ABSTRACTDynamic bandwidth allocation (DBA) will

play an important role in future broadbandwireless networks, including the 3G and 4GWCDMA systems. A code-division generalizedprocessor sharing (CDGPS) fair schedulingDBA scheme is proposed for WCDMA sys-tems. The scheme exploits the capability of theWCDMA physical layer, reduces the computa-tional complexity in the link layer, and allowschannel rates to be dynamically and fairlyscheduled in response to the variation of trafficrates. Deterministic delay bounds for heteroge-neous packet traffic are derived. Simulationresults show that the proposed CDGPS schemeis effective in supporting differentiated QoS,while achieving efficient utilization of radioresources.

INTRODUCTION

Future wireless networks are expected to supportIP-based multimedia traffic with diverse quality ofservice (QoS) requirements. One of the promis-ing approaches to support differentiated QoS isto employ a central controller that can dynamical-ly allocate bandwidth to mobile users according tothe variation of channel condition and trafficload. Such a dynamic bandwidth allocation (DBA)scheme should be efficient in utilizing radioresources, and fair in scheduling services. By effi-ciency, we mean that a user can get as much ser-vice as needed whenever there are idle resourcesavailable in the networks. By fairness, we meanthat each user can be guaranteed the agreed-onservice rate, even though other users are greedyin demanding bandwidth. Fair scheduling schemeshave been studied for packet networks where het-erogeneous traffic flows may share the same link[1, 2]. One of the idealized fair scheduling disci-plines is called generalized processor sharing(GPS) [1]. It is evident that a fair distribution oflink capacity to heterogeneous traffic flowsaccording to the GPS discipline will not onlyguarantee QoS to each individual user, but alsobenefit the overall utilization of network re-sources. However, conventional implementationof GPS is based on a time scheduling approach

which may only be suitable for time-divisionmultiple access (TDMA)-based wireless networks[3-4], and cannot simply be adopted for code-divi-sion multiple access (CDMA) systems withoutsignificantly increasing system complexity [5].

Wideband CDMA (WCDMA) has been pro-posed as a key air interface technique for third-generation (3G) wireless systems, and will likelycontinue to be adopted as a candidate for 4G sys-tems that will provide differentiated service tomultimedia IP traffic [6]. With the capability ofdynamically varying user channel rates, WCDMAsystems can provide more flexibility in bandwidthallocation. Although the idea of DBA by varyingchannel rate in WCDMA systems was recentlyadopted for supporting multiple QoS [7-8], thefair scheduling issue has not been well addressed.In this article a code-division generalized proces-sor sharing (CDGPS) fair scheduling scheme isproposed that employs both DBA and the GPSdiscipline to provide fair services in a WCDMAsystem. The CDGPS scheduler makes use of boththe traffic characteristics in the link layer and theadaptivity of the WCDMA physical layer toachieve efficient utilization of radio resources. Itadjusts only the channel rate (service rate) of eachtraffic flow in the WCDMA system by varying thespreading factor and/or using a multiple of orthog-onal code channels, rather than allocating servicetime to each packet. This results in lower imple-mentation complexity of the CDGPS scheme thanfor a conventional GPS-based time schedulingscheme, while the performance is comparable tothat of the ideal GPS in terms of bounded delayand guaranteed bandwidth provision.

The rest of this article is organized as follows.The system model is briefly described in the fol-lowing section. The principle of dynamic band-width allocation with fair scheduling is discussed.We then describe the proposed CDGPS schemein detail. Performance of the proposed scheme isdemonstrated by simulation results, followed byconcluding remarks.

THE SYSTEM MODEL

We consider a frequency-division duplex (FDD)WCDMA cellular network where mobile stations(MSs) are interconnected with the wireline

IEEE Wireless Communications • April 200226 1070-9916/02/$17.00 © 2002 IEEE

DYNAMIC BANDWIDTH ALLOCATION WITHFAIR SCHEDULING FOR WCDMA SYSTEMS

LIANG XU, XUEMIN SHEN, AND JON W. MARKUNIVERSITY OF WATERLOO

Radio l

BS

MSC/router

Wireline backbone(Internet)

Uplinkscheduler

Downlinkscheduler

Dynamic bandwidthallocation will playan important role infuture broadbandwireless networks,including thethird- andfourth-generationwidebandcode-division multipleaccess systems.

