The design of an efficient traffic scheduler with fair bandwidth-sharing for wireless multimedia services

(1)

The Design

of

A n Efficient

Traffic

Scheduler with Fair Bandwidth-Sharing

for Wireless Multimedia Services*

Fu-Ming Tsou Hong-Bin

C%iou

Gradua1.a Institutc

of

Commun. Enginccring

National Ta.iwan

University

d694200X?cc.

ntu.

(h.

trii

Northcrn

Ta.iw8.n Business

Group

Churighw-a Tclccommunicn.tioii

Co.

~

Ltjd.

hhchioii@ch.t.

com..

tu:

Zsehoiig

Tsai

G:raduat,c lnstitutc of C o m m u n . Enginccring

Nia.tional

Taiwan University

ztsniiil7cc. el?.

ntu.

l?dll.t?Il

Abstract- In this paper, the G e n e r a l i z e d W i r e l e s s D z n e r e n t i a t e a

F a z r Queuezny (GWDFQ) algorithm is proposed to accommodate

delay/jitter controls, and fair residual bandwidth sharing for real-time and non-real-time traffic streams simultaneously. The location-dependent channel error property, as appeared in most wireless networks, are considered in the algorithm and the tem- porary short error burst are compensated by the design of cred- its. The simulation results show GWDFQ can achieve excellent performance, including timely delivery of real-time traffic, virtu- ally loss-free transmission of non-real-t,ime t.r&c, and fair usage of channel bandwidth among remote stations.

I . INTFLODUCTION

NEVITABLY

,

t.he demands from customers in wireless access

I

will lean to high volume of real-time (RT) and non-real-time (NR,T) connectivit,y, which could often he beyond the available bandwidth. As a result, the issues of Quality of Service, fairness and pricing strategies should have expedited the emergence of service differentiatiori in such wirelebs access networks. How- ever, due to the fact that, characteristics of wireless channels can be very different from the wireline links, the traffic scheduler that satisfies the needs of wireless multimedia demands forms a complete new design.

A representat,ive of the scheduling algorithms that specifically design for wireless access and to handlc location-dependent error bursts is the idealized Wireless Fair-Queueing (IWFQ) algorithm proposed by Lu, Bharghavan and Srikant[l][:!]. However, IWFQ does not consider the delay/jit,ter requirements in wireless multimedia applications. In addition, the guarantees for throughput and delay in IWFQ are tightly coupled, and may not satisfy the requirements of multimedia applications[3]. The Channel-condition Independent packet, Fair Queueing (CIF-Q) algorithm[4] proposed by Ng, Stoica and Zhang, the enhanced Class-Based-Queueing (enhanced CBQ) scheme[5] proposed by Fragouli et al., and the effort,-limited fair (ELF) scheduling algorithm proposed by Eckhardt and Steenkiste[G] all provides long-term fairness and ensures delay and throughput guarantees for loss-free flows. However, either their implementat,ion complexity is still too high for a cost-effective implementation

or the flexible delay/jitter bounds may not be accommodated.

Hence, when fairness, differentiated QoS in delay/jitter, and link utilization are all taken inlo consideration, it is necessary to redesign a new scheduling algorithm for wireless multimedia. In this paper, we generalize the scheduling algorithm Wireless

Di#erentiatecif.irir Queueing (WDFQ) proposed in [7] to accommodate flexible delayljitter controls and fair residual bandwidth ‘This work was in part aupporaed by Ylstianal Science Council o f the R . 0 . C

under G r a n t ?ISC89-2213-E-002-087, and by Ministry of Education of t h e R.0.C

under G r a n t 89E-FA06-2-4-7.

sharing for

RT

and NRT traffic streams simult,aneously. The GWDFQ not only possesses the advantages of WDFQ, such as 1) timely delivery of delayljitter constrained RT traffic with cont,rolled packet, losses; 2 ) virtuallJi error-free transmission of

NRT

traffic; 3) shared utilization of the residual bandwidth for both

RT

and

NRT

traffic streams, but also provides flexiblc delay/jit,ter controls for

RT

Traffic streams via incurring limited FIFO queues. In additmion, t,he implementation of such t,raffic scheduler shall not require t,he use of sorter circuit and thus in- volve very limited complexity issues. The organization of this paper is as follows. In Section 3: the proposed traffic scheduler with fair residual bandwidth sharing is presentfed. Simulation results of R T and NRT traffic under various scenarios are shown in Section 3 . Our conclusions and future work are drawn in Sec- tion 4.

