. For real-timeservice classes, including both handoff-guaranteed andhandoff-prioritized services, the minimum required bandwidthto meet the delay requirement is necessary. The applicationprofile also includes the required handoff-dropping probabilityfor real-time service classes. For the handoff-guaranteedservice, the target handoff-dropping probability should be0. For the handoff-prioritized service, the target handoff-drop-ping probability is bounded by .Moreover, an application requesting the handoff-guaranteedservice should also provide its mobility information so that thenetwork could predict the cells that the mobile is going to visitduring its lifetime. For the best-effort service class, there is nominimum bandwidth requirement. In this class, the traffic loadis described by the packet-generation rate and the packet size.

The service model covers a wide range of applications. Somecalls, such as emergency rescue or business transactions, cannotbe dropped before completion. These applications will requirethe handoff-guaranteed service. Since priority is usually asso-ciated with pricing, some applications such as normal conver-sation, which are not so critical, may be willing to be servedas handoff-prioritized service at a lower price. Another con-sideration could be user’s mobility, e.g., moving range and/orspeed. Mobile users with high mobility move quickly withina large area across many cells, e.g., a moving vehicle on thehighway. Handoff occurs frequently in this case. As a result,the probability that the call is dropped before its completionwould be high even if the handoff-dropping probability in eachindividual cell is relatively low. Thus, users with high mobility

would prefer the handoff-guaranteed service class or a lowertarget handoff-dropping probability.

C. Resource Allocation

For real-time service classes, resource allocation includescall-admission control (CAC) and resource-reservation (RR)mechanisms. These two mechanisms are closely related toeach other to achieve the desired QoS for a given application.A different resource-allocation scheme, as explained below, isused for each service class to provide appropriate QoS to thecorresponding applications.

• For the handoff-guaranteed service class, it is necessaryto reserve resources in other cells the mobile host mayvisit, which is indicated by its “application profile.” Thereserved resources can only be used by the correspondinghandoff-guaranteed call or the best-effort data call until thereserving handoff-guaranteed call arrives. This guaranteesresources to each handoff-guaranteed call upon handoff.CAC is simply based on whether resources are reservedsuccessfully.

• For the handoff-prioritized service class, aggregate re-sources are reserved for the handoff calls of this classto maintain a reasonably low target handoff-droppingrate. We design a measurement-based algorithm thatdynamically adjusts the threshold in the guard-channelscheme, which will be presented in Section III.

• The best-effort service class is serviced with the remainingresources, including those reserved but not being used byreal-time service classes. The above two real-time serviceclasses can preempt this service class, thus improving theoverall utilization of the network without sacrificing theQoS guarantee for real-time traffic.

In our scheme, there are four types of real-time traffic listedin the decreasing order of priorities: the handoff-guaranteedhandoff call (HGH), the handoff-prioritized handoff call (HPH),the handoff-prioritized new call (HGN), and the handoff-guar-anteed new call (HPN). The resource sharing among thereal-time traffic is performed by our resource-allocationscheme, as follows. The reservation for HGH calls is similar tothe complete partitioning (CP) policy [15]. It ensures enoughresource for admitted HGH calls. For HPH calls, a postreserva-tion [15] for a lower handoff-dropping probability requirementis achieved by the proposed dynamic guard-channel scheme.The HPN and HGN calls adopt the complete sharing (CS) [15]policy to use the remaining available resources.

Within each of the above four real-time traffic types, ourresource-allocation scheme does not differentiate media types.For example, narrow-band voice and wide-band video be-longing to the same traffic class will be treated equally. Inother words, we use the CS policy among different media typeswithin each traffic class. As indicated in [15], this scheme maydiscriminate against wide-band traffic, because a posteriorihigher priority may be given to narrow-band applications,especially when the overall traffic load is heavy. Since weuse the bandwidth adaptation feature of modern multimediaapplications, this shortcoming would be alleviated to a greatextent. Note that our system can be extended to employ other


schemes, such as prereservation, to wide-band traffic accordingto different design objectives.

To prevent bad calls that are violating their application profilefrom degrading the QoS of other conforming calls, a policingmechanism should be enforced by the network. In our system, ifa call of the handoff-guaranteed service class tries to enter a cellnot covered by its mobility pattern, it violates its service speci-fications. The policing mechanism will release the resources re-served for this call in all cells included in its application profileand treat it as a handoff-prioritized call. If this call still needs thehandoff-guaranteed service, it has to make a new request with anew application profile.