TE C H N O L O G I E S F O R 4G MO B I L E CO M M U N I C AT I O N S

IEEE Wireless Communications • April 2002 27

Internet backbone network through base stations(BSs) and mobile switching centers (MSCs), asshown in Fig. 1. The wireless link in the FDDWCDMA system is characterized by orthogonalchannels in the downlink (from BS to MS) andmultiple access channels in the uplink (from MSto BS). A pair of bandwidth schedulers areassumed to reside in each BS. The schedulersallocate the power and rate of the channels inthe uplink and downlink, respectively, to allmobile users in the same cell covered by the BS.In this article our discussion focuses on theuplink, while similar approaches can be appliedto the downlink.

The physical data channels in the uplinkcomprises a small number of random accesschannels and a large number of dedicated chan-nels. Each mobile user is assigned to at leastone dedicated data channel, and shares the ran-dom access channels with other users in thesame cell. While signaling and short messagesmay be transmitted freely via the random chan-nels, most IP data flows are scheduled for trans-mission on the dedicated channels. The rate ofeach dedicated channel can be changed dynami-cally, by varying the spreading factor and/orusing multicode CDMA (MC-CDMA) [9]. How-ever, the total uplink capacity, in terms of thesum of all the uplink channel rates, is limited byintracell and intercell multiple access interfer-ence (MAI), and varies as the interferencevaries. It is assumed that the total capacity ofthe rate-variable dedicated channels is known bythe scheduler since the interference level can bemeasured and estimated at the BS. The totalcapacity of these dedicated channels is consid-ered the uplink radio resource shared by themobile users in the same cell.

Multimedia IP traffic (e.g., voice, video, anddata) is considered. QoS requirements of IPtraffic generally consist of two parts: delay andloss rate. In order to guarantee the QoS for allusers, each traffic source is shaped by a leakybucket regulator [10] with parameter (s, r),where s and r are token buffer size and tokengenerating rate, respectively, of the leaky bucket.In this article we use the (s, r) model to charac-terize IP traffic source in the uplink system.

DYNAMIC BANDWIDTH ALLOCATIONGPS FAIR SCHEDULING DISCIPLINE

The DBA scheduler is the core of the link-layercontrol entity, which in general:• Supports differentiated QoS for IP traffic• Makes efficient use of physical layer resources,

by taking into account the variation of channelconditions as well as the traffic loadIn WCDMA-based cellular networks, both

the channel rate and transmission power of eachuser should be dynamically adjusted, such thatthe user QoS requirements can be guaranteedwith high utilization of the radio resource. Whiledynamic rate allocation aims to satisfy the delayrequirement, dynamic power assignment (powercontrol) is applied to compensate for channelimpairment and to maintain a satisfactory lossrate. In practical systems, the signal-to- interfer-ence-ratio (SIR)-based power control mecha-

nism, which maintains a target SIR at the receiv-er side of each dedicated uplink channel, can beapplied to guarantee the loss rate requirement.As a result, the DBA scheduler can only be con-cerned with the delay due to the variation oftraffic. If the instantaneous arriving rate of atraffic flow exceeds its allocated bandwidth,packet delay becomes significant. Conversely, ifa traffic flow is allocated more bandwidth than itactually uses, the overall system utilizationdrops. Furthermore, the overall system band-width should be distributed to users in a fairmanner according to a fair scheduling disciplineso that each user is guaranteed its requiredbandwidth, and greedy users cannot affect theQoS performance of well-behaved users. Thus,the overall goal of the DBA is to effectively andefficiently perform fair scheduling for trafficflows in WCDMA networks.