I 1. G E N E R A L I Z E D W I R E L E S S DIFFER.ENTIATED FA^

QUEUEING DISCIPLINE

.4. Minimum Ban.dwidth Guarantee for RT and NRT T r a f i c

Streams

The concept of minimum bandwidth guarantees and residual

bandwidth sharing[S] is adopted in GWDFQ. In the following, we define some necessary notations. Due to t,he limitation of space, the detailed definitions of each notation and the work- load calculation procedures are described in [7]. Note that, un- derlying layer-:! PDU is assumed t,o be fixed and is called the air packet for simplicit,y in the following context. The time unil in GWDFQ is “slot” which is the time interval to transmit an air packet.

(Bi,

Mi,

4;

): the traffic profile of flow i used in the service level agreement with respect, to the air interface, consist- ing of the maximum burst size

B;?

guaranteed minimum bandwidth

Mi

and the share weighting factor of residual bandwidth + i ;

T : t.he length of a refreshing period, which is the period for service workload calculation;

t,: the starting epoch of n-th refreshing period, and t, = t,-i

+

T for a

2 I;

W[(tn, t n , + l ) : the reserved workload for a backlogged flow

z within the a-th refreshing period, which is the service workload satisfying the traffic profile of flow i ;

L’Vf(ln, t n t l ): the m t e n d e d v~orkload for. a backlogged flow i within the w t h refreshing period, which is (.he workload contributed by the residual bandwidth observed within the n-th refreshing period;

W ; ( t n , t n + l ) : the total granted workload for a backlogged flow

i

within the n-th refreshing period, which is the sum of W,’(t,,t,+l)and W;(t,,t,+l);

(2)

Vi’,E(t,. t,+l): the eatro worklorsd for a backlogged flow z, which is contributed by the total granted workloads of the flows under bad channel stat,es at the stfartirig epoch of the refreshing period;

tji(tn,tv+l): the number of eligible air packets of flow I within the la-th refreshing period.

For convenience,

VVc(kn3

t , + l ) , Wt(t,,, t , , , + l ) , W F ( t l l , t,+l) and

W i ( t T b , t n + l ) are all normalized by the size or a single air packel. When the n-th refreshing period starts and the integrity of an air packet is taken into considerations, the reserved workload of’ a backlogged flow i can be calculated via the following recursive equation at trb:

~ W , T

+

B;] ,

+

~ ) A B , T

+

~ i

-It can be shown that eq. (1) is equivalent lo the resdt of

the leaky-buckpt, policing algorithm in [9]. However, the ac- tual arrival 1at.e may fluctuate, and the exlended workload of d

backlogged flow 1 at t,, is:

However, if flow J is found to be with the bad channel state

al t,. the workload it granted,

VV;(tn,

t T b + l ) and W;(t,,, t n + l ) ,

sliould be distributed fairly to backlogged flows undcr good chdnnel states, in order to achieve higher link utilization. Hence, backlogged flow z with good channel state can obtain extra workload, W,”(tn, t,,+l), which is expressed as

where G(t,) and E ( t n ) are the sets of flows under good channel state and under bad channel state a t time t,, respectively.

In addition, we use the concept of “credit” to compensat,e the loss or the overuse of bandwidth due to location-dependent emors and tcmporary short error burst. Detailed credit calculation algorithm arid how it works can be referred to [ 7 ] . B . Queueing ModeE of the GWDFQ Algorithm

The queueing model in this paper is a generalized version of model in [7] and is shown in Fig. 1. Some important mech- anisms in wireless networks, such as ackn,owIedgement, channel state detection, etc., are assumed to he supported by the underlying MAC protocol. The jitter bounds’ of all flows in the Group RT4 are within

((i

- 1)T,

ZT]

slot times, where

i

= 1,

.

. , N . In GWDFQ, each flow is associated with a class of service with a set of pre-determined air-packet level QoS pa- rameters, including delay/jitter, packet loss ratio, and residual bandwidth share, etc’.

In this model, non-real-time (NHT) traffic is assigned a spe- cial dedicated group, called Group NRT. The Head-of-Line (HOL) packet. of a flow queue is called an eligible packet if it can be transmitted immediately without violating its delay bound

‘In this paprr. we follow t h e definition of jrfter dracribed in [lo], w h r r e thr jitter of a flow (or I connection) is defined by t h e maximum absolnte difference in the delays rxperienced by any t w o packets on tha.t R o w .

3For NRT traffic, i t s delay and jit,ter limits arc aaaigned infinite.