D. Rate-Adaptive Applications

The wireless network is a highly variable environment whereavailable link bandwidth may vary with network load andchannel condition. By using rate-adaptive features of manymultimedia applications, our proposed resource-allocationscheme can be adaptive to network conditions. For example,voice applications can be encoded at a rate ranging from 2 to128 KB/s by choosing appropriate encoding mechanisms ordynamically modifying the encoding parameters. Similarly,video applications can be made rate adaptive by using alayered coding method. For example, the MPEG-2 video/audiocompression standard [16] defines different layers and profilesto achieve SNR and spatial scalability. The lowest layer (i.e.,the base layer) consists of critical information for decoding theimage sequence at its lowest visual quality. Additional layersprovide increasing quality. Another promising approach foradaptation is the use of embedded coding schemes, such as thewavelet-based JPEG-2000 image-coding standard [17]. Insteadof a few discrete coding rates provided by a layered codingscheme, continuous bit rates can be achieved by cutting a singlecoded bit stream at almost any bit. Similarly, MPEG-4 [18],which is the new generation multimedia communication-codingstandard, has the fine-granular scalability (FGS) mode.

There has been a large amount of research in the bandwidthadaptation of multimedia services for wireless networks[19]–[23]. In our system, when rate-adaptive applications makea connection request to the network, they specify the rangeof bandwidths required to be supported by the network as

, where and denote theminimum and maximum bandwidth requirements, respectively.Adaptation first takes place while admitting a new call. If thenetwork has enough resources available, the request is admittedat ; otherwise, it is admitted at a lower bandwidth.If the network is overloaded and cannot be satisfied,the call is blocked. Bandwidth adaptation also takes place atthe time of handoff occurrence. A rate-adaptive connectionadmitted at could be handed off at a lower rate ifthe cell it is entering is heavily loaded. On the other hand, acall admitted at could be upgraded to a higher rateif the cell it is going to enter is underutilized. We use the rateadaptation for a call only upon its admission and handoff,because frequent changes in the quality are not desirable foraudio/visual applications.

Fig. 2. State-transition diagram for the handoff-prioritized service.

III. MEASUREMENT-BASED DYNAMIC

GUARD CHANNEL SCHEME

In this section, we describe a measurement-based dynamicguard-channel scheme that is designed for the handoff-priori-tized service class. Note that this scheme can be easily extendedto any other system with the objective of achieving a targethandoff-dropping probability.

A. Guard-Channel Scheme

Considering a single cell with a fixed amount of bandwidthcapacity of channels, the traditional guard-channel schemegives a higher priority to the handoff-call request as comparedto the new call request by reserving a portion of the channel re-source for handoff calls. More specifically, a new call request isadmitted only when there are less than channels occupied,where is a threshold between 0 and . On the other hand,a handoff request is rejected only when all C channels are oc-cupied. As a result, channels are the guard channels,which are used only by handoff calls.

The new call and handoff requests to a given cell are Poissonprocesses with rate and , respectively. The cell-residencetime for each call is exponentially distributed with mean .Each call requires one unit channel bandwidth. The arrivals ofnew- and handoff-call requests are independent of each other.Based on the above assumptions, a cell deploying the guard-channel scheme can be modeled by a M/M/C/C queuing systemwith a threshold state Th, as illustrated in Fig. 2.

The state space can be denoted by ,where is the number of occupied channels in the cell. Thesteady state distribution can be derived as [24]

,(1)

where

(2)Given the steady state distribution, the new-call-blocking

probability can be expressed as the probability of thesystem in the states where or more channels are occupied,i.e.,

(3)


The handoff-dropping probability is the probability that allchannels are occupied, i.e.,

(4)

Based on (3) and (4), it is obvious that .

B. Proposed Dynamic Guard-Channel Scheme

It can be seen from the above analysis that for the guard-channel scheme plays an important role in the new call-blocking

and handoff-dropping probabilities. For a singleservice-class network, increasing results in less resourcesreserved for handoff calls, thus strictly increasing the . Atthe same time, the decreases accordingly because more re-sources become available for new arriving calls. On the otherhand, decreasing has the opposite effect [25]. Moreover,the overall system utilization could also be influenced by thevalue of . Reserving more resources than needed by handoffcalls results in lower system utilization since reserved resourcescannot be used by new call requests. Thus, the selection of iscritical to system design. The value of should be selected sothat necessary and sufficient resources are reserved for handoffcalls to meet their QoS requirements in terms of . Appar-ently, the selection of , i.e. the shared reservation of resourcesfor handoff calls, should dynamically vary with changing trafficconditions.