GPS [1] is an ideal fair scheduling disciplinethat has attracted much attention to its applica-tion in both wireline [2] and wireless packet net-works [3–5]. A basic principle of GPS is to assigna fixed positive real number (namely weight),instead of a fixed bandwidth, to each flow, andto dynamically allocate bandwidth for all flowsaccording to their weights and traffic load. Con-sider N packet flows sharing a network link withtotal bandwidth C. Let the weight for flow i befi. Let Si(t, t) denote the amount of flow i trafficserved during an interval (t, t]. Then a GPS serv-er is defined as one for which the followinginequality holds for any flow i that is continuous-ly backlogged in any interval (t, t] [1]:

(1)S t

S tj Ni

j

i

j

( , )

( , ), , ,

tt

ff

≥ = º1 2

� Figure 1. The network structure.

Radio link

BS

MSC/router

Wireline backbone(Internet)MSC

BS

MS MS MS

Uplinkscheduler

Downlinkscheduler

IEEE Wireless Communications • April 200228

If all the flows are continuously served by theGPS server during interval (t, t], then Eq. (1)holds with equalities, and the allocated band-width of each user is exactly proportional to itsweight. Due to the burstiness of packet traffic,sometimes a user may not have packets to trans-mit and gives up its bandwidth for a while. Theexcess bandwidth can be distributed among allbacklogged sessions at each instant in propor-tion to their individual weight fi. This makes theGPS server efficient and fair in bandwidth allo-cation. Furthermore, it has been proven [1] thata tight delay bound can be guaranteed by theGPS server if a traffic flow is shaped by theleaky bucket regulator. Given that shaped flow iis modeled by (si, ri) and the guarenteed rate gi≥ ri, the delay bound is˙si/ri. With the bound-ed delay guarantee, the GPS server can effec-tively support user QoS requirements. It shouldbe mentioned that the ideal GPS server requiresEq. 1 to be held at any time interval, which isnot realizable in practical systems. In the time-division multiplexed link, for instance, only onepacket can be transmitted at a time; hence, Eq.1 cannot hold in a small timescale. In practice,

the well-known packet-by-packet GPS (PGPS)time scheduling scheme [1], also known asWeighted Fair Queueing (WFQ), can onlyapproximately satisfy Eq. 1 over a long intervalrelative to packet transmission time and providea larger delay bound than that of the ideal GPS.Nevertheless, the ideal GPS server can be founduseful to serve as a hypothetical reference sys-tem to facilitate the implementation of thePGPS scheme.

There are two approaches to realize DBAwith the GPS discipline in WCDMA systems:time scheduling and rate scheduling. Figure 2illustrates the difference between the two. Letthree packet flows share the WCDMA uplink. Inorder to facilitate the DBA, the uplink is dividedinto time slots, and the scheduling is performedslot by slot. With time scheduling, the channelrate of each user is usually fixed, and the DBAscheduler arranges the service time of each pack-et. With rate scheduling, in contrast, the DBAscheduler arranges the channel rate of each userfor each slot, and different packets in the sameflow may be transmitted at different rates. Thetime scheduling approach was adopted in [5] torealize GPS in a hybrid CDMA/TDMA system,which is an extension of the PGPS scheme. Withthe PGPS scheme, a hypothetical GPS serverneeds to be simulated for calculating a virtual ser-vice time for each packet. Based on the simulatedvirtual time, packets from different flows aresequenced and arranged for transmission. Howev-er, this packet-by-packet time schedulingapproach induces excessive computational com-plexity and signaling overhead for CDMA sys-tems, which makes it difficult to be implemented.In the following section, a dynamic rate schedul-ing approach is proposed, which is aimed at mak-ing use of the physical layer capability of theWCDMA system to reduce the implementationcomplexity of GPS.

� Figure 2. Rate scheduling and time scheduling.

Time scheduling (fixed rate)

Slotted radio link

Flows Packet

Rate scheduling (variable rate)

Slotted radio link

Flows Packet

� Figure 3. A rate-scheduled WCDMA system.