. . . . . . lc ... ... : . . . . . . . . . Legend:

Fig. 1. T h e queueing model of the OWDFQ algorithm.

and packet jitter constraint. Each flow s is assigned two dedi- cated FIFO queues, called f l o w q u e u e i and rets-queue i. The function of f l o w q u e u e 1. is t o buffer the arriving air packets im- til they become eligible, while reta-queue i buffers t,he eligible

air packets whose channel states observed by the scheduler are under the “error” state now.

As far as RT traffic is concerned, three FIFO queues are re- quired for each

RT

group a t the output port, called C.7-queue, N-queueand R-queue, where C , N and R stand for conforming, nonconforming and retransmission, respectively. ‘The C- pueue.r

of Group i, denoted as Q!:.)? buffers the air packets conforming t o their service level agreements and the N-queue of Group

i,

denoted as

QC),

buffers the eligible but nonconforming air packets exceeding the service level agreements. In turn, the R - q u e u e of Group i (denoted as Qg,)) hufTers the eligible air

packets whose flows encountered bad channel states previously and are ready t o retransmit. The function and operation of

Q(R‘

will be described in details later.1t is noted that, the delay bound and jitter bound of each RT flow hence has t,o be ceiled as an integer mulliple of 7’.

On the other hand, for NRT traffic streams, only two extra FIFO queues, Q(;. and Qk:yRT), are needed. Nonconforming air packels of NRT traffic are still buffered in the corresponding

flow queues. Hence, the N-queue is not necessary in Group

NRT.

This special mechanism for NRT traffic is t,o assure the ( A’FlT)

(3)

air packet. sequence integrit.y of each flow. Each time when the refreshing procedure starts, the air packets whose total traffic load is within the wnerued workload and the extended workloud of each NRT flow are then moved to queue Q:”””, before their transmission.

Due to lack of space, the mechanism of delay a i d jitter control for RT traffic streams is riot described in lhis paper. However, it can be found in [7].

C. Operations of the G’CVDFQ .4lgorithm

Before describing the detailed operations of the GWDFQ algorithm, we briefly depict how GWDFQ deal with the retransmission mechanism under bad channel states.

Suppose that an air packet belonging to flow

i

is picked from certain C-yueue or R-queue t o transmit when the current chan- nel state is had. Then, this air packet is moved back to retz- yaeup i and the retransmission timer of flow

i

with period T: is st,artcd. Before the retransmission timer expires, all eligible air packets belonging to flow i are moved t,o reta-queue i . There- fore, it is not, necessary for the scheduler to check t.he channel stat.e slot by slot. Once the retransmission timer expires, the channel state of flow i is updated and t,he eligible air packetas in rets-queuei are then moved to the corresponding R-queues. Air packets in R-queues are the eligible and conforming air packets that were not be served due to the bad channel states in previ- ous refreshing periods. Hence, the R-queue is assigned highest service priority in a group to compensate their loss in bandwidth share. Last but not lcast, we have t~o note that t.he eligible air packets in N-yueuesare the air packets violating the traffic con- t,racts. Thus, air packets in N-queues axe of the lowest, priority and may he subjected to packet disr:arding if coriforming flow must be protected.

In the following, we describe the operations of GWDFQ. At each starting epoch of the refreshing period t,,, all eligible air packets in Q g ’ , Q:;’ and Q$) are discarded due to dclay/jitter

violations. Then, t,he eligible packets in Q$,)? Q$) and Qk’ are shifted to Q(H-l), Qg-’) and Q$-’), for

i

= 2, ’ . , N . In turn, the rcserved workload and the extended workload of each flow are calculated. Last, the eligible air packet,s from each flow queue are moved to the corresponding C-queues and N-queues

according to their reserved workloads and the extended workloads. The eligible air packets conforming to traffic contracts are moved to C- queues while nonconforming air packets are moved to N-queues. For

NRT

traffic, conforming

NRT

packets are moved to

Q,“

’’

while nonconforming packets are buffered in the flow queues. Then, the eligible packets are serviced in the se-

Q‘,“’. Via this service sequence, the delay/jit,ter requirements and the residual bandwidth shares cam be accommodated simultaneously t o

RT

multimedia streams.

quenceofQ$),

9:’

;...,

Q‘,“’,

QLN),

Q R (ArR.T)

, Qc

(NRT),QG),,.,,

111. SIMULATION RESULTS

In this section, we evaluate the performance of the GWDFQ scheme for

RT

and NRT traffic streams. The examined performance metrics include the packet loss ratio, t,he bandwidth usage. Note that the packet loss ratio only accounts for those packets discarded due to delay or jitter violations. Because the buffer size is assumed infinite, no packet losses are due t,o buffer overflow.