The above discussion is based on the assumption of a singletraffic class. We extend it to multiple traffic classes by applyinga single threshold to reserve resources for HPH calls in apostreservation fashion. The remaining resources are used bythe new call requests from both HG and HP classes (i.e., HGNand HPN calls) in a CS fashion, as discussed in Section II-C. De-creasing the in our multiple traffic class network will reservemore resources for HPH calls. Unlike the single traffic networkconsidered in [25], this will intuitively result in the lower or thesame for HPH calls and higher or the same for HGN andHPN calls. Similarly, increasing has the opposite effect.

Thus, the objective of our dynamic guard-channel schemecan be stated as follow. For a given system with a targethandoff-dropping probability for HPH calls, threshold

should be selected to keep the resulting for HPH callsas close to as possible without exceeding it, while the

for new HGN and HPN calls should be minimized. In thefollowing sections, we propose a scheme to achieve this goal.

1) Measurement of Handoff-Dropping Proba-bility: Generally speaking, there are two approaches used toestimate the handoff-dropping probabilities: modeling basedand measurement based. The modeling-based approach [26],[27] uses theoretical models to deduct the probability viamathematical analysis by assuming some parameters of themodel. This approach gives a theoretical reference to theestimation such that the system can be designed in advance.However, the performance of this resulting algorithm dependson the conformance of the real system to the theoreticalmodel and the accuracy of these assumptions made forthose parameters. Generally, a real network system cannotbe approximated by a simple model without making some

Fig. 3. Proposed dynamic guard-channel CAC algorithm.

unrealistic assumptions. Thus, more elaborate models areusually needed to make better estimation. However, the moreelaborate the model is, the more complex it is to analyzeand the more sensitive it is to the accuracy of assumptions.The measurement-based approach uses observed networkconditions, obtained by some measurements, to do simpleestimation. With this approach, system design is conductedduring run time along with updated measurements. Theperformance of this approach is usually good considering itslight computational complexity and its adaptability to changingconditions of practical networks. For these reasons, we designa measurement-based algorithm that aims at achieving theobjective defined above by dynamically adjusting .

Basically, our proposed dynamic call-admission scheme isbased on measurements of the current . Compared withother measurement-based schemes relying on the measurementof current traffic conditions [3], [28], measuring the directlygives more accurate information and enables more efficientcontrol in order to achieve our objective. The proposed schemeworks as follows. At the beginning, an initial value of isselected for a given cell. The BS of this cell monitors itsfor HPH calls. When the measured reaches or exceeds thetarget value , is decreased by one unit (channel) sothat more guard channels are reserved for handoff requests.Otherwise, we increase by one unit (channel) to admit morenew call requests. The proposed algorithm is summarized withthe pseudocodes given in Fig. 3, where is the cur-rently occupied bandwidth and is the required bandwidthby the call request. Some design issues are discussed in thefollowing subsections.

2) Update Frequency: An important design issue that af-fects the performance of the scheme is how often and when toupdate the measurement of to dynamically adjust . A fre-quent update can keep pace with changing traffic conditions. Ifthe is not updated quickly enough to match the variation insystem conditions, it could result in lower channel utilization


(due to late increase in ) or higher (due to late decreasein ). However, a new value of does not immediately affectthe measured . Thus, very frequent updates could causeunnecessary fluctuations and burden the system. Therefore, atradeoff between fast response and system stability is desired.

We use a “prompt-decreasing/timer-increasing” strategy tochoose the timing for updating the value of in the proposedscheme. Whenever a handoff request is dropped, the BS checksthe current value of and decreases if .Since only handoff-dropping events could drive the higherthan the target value, this “prompt-decreasing” strategy can im-mediately take the corrective action by decreasing to avoidfurther handoff drops. At the same time, a timer is set. If thereare no further handoff drops upon the expiration of the timer, theBS checks if . If it is true, the threshold is in-creased to improve channel utilization. Otherwise, it renews thetimer. Before the timer’s expiration, if there are further handoffdrops, the BS again checks . If necessary, it further decreases

and resets the timer.3) Timer Setting: The timer setting allows the effect of de-

creasing being reflected in the change of , since it maytake several successful handoff calls before the measuredfalls below the target or before it gets to the steady state.The timer is renewed when the continues to drop (indicatedby no handoff dropping) until it falls below , which trig-gers an increase in , or further handoff dropping happenswhen , which means reserved resources are notenough for handoff and a further decrease in is needed. Thus,it avoids the unnecessary fluctuation of and provides neces-sary updating. The value of the timer is the maximum time thatthe system could be underutilized.

4) Initial Threshold: Another issue is the selection of theinitial threshold . Theoretically, it can be derived from math-ematical analysis conducted in the previous section accordingto given . This approach can quickly lead the systeminto a steady state. However, it requires modeling parame-ters such as the new call and the handoff-arrival rates, thecell-residence time, etc. Practically, when these parameters arenot available, we can simply set to the capacity of thecell . By applying the proposed dynamic scheme, it willreach a steady state after some time. Starting withmaximizes system utilization from the beginning. However, itmight be achieved at the cost of violating the boundin the initial period.