Request from other users

Rate scheduler(select spreading factorand number of codes)

Bandwidthrequest

Estimatebandwidth

Source(shaped)

Coding

Buffer state

BufferMS

transmitter(spreading)

Adjustpower

P

Wireless channel

Interference, noise

Adjust spreading factor and number of codes

BSreceiver

(despreading)Decoding

Estimate SIR SIR threshold

Power control

IEEE Wireless Communications • April 2002 29

THE CODE-DIVISION GPSFAIR SCHEDULING SCHEME

The proposed CDGPS scheme uses the GPSfair scheduling discipline to dynamically allo-cate channel rates. The scheme is developedbased on a rate-scheduled WCDMA system, asshown in Fig. 3. The source of each MS gener-ates a sequence of packets that enter a bufferafter error control coding. The buffered infor-mation bits are then converted to a direct-sequence CDMA (DS-CDMA) signal withsymbol rate Rc/Ns,where Rc is the chip rate andNs the spreading factor. Since the chip rate(spread bandwidth) is fixed, the spreading fac-tor Ns (number of chips per symbol) deter-mines the service rate (or channel rate) atwhich the buffer is emptied. If MC-CDMA isemployed (i.e., more than one spreading codecan be used by one user), the service rate canbe mRc/Ns, where m is the number of codechannels.

The channel rate of each MS is scheduledslot by slot. By the end of each time slot, theMS sends a short message through the randomaccess channel or its own dedicated channel tothe BS, reporting buffer state and the amountof traffic that arrived in the previous slot.Upon receiving bandwidth requests from allactive users, the scheduler allocates the chan-nel rate for the next slot, taking into accountuser QoS requirements and the availableuplink capacity. The resource allocation mes-sage, in terms of the spreading factor and thenumber of codes of each user, will be sent viadownlink channels before the start of the nextslot. Upon receiving a resource allocation mes-sage indicating that the rate needs to bechanged, the MS immediately adjusts i tsspreading factor and/or number of codes atthe start of the next slot. In order to maximizethe utilization of uplink capacity, the rates ofdifferent channels are changed simultaneously,and all the uplink channels are synchronizedat the slot level . This, however, does notrequire strict bit synchronization among differ-ent channels, and can be supported by asyn-chronous WCDMA systems. SIR-based fastpower control is performed to guarantee a tar-get bit error rate (BER) at the output of thereceiver at the BS by adapting the transmis-sion power of the MS in response to variationsof MAI and channel fading. When the channelrate is changed, the SIR threshold is adjustedaccordingly so that the target BER can bemaintained.

Figure 4 shows the queuing model of theproposed CDGPS scheme, where total linkcapacity C is shared by N flows. Each flow main-tains a connection with link rate Ci(k) duringthe kth time slot. The sum of Ci(k) over allusers should not exceed C. From the viewpointof any flow i, it is entering a single-server queu-ing system with service rate Ci(k). Differentfrom a conventional single-server queue, theservice rate Ci(k) may vary periodically. Foreach slot, the scheduler allocates adequate ser-vice rates to the N flows, using the followingscheduling procedure.

THE RATE SCHEDULING PROCEDURE

Let the pre-assigned weight for flow i be fi, i = 1,2,…, N. Let Si(k) denote the amount of session itraffic that would be served during time slot k.Then, according to the GPS scheduling discipline,Eq. (1) should hold for any flow i that is continu-ously backlogged in the time slot k. Let thescheduling period of the CDGPS scheme, that is,the slot length, be T. The proposed CDGPS serv-er allocates each Ci(k) using the following steps:Step 1. Let Bi(k) be the total amount of back-

logged traffic of flow i during time slot k. Esti-mate Bi(k), i = 1, 2,… N, as follows:

Bi(k) = Qi(tk) + ri(k)T, (2)

where tk is the end time of slot k – 1, and Qi(tk)is the backlogged traffic at time tk; ri(k) is theestimated traffic arrival rate of flow i duringtime slot k and can be estimated from past traf-fic measurement. The following two approachesto rate estimation can be used:• One-step estimation: Let ri(k) = ai(k – 1)/T,

where ai(k – 1) is the total amount of thearrival traffic (in bits) during time slot k – 1.