Here, we assume the wireless channel follows IEEE 502.11[12] with link bandwidth is 10 Mbps at, MAC: layer. According to [la], we assume the overhead of MAC layer is 30 bytes in all simulation experiments. In addition, we assume the pay- load of the air packet is 1 2 8 bytes, which includes RTP header

F r a m e Type

V a r i a n c e of f r a m e length M e a n f r a m e length ( a i r packers) M a x . f r a m e length ( a i r packets)

12 bytes, UTIP header 8 bytes, IP header 20 bytes and video frame data 88 bytes. Each video stream is a replay of “James Bond: Goldfinger” MPEG-1 video trace obtained from Uni- versit,y Wuerzburg[ 131, with equally separat,ed starting points within the 39996 frame positions. Since t,he frame rate is 24 frames/sec, each stream is equivalent, t,o a video of the length

1666.5 seconds. As for the sbatistical information of the video script, trace are shown in Table I. The refreshing period is set t,o he 1.0 ms and all simulations lasts for 5 x LO’ time slots, equivalent to 6320 seconds. Wc will show various of target delay constraints and jitter constraints can be supported easily. We have to note that as we mention the “bandwidth” or ar- rivalldeparture “rate,” t,he prot,ocol overheads from RTP layer. t.0 MAC layer are included.

T P R

118.82 59.34 1 6 . 4 3

1371.7a 820.63 35.57

348 296 79

TABLE I

THE GENERAL INFORM.4TION OF THE MPEG VIDEO T R I C E IN SIMULATIONS ALL STATICS H A V E INVOLVED RTP, TKJP A N P IP LAYER OVERHEADS.

The error characteristic of the wireless channel is modeled by a 2-state Markov chain. If the channcl state changes from GOOD state to BAD state suddenly during the air packet transmission period, the packet is received in error. The packet, is received correctly otherwise. Every air packet received in error is assumed t o be detected by the decoder.

In addition, we adopt the Priority FIFO algorithm as the baseline comparison, whose queueing model is shown in Fig. 2. The regulators, which serve as the front end packet proces- sor, perform nothing except forwarding packets conforming the traffic profile

(Bi,

Mi)

to the high-priority output queue and forwarding nonconforming packets to the low-priority output queue, where

Bi

is the maximum burst size and

Mi

is the guaranteed minimum bandwidth.

r-

4

... ~ ... .

rcgulaturs Prioritg PIP0 schcduler

F i g . 2 . T h e queueing model of t h e P r z o r l t y F I F O algorithm

A . E x p e n m e n t 1: Integrated Services with RT T ~ a f i c Streama 31, we examine the transient behavior of the GWDFQ algorithm for NRT traffic streams. ‘Two CBR flows (flow 1 and flow 2) are employed to model the

RT

t,raffic st,reams and their configurations are shown in Table 11. Flow 3 , serving as the background traffic, carries the NRT traffic stream which is a replay of LAN traffic trace obtained Irom the Lawrence Berkeley National Laboratory[l4]. In total, the original trace is used to generate the traffic with

and A‘RT

TTQBC

Streams

In this simulation scenario (see Fig.

(4)

average load equal t,o 7 . 2 Mbps. The oiit,put, link bandwidth is also assumed 10 Mbps. LTe assume that, the average time dura- tion of the wireless channel a t the GOOD (BAD) st,ate is 10000

( 1 000) time slots. The retransmission periods of three flows are

all 10 t,ime slot8s and credit limits are 20 (air packets). I n order

t,o observe t,he behavior of bandwidth sharing more clearly! we set the average arrival rate of Flow 2 and Flow 3 much higher than their guaranteed minimum bandwidth M ; , and the packet loss ratios for two

RT

flows are not considered in this simulation scenario.

Flow 1

Flow I

4 , = 0.2

Arrival Reserved Delay J i t t e r

1

Race R W Mi MBS Borind Roiind

(Mbps) (Mbps) R z 4% (msec) ( m s e c )

2.2 2.0 3 0 0 . 2 2 4 6

nurlc 1

Flow 3

I

9.2

I

2.0

I

2 0

I

0.6

I

N.A.