Note that the proposed CAC scheme, as illustrated in Fig. 3,is applied to the handoff-prioritized service class to achieve itstarget handoff-dropping probability in our system. However, thenew call admission of the handoff-guaranteed service class isalso subject to the dynamic changing threshold derived here.

IV. ANALYSIS FOR REAL-TIME SERVICE CLASSES

In this section, to further explain connection-level QoS provi-sioning for the real-time service class, we present mathematicalanalysis of our scheme. First, the handoff-guaranteed serviceclass is modeled individually and then the system is modeled byconsidering both real-time service classes, i.e., handoff-guaran-teed and handoff-prioritized services.

A. System With Handoff-Guaranteed Service

Let us consider a single cell with a fixed amount of band-width (a total of channels of the same bandwidth); we derivethe model for the handoff-guaranteed service. Let denote thenumber of cells a call is going to enter while traveling alonga certain path during its lifetime. Upon admission, the channelresource is required to be reserved in each of the cells that itis going to enter. We define random variable as the channel-occupation time of the call in the th cell, whichis equal to the time from the admission of the call until it leavesthe th cell. Thus, equals to the sum of independent randomvariables of the cell-residence time in cell , i.e.,

(5)

Assuming that the cell-residence time is exponentially dis-tributed with the mean cell-residence time , it can be shownthat has an -stage Erlangian distribution with the densityfunction

(6)

Here, the stage parameter decides the shape and moments ofthe Erlangian distribution. For , this is the same as theexponential distribution with the density function

(7)

The first cell is the originating cell where the resource is onlyoccupied for the period of the cell-residence time, i.e. .

It is assumed that the maximum number of cells a handoff-guaranteed call can traverse is . Then, a cell with such aservice class can be modeled as an -dimensional model,where dimension is for all the calls enteringthis cell as their th cell. Thus, the th dimension is an

queuing model with the Poisson arrival and the-stage Erlang departure . Let the number

of calls in the th dimension be . The state is the-tuple vector and the state space

is .Given the steady state distribution of the model, the

blocking probability in each dimension is

(8)

The new-call-blocking probability ofa handoff-guaranteed call, which would visit cells during itslifetime, can be computed as the probability that at least one ofthe cells has no available channel for the new call request.Under the assumption of a unified cell load in all cells, we have

(9)

Since handoff is guaranteed not to be dropped for this type ofcalls, the handoff-dropping probability .


Fig. 4. Comparison for adaptive and nonadaptive applications under a typicalscenario. (a) Handoff-dropping probability of the HP class and (b) threshold.

B. System With Both Handoff-Guaranteed andHandoff-Prioritized Services

Let us consider a cell with channels offering both thehandoff-guaranteed and handoff-prioritized service classesdescribed previously. The system can be analyzed by using amultidimensional model as follows. The state of a system is thevector , where and are the numbers ofoccupied (being used or reserved) channels by handoff-guar-anteed and handoff-prioritized calls, respectively. Here, wereserve channels exclusively for the handoff calls ofthe handoff-prioritized service class. Therefore, the number ofhandoff-guaranteed service class calls is subject to the upperbound of . The state space can be written as

and

The handoff-guaranteed service can be divided intosubdimensions, where the th subdimension has an -stageErlangian departure , denoting that the handoff-guaranteed call that is going to enter this cell as its th cell.

Fig. 5. Comparison of the new-call-blocking probability for adaptive andnonadaptive applications under a typical scenario. (a) HG class and (b) HPclass.

The state of the th subdimension denotes the number ofcalls in this subdimension. Then, .

Assuming that the arrival and departure of each dimensionare independent of each other, the steady state distributioncan be obtained numerically. Given the steady state distribution

and threshold for handoff-prioritized calls, we can de-rive the new-call-blocking probability and the handoff-droppingprobability for both types of calls as follows.

For the handoff-prioritized call, the new-call-blocking prob-ability for a given cell is the probability that or morechannels of this cell are occupied, i.e.,

(10)

The handoff-dropping probability is the probability that allchannels are being occupied, i.e.,

(11)


Fig. 6. Performance of real-time services for adaptive and nonadaptiveapplications under a typical scenario. (a) Handoff-dropping and(b) new-call-blocking probabilities.

For the handoff-guaranteed call, the new-call-blocking prob-ability is the probability that at least one of the cellsit is going to enter has no available channel for the new call re-quest. Under the assumption of a unified cell load in all cells,we have

(12)

where is calculated in (10). Here, the handoff-droppingprobability .