• Exponential averaging: Let tni and ln

i be thearrival time and length of the nth packet offlow i, respectively. The estimated rate of flowi, ri, is updated every time a new packetarrives:

(3)

where Tni = tni – tn–1

i and K is a constant.Step 2. Based on the estimated Bi(k), i = 1, 2,

…, N , determine each Si(k) which is theexpected amount of service received by the ithuser.

• If Bi(k) = 0, then Si(k) = 0.• If Bi(k) > 0, then Si(k) = giT, where

(4)

is the minimum rate guaranteed to user i.• If S i Si(k) < CT , where C is the maximal

amount of service rate that can be provided bythe network, the remaining network resourceis distributed to users who expect more thantheir guaranteed service giT. The distributionof the remaining network resources should be

gC

ii

jN

j

==Â

f

f1

r el

Te ri

new T K in

in

T KioldI

nIn

= - +- -( ) ,/ /1

� Figure 4. A queuing model of the CDGPS scheme.

Radio link (capacity = C)

Scheduler

Flow 1

CFlow 2

Flow N

Connection 2 (C2(k))

Connection 1 (C1(k))

Connection N (CN(k))

IEEE Wireless Communications • April 200230

in proportion to each user’s weight fi accord-ing to the GPS service discipline.

Step 3. The allocated channel rate to user i canbe determined by Ci(k) = Si(k)/T.

DELAY BOUNDLet traffic flow i be regulated by a leaky bucketwith token buffer size si and token generatingrate ri. Let Q*

i be the maximum flow i backlog(i.e., the maximum of Qi(t)). The delay boundfor the traffic flow i can be derived with the fol-lowing lemma:

Lemma 1: If gi ≥ ri, where gi is given by Eq.(4), then the maximum backlog Q*

i, £ si + riT.With lemma 1, the following delay bound for

flow i can be obtained.Theorem 1: If gi ≥ ri, then the packet delay of

flow i is bounded by

(5)

FAIRNESS

Similar to the ideal GPS server, a CDGPS sched-uler ensures weighted fairness to heterogeneoustraffic. The CDGPS server can guarantee eachbacklogged flow with at least the minimum rategi which is proportional to its weight fi. In addi-tion, the excess bandwidth available from flowsnot using their reserved rates is distributedamong all of the backlogged flows in each timeslot in proportion to their individual weights.This results in short-term fairness during a slotperiod T. The long-term fairness property of theCDGPS scheme can be explained by comparingthe delay bound of Eq. (5) with the ideal GPSdelay bound si/gi given in [1]. It can be seen thatthe difference between the two bounds is nomore than T, which implies that in the long termthe services provided to user i by CDGPS areequivalent to those provided by the GPS. Thisexplains the fact that the CDGPS ensuresweighted fairness, as does the ideal GPS.

The weighted fairness feature of the pro-posed CDGPS scheme is important for provid-ing differentiated services in WCDMA networks,since different traffic flows can be isolated whilesharing the same radio resource.

SIMULATION

In this section we present simulation results todemonstrate the delay performance, fairness fea-ture, and system utilization of the proposedCDGPS scheme. In the simulation, WCDMAuplink channels with a total capacity C = 2 Mb/sare used. Perfect power control on the WCDMAuplink and error-free channels are assumed. Thesimulated WCDMA system can support threeclasses of packetized traffic: voice, video, and besteffort data. A one-step traffic rate estimationapproach is adopted in the CDGPS rate schedul-ing procedure. The scheduling period T is 10 ms.

We first simulate four homogeneous besteffort packet data flows. Each of these flows ismodeled by a Poisson process with average arriv-ing rate l and packet length L, shaped by a leakybucket regulator with s = 20L and r = C/4. In

Delayg

Ti

imax .£ +s

� Figure 5. Average delay comparison.

Load

0.100

0.1

0.12A

vera

ge d

elay

(s) 0.08

0.06

0.04

0.02

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

CDGPSGPSStatic

� Figure 6. Maximal delay comparison.