I

N.A. Flow 3

*3 = u.2

flow 1

flow 2

Fig. 3 . Simularion m o d e l for E x p e r i m e n t I

Arrival Reserved Delay J i t t e r

Rate R W M ,

’’’‘

RRS Round B o u n d

(Mbps) (Mbps) (msec) ( m s e c )

3.34 2.0 10 0.6 24 24

6 . 6 6 2 . 0 1 0 0.4 N.A. N.A.

TABTX IT

SIMULATION CONFIGUR.4TION F O R EXPERIMENT 1 , W H E R E N.A. STANDS FOR

“NOT AVAILABLE ” MBS STANDS FOR “M.4x1biuh.1 BURST SIZE” A N D ITS UNIT IS

I I R P 4 C K E T RBS R E P R E S E N T S ‘[RESIDUAL B.4NDWIDTH S H A R E ’‘

If the channel states of three flows are all GOOD, according to the concept, of GPS algorithm, it is easy t o derive the the ideal granted-service rate of flow 2 and flow 3 should be 2.95 Mbps and 4.85 Mbps. From Fig. 4 (a), we can observe that the channel states of three flows are all under GOOD states during the interval [18.8,19.2] sec. And Fig. 4 (b) shows that three flows approaches their ideal granted-service rates under GWDFQ dur- ing the this interval. During the interval [19.2, 19.31 seconds, flow 1 enters

BAD

channel state. Hence, the ideal granted- service rates of flow 2 arid flow 3 during within this interval should be 3.5 Mbps and 6.5 h‘lbps, respectively. Siniulation results shown in Fig. 4 (b) verify that the gpmted-service rate of flow 1 is distxibuted fairly to flow 2 and flow 3 when flow 1 is under BAD channel state. After time 19.3 sec, the channel state of flow 1 becomes “GOOD” once again arid it receive its granted-service rate right away. Other two flows also release their bandwidth granted from flow 1 when flow 1 was in BAD channel states. On the other hand, Fig. 4 (;) shows t,hat al- though Priority FIFO can guarant,ee the minimuni bandwidth, it cannot distribute the residual bandwidth to every backlogged flow fairly. Based on the simulation results obtained in

Ex-

periment 1, we coiiclude that GWDFQ not only provides QoS guarantees for RT traffic streams but also guarantees the bandwidth usage of NRT user groups following pre-determined traffic profiles.

(a) C h a n n e l state diagram of all flows.

I O ,

P I

1

(h) Transient bandwidth sharing b e h a v i o r s of GWDFQ

w i t h retransinissioii period 10 alotn.

( c ) Transient bandwidth sharing b-hauiors of P r i u r i l y

F I F O , assuming perfect channel knowledge.

Fig. 4 Simrilarion reaulrs of E x p e r i m e n r 1.

B. Experzrnenl 2: The Influence of the Length of the Retmns-

In this section, we study the influence of the length of the retransmi5sion period to find a best point t o achieve the best performance via minimum processing overhead.

mzsszon Period

TABLE I11

SIMULATION CONFIGURATION FOR EXPERIMENT 2 , WHERE N.A. REPRESENTS “NOT AVAILABLE ‘I &fBS STANDS FOR “ h 4 A X I M U M BURST SIZE” AND ITS UNIT IS

AIR PACKET RBS REPRESENTS “RESIDUAL BANDWIDTH S H A R E ”

The simulation model and configuration parameters are shown in Fig. 5 and Table 111, respectively. The average lengths of “GOOD” period and “error” period are fixed as 1000 slots and 25 slots, respectively. The length of the retransmission period varies from 5 slots to 50 slots to investigate the inhence of the length of the retransmission period. Flow 1 is a test video stream while flow 2 is an aggregat*ed regular

NET

flow driven by a LAN t,raffic trace.

(5)

I ~ l O U ~ 1

h

Fig. 5 . Packer loss r a t i o of flow 1 rinder v a r i o u s l e n g t h s of r e r r a n s m i s s i o n p e r i o d s .

if no control niechanisni is adopled, the performance of the flow

1 is much worse under Priority FIJ'O than under GWDFQ in t,his comparison. Hence, in the following discussions, thc perfor- mance of P r i o r i t y FIFO are not, included. On the other hand, in GWDFQ the packet loss ratio due to channel errors increases slightly as the retransmission period is less than 25 slots, the average length of the error period. Therefore, we conclude that, if we sel the retransmission period too small compared to the average length of error period, the processing overhead will be high and the performance enhancement will not be sufficient. On the other hand, if the retransmission period is set too large, the performance degradation due t,o error period becomes sig- nificant. Hence, we recommend that the retransmission period should be set close to the observed average length of the error period as much ax possible.