The above multidimensional M/Er/C/C queuing model isprimarily used to explain the proposed system, especially indifferentiating the handoff-dropping-probability requirementsof different service classes. However, we are not aware ofany generalized closed-form solution to such a system. Insteadof providing an analytical solution, it is possible to obtain anumerical solution with queuing model software such as QTS[29]. An alternative numerical solution can also be obtainedwith discrete-event simulation, such as the OPNET simulatorused in this paper, as described in the following section.

Fig. 7. Channel utilization under a typical scenario. (a) Real-time serviceclasses and (b) nonreal-time service class.

V. SIMULATION RESULTS

To evaluate the performance of the proposed QoS manage-ment system, a network model of a single cell with channelcapacity was constructed in OPNET, which is a discrete-event-driven simulator. Based on the analysis conducted in Sec-tion IV, a number of call generators generated Poisson arrivalsof new and handoff call requests from different service classes.We use the following notations for the simulation parametersthroughout this section.

• For handoff-guaranteed service:— : maximum number of cells a handoff-guaranteed

call traverses in its lifetime.— : mean arrival rates of new call requests from

handoff-guaranteed calls entering this cell as its thcell.

— : mean of the exponentially distributed cell-resi-dence time of the handoff-guaranteed calls. Thus, thecorresponding channel holding time of handoff-guar-anteed calls have -stage Erlangian distributions,

, and the mean channel holding times are .


• For the handoff-prioritized service:— : mean arrival rate of new call requests from

handoff-prioritized service class.— : mean arrival rate of handoff requests from

handoff-prioritized service class.— : mean of the exponentially distributed cell-resi-

dence time of the handoff-prioritized calls.• For the best-effort service:

— : mean packet arrival (Poisson) rate.— : packet size in bytes per packet.

A. Typical Scenario

To simulate a typical scenario, system parameters were setas follows. The capacity of each cell is channels withdata rate 4800 b/s per channel. For the handoff-guaranteed ser-vice, , call/s,

call/s, and call/s, s. For thehandoff-prioritized service, call/s, call/s,

s. The target handoff-dropping probability ofthe handoff-prioritized service was set to 0.01. For the best-ef-fort service, packets/s with fix-sized packets of

B/packet.We simulated three types of real-time multimedia traffic,

i.e., voice, audio, and video, each requiring one, two, andfour channels from the network, respectively. Among thegenerated handoff-guaranteed and handoff-prioritized calls,we randomly select 50% as voice, 25% as audio, and the re-maining 25% as video applications. The performance in termsof the new-call-blocking probability, the handoff-droppingprobability, and channel utilization were compared for adaptiveand nonadaptive applications. In the experiments for adaptiveapplications, audio and video calls were assumed to reducetheir rate to one channel under congestion. The results for a20-h simulation are shown in Figs. 4–7. We use HG and HPto denote the handoff-guaranteed and the handoff-prioritizedservice classes, respectively.

By using our proposed dynamic guard-channel scheme forthreshold selection, the HP service class achieves the targethandoff-dropping probability after a short initial unstable stage,as shown in Fig. 4(a), for both adaptive and nonadaptive ap-plications. The dynamic change of the threshold for achievingthe target handoff-dropping probability is shown in Fig. 4(b).The value of the threshold fluctuates more for nonadaptiveapplications, since rate-adaptive applications have the abilityto adapt to network conditions.

There is no handoff dropping for HG service class due tobandwidth reservation. Figs. 5(a) and (b) show the new-call-blocking probability for HG and HP service classes, respectively.The HG service has a higher new-call-blocking probabilitythan does HP service, since it requires more resources to bereserved to ensure the zero-handoff-dropping probability. Thecomparison between adaptive and nonadaptive applications inFig. 5 shows that the new call blocking of both real-timeservice classes are reduced significantly by exploiting the rateadaptability of multimedia applications. Similarly, the overallhandoff-dropping probability [Fig. 6(a)] and new-call-blockingprobability [Fig. 6(b)] of real-time services are lower for adaptiveapplications.

Fig. 8. Performance under varying real-time traffic load. (a) Handoff-droppingprobabilities, (b) new-call-blocking probabilities, and (c) channel utilization.

Fig. 7 shows the channel utilization of real-time and non-real-time service classes. The channel utilization of real-timeservice classes [Fig. 7(a)] increases slightly when incorporatingadaptive applications. Due to resource reservation, real-timeservice classes cannot fully utilize channels to maintain the


desired connection-level QoS. However, the nonreal-timebest-effort service class can exploit the reserved but unusedresources, as shown in Fig. 7(b), thus improving the overallutilization for both adaptive and nonadaptive cases.