Load

0.200

0.05

Max

imal

del

ay (

s)

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.5

0.45

0.1 0.3 0.4 0.5 0.6 0.7 0.8 10.9

CDGPSGPSStatic

� Figure 7. Throughput comparison of CDGPS and GPS (same weights).

Time (s)

Flow 1

Flow 4

Flow 2

Flow 3

1.0512.5

3

Thro

ughp

ut (

bits

)

3.5

4

4.5

5

5.5x105

1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.51.45

GPSCDGPS

IEEE Wireless Communications • April 2002 31

the simulation, L = 5120 bits. l can be varied inorder to change the system load. A hypotheticalGPS scheduler with service rate C = 2 Mb/s issimulated, providing a performance benchmark.We also simulate a static rate allocation schemecompared to our proposed dynamic rate schedul-ing scheme. Unlike the CDGPS scheme, the stat-ic rate allocation scheme will assign a fixedchannel rate equal to the guaranteed rate gi inthe CDGPS scheme to each packet flow. Figures5 and 6 show average delay and maximal delay,respectively, with different system load. In thefigures the system load is normalized to be Load= 4 * l/C. We vary the packet arriving rate l ofeach flow to change the system load. From Figs.5 and 6 it can be seen that the delay performanceof CDGPS is close to that of GPS, and betterthan that of the static scheme. As expected, theidealized GPS can achieve lower average andmaximal delay. This is because the GPS is sup-posed to perform, hypothetically, bit-by-bitscheduling that can instantly respond to trafficvariation. The difference between the delay per-formance of CDGPS and GPS is mainly due tothe scheduling period T and does not significant-ly change with load. The maximal delay withCDGPS is less than a theoretical delay boundcomputed from Eq. (5), which is 0.21 s in thissimulation. The CDGPS is superior to the con-ventional static rate allocation scheme withrespect to delay performance. As shown in Fig. 5,the average delay with the static scheme is closeto that with CDGPS when the system load islight, but grows much faster than that withCDGPS when the system load increases. Asshown in Fig. 6, when the system load is light, themaximal delay with CDGPS is less than that withthe static scheme, and far below the theoreticaldelay bound. The theoretical delay bound ofCDGPS is only reached when the system loadapproaches 1. This is because CDGPS is moreflexible in allocating bandwidth and can make

use of idle network resources to improve delayperformance when the system load is light.

Figures 7 and 8 give the throughput compari-son of CDGPS and GPS. Figure 7 shows fourflows assigned to the same weights (f1 = f2 = f3= f4), while Fig. 8 shows them given differentweights (f1 = f2/2 = f3/3 = f4/4) (i.e., differentpriorities). It can be seen that both CDGPS andGPS schedulers can fairly allocate service ratesto different flows, according to their assignedweights, and the throughput using the CDGPSscheduler is very close to that of the ideal GPSscheduler, which demonstrates the weighted fair-ness property of CDGPS.

To demonstrate the performance of CDGPSin a heterogeneous traffic environment, 10 voiceflows, three VBR video flows, and four best

� Figure 8. Throughput comparison of CDGPS and GPS (different weights).

Time (s)

Flow 1

Flow 4

Flow 2

Flow 3

1.0512

2.5

Thro

ughp

ut (

bits

)

3

3.5

4

4.5

5

5.5x105

1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

GPSCDGPS

� Table 1. Delay performance for heterogeneous traffic.

Flow ID Traffic type Throughput (packets) Average delay (s) Max. delay (s) Delay bound (s)

Flow 1 3019 0.003850 0.011976 0.01258Flow 2 2253 0.004035 0.011875 0.01258Flow 3 1844 0.004375 0.010512 0.01258Flow 4 2425 0.004107 0.009738 0.01258Flow 5 Voice 3492 0.004246 0.012416 0.01258Flow 6 2798 0.004013 0.011944 0.01258Flow 7 2401 0.004260 0.012049 0.01258Flow 8 2360 0.004007 0.011951 0.01258Flow 9 2509 0.004772 0.012424 0.01258Flow 10 1530 0.004052 0.011853 0.01258

Flow 11 24017 0.002916 0.012382 0.01996Flow 12 Video 24182 0.002814 0.011012 0.01996Flow 13 23792 0.002864 0.010598 0.01996

Flow 14 2917 2.1883 3.8571 N/AFlow 15 Data 3025 2.7816 4.7348 N/AFlow 16 3034 2.8141 4.7192 N/AFlow 17 3021 2.2572 4.3340 N/A

Utilization 0.951

IEEE Wireless Communications • April 200232

effort data flows are simulated. Voice traffic isgenerated using an on-off model. In the on state,the voice packets arrive at a constant rate Rvo =16 kb/s = 200 packets/s, where the packet lengthLvo is 80 bits. The voice activity factor is assumedto be 0.4. VBR video traffic is generated byusing an 8-state MMPP model. In each state,packet arrival is assumed to be a Poisson pro-cess. The average duration in each state is cho-sen to be 40 ms,which is equivalent to the lengthof one frame of the video sequence with a framerate of 25 frames/s. A Poisson process is used togenerate the best effort data traffic with packetsize 2560 bits and average arrival rate 256 kb/s.Simulation results are summarized in Table 1,where the average and maximal delay of eachflow are compared to the analytical delay boundobtained from Eq. (5). Network resource (band-width) utilization is also calculated. It can beseen that low average packet delays are achievedby the real-time traffic flows, voice and video,under the proposed CDGPS scheduling scheme.The maximal packet delays are less than the the-oretical bounds. Table 1 also shows that the pro-posed scheduling scheme gives a high utilizationof the total uplink capacity.

CONCLUSION

We have proposed an efficient dynamic fairscheduling scheme to support QoS in WCDMAmultimedia packet networks. The analysis andsimulation results show that bounded delay canbe provisioned for real-time application by usingGPS service discipline, while high utilization ofsystem resources is achieved. The dynamic ratescheduling approach used by the proposedCDGPS scheme improves the delay performanceover the conventional static scheme. The CDGPSscheme is simpler to implement than convention-al GPS-based time scheduling schemes, and moresuitable to WCDMA systems. Research to designthe CDGPS fair scheduler by considering chan-nel impairments such as fading, imperfect powercontrol, and the fluctuation of intercell co-chan-nel interference are underway.

ACKNOWLEDGMENTSA preliminary version of this work was presentedat 3Gwireless 2001, San Francisco, California.The work has been supported by a grant fromthe Canadian Institute for TelecommunicationsResearch (CITR), a Network of Centers ofExcellence under the NCE program of the Fed-eral Government of Canada.

REFERENCES[1] A. K. Parekh and R. G. Gallager, “A Generalized Proces-

sor Sharing Approach to Flow Control in IntegratedServices Networks: The Single-Node Case,” IEEE Trans.Net., vol. 1, June 1993, pp. 344–57.

[2] D. Stiliadis and A. Varma, “Efficient Fair Queueing Algo-rithms for Packet-Switched Networks,” IEEE/ACM Trans.Net., vol. 6, Apr. 1998, pp. 175–85.

[3] P. Ramanathan and P. Agrawal, “Adapting Packet FairQueuing Algorithms to Wireless Networks,” Proc.ACM/IEEE MOBICOM’98, Dallas), TX, Oct. 1998, pp. 1–9.

[4] S. Lu, V. Bharghavan, and R. Srikant, “Fair Scheduling inWireless Packet Networks,” IEEE/ACM Trans. Net., vol.7, Aug. 1999, pp. 473–89.

[5] M. A. Arad and A. Leon-Garcia, “A Generalized Proces-sor Sharing Approach to Time Scheduling in HybridCDMA/TDMA,” Proc. IEEE INFOCOM ’98, San Francisco,CA, Mar. 1998, pp. 1164–71.

[6] W. W. Lu, “Compact Multidimensional Broadband Wire-less: The Convergence of Wireless Mobile and Access,”IEEE Commun. Mag., Nov. 2000, pp. 119–23.