B. Varying Traffic Loads

To investigate the system performance under different trafficloads, we conducted two sets of simulations, one with a varyingreal-time traffic load and the other with a varying nonreal-timetraffic load, as shown in Figs. 8 and 9, respectively.

In the first case, the new call arrival rate of thehandoff-prioritized service was set to vary from 0.05 to0.95 calls/s, while maintaining all other settings as those inthe above typical scenario. The results are shown in Fig. 8.Fig. 8(a) shows that the handoff-dropping probabilities of HPand HG service classes remain at the target with increasingreal-time traffic. Fig. 8(b) shows that the new-call-blockingprobability of both real-time service classes increases withincreasing real-time traffic. Fig. 8(c) shows that the bandwidthutilized by real-time service classes increases with the increasein the real-time traffic, while the throughput of nonreal-timetraffic decreases. As a result, channel utilization in ourscheme is quite high for real-time traffic, even without thebest-effort service. This is achieved by the proposed dynamicguard-channel scheme and rate adaptation. Also, the decreasein the throughput of the best-effort class traffic shows that thereal-time HG and HP traffic effectively preempts the ongoingbest-effort traffic.

In the second case, the packet arrival rate of the best-effortservice was set to vary from 5 to 100 packets/s, while keepingall other parameters the same as the above typical scenario. Thesimulation results are shown in Fig. 9. We see from this figurethat the traffic load of the nonreal-time service class does notdegrade connection-level QoS parameters of real-time serviceclasses, but influences the throughput of nonreal-time trafficitself.

The results shown in Figs. 8 and 9 demonstrate that the systemgives preference to real-time services to achieve their requiredQoS. On the other hand, the nonreal-time service improves thetotal utilization of network resources. Here, we assume that amechanism exists in the multiple access control (MAC) layer topermit the best-effort class packet to occupy the unused chan-nels, preempt them by real-time traffic, and take care of colli-sions. The results presented in Figs. 8(c) and 9(c) are obtainedunder the assumption of an ideal MAC layer and packet-sched-uling policy, which does not sacrifice the channel utilization inorder to resolve collision. Practically, the channel utilization bynonreal-time service and the resulting total channel utilizationmight be lower due to the limitation of practical MAC layermechanism.

C. Dynamic Guard-Channel Scheme

Results in previous subsections show that the proposeddynamic guard-channel scheme can achieve the targethandoff-dropping probability for real-time services undervarying traffic loads. In this subsection, we further investigatethe performance of the proposed scheme by setting differenttargets for the handoff-dropping probability of the HP serviceclass, while the handoff-dropping probability for the HG class

Fig. 9. Performance under varying nonreal-time traffic load. (a) Handoff-dropping probabilities, (b) new-call-blocking probabilities, and (c) channelutilization.

remains zero. The results are shown in Fig. 10. The nearlystraight line in Fig. 10(a) indicates that our measurement-baseddynamic guard-channel scheme can always achieve the targethandoff-dropping probability of the handoff-prioritized ser-vice class. Fig.10(b) indicates that, with an increasing target


Fig. 10. Performance of the dynamic guard-channel scheme under a varyingtarget handoff-dropping probability of the handoff-prioritized service class.(a) Handoff-dropping and (b) new-call-blocking probabilities.

handoff-dropping probability, the new-call-blocking probabili-ties of both real-time service classes tend to decrease slightly.

VI. CONCLUSION AND FUTURE WORK

In this paper, we proposed an adaptive QoS managementsystem in wireless multimedia networks. The proposed systemis based on a service model designed for both connection- andapplication-level QoS. Wireless multimedia applications areclassified into different service classes in the service modelby their application profiles. Based on the service model,adaptive resource allocation is performed for each service classby employing the appropriate CAC and RR schemes tailoredto the QoS requirements of the service class. Rate-adaptivemultimedia applications can be incorporated for further im-provement of the system performance. Through analysis andsimulations, it was demonstrated that the proposed system canmeet the QoS requirements of different service classes andachieve reasonably high network utilization.

We are implementing a more comprehensive service modelwith further considerations of application-level QoS require-ments. More investigation of the packet-level QoS, such as thedelay of packets, packet-loss probability, and their impact onthe application-level QoS perceived by end users, will be car-ried out.

REFERENCES

[1] D. Hong and S. S. Rapport, “Traffic modeling and performance analysisfor cellular mobile radio telephone systems with prioritized and non-prioritized handoff procedures,” IEEE Trans. Veh. Technol., vol. 35, pp.77–92, Aug. 1986.

[2] R. Ramjee, D. Towsley, and R. Nagarajan, “On optimal call admissioncontrol in cellular networks,” Wireless Netw., vol. 3, no. 1, pp. 29–41,1997.

[3] M. Naghshineh and S. Schwartz, “Distributed call admission control inmobile/wireless networks,” IEEE J. Select. Areas Commun., vol. 14, pp.711–717, May 1996.

[4] S. Wu, K. Y. M. Wong, and B. Li, “A new distributed and dynamic calladmission policy for mobile wireless networks with QoS guarantee,” inProc. 9th IEEE Int. Symp. Personal, Indoor and Mobile Radio Commu-nications, vol. 1, Sept. 1998, pp. 260–264.

[5] D. Levine, I. Akyildiz, and M. Naghshineh, “A resource estimation andcall admission algorithm for wireless multimedia networks using theshadow cluster concept,” IEEE/ACM Trans. Networking, vol. 5, pp.1–12, Feb. 1997.

[6] C. Oliveira, J. B. Kim, and T. Suda, “An adaptive bandwidth reservationscheme for high-speed multimedia wireless networks,” IEEE J. Select.Areas Commun., vol. 16, pp. 858–874, Aug. 1998.

[7] A. K. Talukdar, B. R. Badrinath, and A. Acharya, “Integrated servicespacket networks with mobile hosts: architecture and performance,”Wireless Netw., vol. 5, no. 2, pp. 111–124, 1999.

[8] J. M. Capone and I. Stavrakakis, “Delivering diverse delay/dropping qosrequirements in a TDMA environment,” in Proc. ACM MobiCom, vol.1, Sept. 1997, pp. 110–119.

[9] S. Choi and K. G. Shin, “A cellular wireless local area network with QoSguarantees for heterogeneous traffic,” in Proc. IEEE INFOCOM, vol. 3,Apr. 1997, pp. 1030–1037.

[10] N. R. Figueira and J. Pasquale, “Providing quality of service for wirelesslinks: wireless/wired networks,” IEEE Pers. Commun., vol. 6, pp. 42–51,Oct. 1998.

[11] S. Shenker, D. D. Clark, and L. Zhang, “Services or infrastructure: whywe need a network service model,” in Proc. 1st Int. Workshop on Com-munity Networking: Integrated Multimedia Services to the Home, 1994,pp. 145–149.

[12] R. Handel, S. Schroder, and M. Huber, ATM Networks: Concepts, Pro-tocols, Applications. Reading, MA: Addison-Wesley, 1998.

[13] R. Braden, D. Clark, and S. Shenker, “Integrated Service in theInternet Architecture: An Overview,” Internet Engineering Task Force,RFC1633, June 1994.

[14] K. Nichols, V. Jacobson, and L. Zhang, A Two-Bit DifferentiatedServices Architecture for the Internet, draft-nichols-diff-svc-arch-00.txt(Internet draft), Dec. 1997.

[15] B. Epstein and M. Schwartz, “Reservation strategies for multi-mediatraffic in a wireless environment,” in Proc. IEEE Vehicular TechnologyConf., 1995, pp. 165–169.

[16] Generic Coding of Moving Pictures and Associated Audio Information,ISO/IEC International Standard 13 818, Nov. 1994.

[17] JPEG 2000 Part 1 Final Committee Draft Version 1.0, ISO/IEC Interna-tional Standard N1646R, Mar. 2000.

[18] Overview of the MPEG-4 Standard, ISO/IEC N3747, Oct. 2000.[19] T. Kwon, I. Park, Y. Choi, and S. Das, “Bandwidth adaption algorithms

with multi-objectives for adaptive multimedia services in mobile net-works,” in Proc. 2nd ACM Int. Workshop on Wireless Mobile Multi-media, 1999, pp. 51–59.

[20] T. Kwon, J. Choi, Y. Choi, and S. Das, “Near optimal bandwidth adapta-tion algorithm for adaptive multimedia services in wireless/mobile net-works,” in Proc. IEEE Vehicular Technology Conf., vol. 2, 1999, pp.874–878.

[21] V. Bharghavan, K. Lee, S. Lu, S. Ha, J. Li, and D. Dwyer, “TheTIMELY adaptive resource management architecture,” IEEE Pers.Commun., Aug. 1998.


[22] S. K. Sen and S. Das, “Quality-of-service degradation strategies in mul-timedia wireless networks,” in Proc. IEEE Vehicular Technology Conf., May 1998, pp. 1884–1888.

[23] K. Lee, “Adaptive network support for mobile multimedia,” in Proc.ACM Mobicom, Nov. 1995, pp. 62–74.