[7] S.-J. Oh and K. M. Wasserman, “Dynamic SpreadingGain Control in Multiservice CDMA Networks,” IEEEJSAC, vol. 17, May 1999, pp. 918–27.

[8] O. Gurbuz and H. Owen, “Dynamic Resource Schedulingfor Variable QoS Traffic in W-CDMA,” Proc. IEEE ICC‘99, 1999.

[9] H. Holma and A. Toskala, WCDMA For UMTS: RadioAccess For Third Generation Mobile Communications,John Wiley Inc., 2000.

[10] M. Schwartz, Broadband Integrated Networks, Pren-tice Hall, 1996.

BIOGRAPHIESLIANG XU received a B.S. degree from Tsinghua University,China, in 1994 and an M.Sc. degree from Southeast Uni-versity, China, in 1997, both in electrical engineering. He iscurrently working toward a Ph.D. degree in the Depart-ment of Electrical and Computer Engineering at the Univer-sity of Waterloo, Canada. His research interests includedynamic bandwidth allocation, traffic scheduling, and QoSprovision in 3G and 4G wireless networks.

XUEMIN SHEN (xshen@bbcr.uwaterloo.ca) received a B.Sc.degree from Dalian Marine University, China, in 1982, andM.Sc. (1987) and Ph.D. degrees (1990) from Rutgers Uni-versity, New Jersey, all in electrical engineering. FromSeptember 1990 to September 1993 he was first withHoward University, Washington, D.C., and then the Univer-sity of Alberta, Edmonton, Canada. Since October 1993 hehas been with the Department of Electrical and ComputerEngineering, University of Waterloo, where he is an associ-ate professor. His research focuses on control algorithmdevelopment for mobility and re-source management ininterconnected wireless/wireline networks. In specific, hisinterests are traffic flow control, connection admission andaccess control, handoff, user location estimation, end-to-end performance modeling and evaluation, voice overmobile IP, stochastic process and filtering. He is coauthorof the books Singular Perturbed and Weakly Coupled Lin-ear Systems — A Recursive Approach (Springer-Verlag,1990) and Parallel Algorithms for Optimal Control of LargeScale Linear Systems (Springer-Verlag, 1993).

JON W. MARK [F] received a B.S.Sc. degree from the Univer-sity of Toronto, Ontario, Canada, in 1962, and M.Eng. andPh.D. degrees from McMaster University, Hamilton, Ontario,Canada, in 1968 and 1970, respectively, all in electricalengineering. From 1962 to 1970 he was with CanadianWestinghouse Co. Ltd. in Hamilton, Ontario, Canada,where he was an engineer and then a senior engineer. Hetook leave of absence from Westinghouse in October 1968to pursue Ph.D. studies at McMaster University under theauspices of an NRC PIER Fellowship. Since September 1970he has been with the Department of Electrical and Com-puter Engineering, University of Waterloo, Waterloo,Ontario, Canada, where he is currently a professor. He wasdepartment chair from July 1984 to June 1990. He hadbeen on sabbatical leaves at the IBM Thomas J. WatsonResearch Center, Yorktown Heights, New York (1976–1977)where he was a visiting research scientist, AT&T Bell Labo-ratories, Murray Hill, New Jersey (1982–1983) where hewas a resident consultant, Laboratoire MASI, UniversitéPierre et Marie Curie, Paris, France (1990–1991), where hewas an invited professor, and the National University ofSingapore (1994–1995 and 1999) where he was a visitingprofessor. In 1996 he established the Center for WirelessCommunications at the University of Waterloo where he iscurrently serving as its founding director. His currentresearch interests are in broadband communications, wire-less communications and wireless/wireline interworking.

The analysis andsimulation resultsshow that boundeddelay can beprovisioned forreal-time applicationby using GPS servicediscipline, while highutilization of systemresources isachieved. Thedynamic ratescheduling approachused by the pro-posed CDGPSscheme improves thedelay performanceover the conventionalstatic scheme.