[24] L. Klennrock, Queueing Systems Volume 1: Theory. New York: Wiley,1975.

[25] S. H. Oh and D. W. Tcha, “Prioritised channel assignment in a cellularradio network,” IEEE Trans. Commun., vol. 40, pp. 1259–1269, Jul.1992.

[26] T. Kwon and Y. Choi, “Call admission control for mobile cellularnetwork based on macroscopic modeling of vehicular traffic,” in Proc.IEEE Vehicular Technology Conf., May 1998, pp. 1940–1944.

[27] Y.-B. Lin, L.-F. Chang, and A. Noerpel, “Modeling hierarchical micro-cell/macrocell PCS architecture,” in Proc. IEEE Int. Conf. Communica-tions, vol. 1, May 1995, pp. 405–409.

[28] T. Kwon, Y. Choi, C. Bisdikian, and M. Naghshineh, “Measurement-based call admission control for adaptive multimedia in wireless/mobilenetworks,” in Proc. IEEE Wireless Communications and NetworkingConf., vol. 2, May 1999, pp. 540–544.

[29] D. Gross and C. M. Harris, Fundamentals of Queueing Theory, 3rded. New York: Wiley, 1998.

Lei Huang (M’04) received the B.S. and M.S. de-grees from Beijing University of Posts and Telecom-munications, Beijing, China, in 1993 and 1996,respectively, and the Ph.D. degree from the Uni-versity of Southern California, Los Angeles, CA,in 2003, all in electrical engineering.

Since August 2003, she has been an Assistant Pro-fessor in the Department of Electrical Engineeringand Computer Science, Loyola Marymount Univer-sity, Los Angeles, CA. Her research interests includedigital image and video coding, multimedia applica-

tions in wired and wireless networks, quality of service, and network security.

Sunil Kumar (M’98) received the B.E. degree inelectrical engineering from the National Instituteof Technology, Surat, India, in 1988 and the M.E.and Ph.D. degrees in electrical and electronicsengineering from the Birla Institute of Technologyand Science, Pilani, India, in 1993 and 1997,respectively.

From 1997 to 2001, he was a PostdoctoralResearcher and an Adjunct Faculty Member in theDepartment of Electrical Engineering—Systems,University of Southern California, Los Angeles.

From 2000 to 2002, he also was a Senior Consultant in industry on MPEG-4and JPEG2000 related projects and participated in JPEG2000 standardsactivities. Since 2002, he has been an Assistant Professor in the Departmentof Electrical and Computer Engineering, at Clarkson University, Potsdam,NY. He is the coauthor of a book and more than 50 technical publications ininternational conferences and journals. His research interests include robustimage and video compression techniques and QoS-aware resource managementin multimedia wireless and CATV networks.

C.-C. Jay Kuo (F’99) received the B.S. degree fromthe National Taiwan University, Taipei, in 1980 andthe M.S. and Ph.D. degrees from the MassachusettsInstitute of Technology, Cambridge, in 1985 and1987, respectively, all in electrical engineering.

He was a Computational and Applied Mathematics(CAM) Research Assistant Professor, Department ofMathematics, University of California, Los Angeles,from October 1987 to December 1988. SinceJanuary 1989, he has been with the Department ofElectrical Engineering—Systems and the Signal and

Image Processing Institute, University of Southern California, Los Angeles,where he currently has a joint appointment as Professor of both electricalengineering and mathematics. He has guided approximately 50 students to theirPh.D. degrees and has supervised ten postdoctoral research fellows. He is thecoauthor of six books and more than 600 technical publications in internationalconferences and journals. His research interests are in the areas of digital signaland image processing, audio and video coding, multimedia communicationtechnologies and delivery protocols, and embedded system design.

Dr. Kuo is Editor-in-Chief for the Journal of Visual Communication andImage Representation, Associate Editor for IEEE TRANSACTIONS ON SPEECH

AND AUDIO PROCESSING and Editor for the Journal of Information Scienceand Engineering and the RURASIP Journal of Applied Signal Processing. Heis also on the editorial board of the IEEE SIGNAL PROCESSING MAGAZINE. Heserved as Associate Editor for IEEE TRANSACTIONS ON IMAGE PROCESSING

from 1995 to 1998 and for IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS

FOR VIDEO TECHNOLOGY from1995 to 1997. He is a Fellow of SPIE and aMember of SIAM and ACM. He received the National Science FoundationYoung Investigator Award (NYI) and the Presidential Faculty Fellow (PFF)Award in 1992 and 1993, respectively.

Adaptive Resource Allocation for Multimedia QoS Management in Wireless Networks

Documents