I. Stoica. H. Zhang, a n d T. 3 . E . Ng, " A Hirrarchical Fair Service Curve Al- gorithm for Link-Sharing, Real-Time and Priority Services." ACM Compot. Commun., vol. 27, p p . 247-262. Oct,. 1337.

T. S. E. Ng. I. Staica, and H. Zhang, "Packet Fair Queueing Algorithms f o r Wireless Networks with Location-Dependent, Errors," Proceedings nfIEEE

INFOCOM '$8, Mar. 1998.

C . Fragouli. V . Sivaraman, a n d M. B. Srivmtava. "Controlled Multimedia Wireless Link Sharing via Enhaiiced Class-Baaed Q u e u e i n g with C h a , n n & State-Dependent Packet Schrduling," Pracredings IEEE I.VFOC0M '98. Mar. 1998.

U . A . Rckhardt and P. Steenkiste, "Effort-limited Fair (ELF) Scheduling

f o r Wireless Y e t w o r k s , " Proceedings of IEEE INFOCOM 2000, p p . 1037-1106. 2000.

F.-lq. T s o n , H.-B. Chiou a n d 2 . Taai, " W D F Q : An Efficient Traffic Sched- uler with Fair Bandwidth-Sharing for Wireless Multimedia Services," l o a.ppzar in IEILT Tkans. Cornmrm., A p r . 2001.

A T Y Forum Technical Committee, "Tra.ffic Management Specification Ver-

aiori 4 . 1 , " AF-TM-0121.000, Mar. 1333.

I T U - T Recommendation 1.371, TrafXc Control a n d Congestion Curtdrol in B-ISDU, Aug. 1996.

D . Verma, H . Zhang, and D. Ferrari, "Delay J i t t e r Control f o r Real-Time Communication in a Packet Switching Network," Proceedings of IEEE 'W-

conmi '91, p p . 15-46, h p r l . 1991.

J . Liebeherr a n d D . E. Wrege, "Priority Q u e u e S c h r d u l e r s with Approxi- ma,tr Sorting in O u t p u t Buffered Switches," IEEE J . Select. Areas Commun., vol. 17. no. 6 , pp. 1127-1144, June 1999.

IEEE 802.11: I E E E Standa.rd f o r Wireless LAN Medium Access Control ( V A C ) and Physical Layer ( P H Y ) specificat,ions. June 2 6 , 1397. 0. Rose, MPEG-I Vrdeo Trace, James Bond: Goldfinger, Institute of Computer Science H I , University Wuerzburg, f t p : / / f t , p - i n f o 3 . i n f a r m r t i k . u n i -

wupr.burg.de/pub/MPEG/tracea, 1996.

W . Leland aiid D. Wilson, ftp://ita.ee.lbl.gov/traces/BC-pAug89.TL.Z,

1983. ~~ ~ ~ - ~ ~-

.~

*

.

~~~ ~~-~ 5 1" 15 2u 2; 30 36 10 45 50 ".OW1 LSngtll "l Rslranm,s.lm Psrlod j d O l J ,

!

. O o l i , , , , , , , GWDFQ -- ~ GWDFONeal

-~.

Pilarlty f l F 0

Fig. 6. T h e loss r a t i o of a i r p a c k e t s in GWDFQ and Priority F I F O under

varioiis r e t r a n s m i s s i o n p e r i o d s , w h e r e "G\T'DFQ_ideal" s t a n d s for t h e GWDFQ a l g o r i t h m w i t h full channel knowledge.

I V . CONCLUSIOXS

The proposed GWDFQ scheduling algorithm have been de- signed for transporting both RT streaming data and NRT traffic over wireless networks, and this algorithm can accommodate two different service levels (premium or regular) for

RT

or NRT traffic streams. As a result, the premium RT/NR'I service and regular RT/NRT service can be accommodated simultaneously via the same scheduler architecture. We have also illustrated that Lhe GWDFQ scheduler can provide fair access of residual bandwidth among all backlogged flows. As for supporting multiple service levels for RT traffic strcams, we believe the GWDFQ scheme c a n be easily extended to accommodate this requirement without increasing too much implementation cost. To summarize, timely delivery of RT traffic streams and vir- tually error-free transmission of

NRT

traffic are all well supported by GWDFQ. We believe that wireless niultimedia services can be supported more easily by employing GWDFQ-

enabled switches or base stations.

REFERENCES

[I] (21

S . L u ? V. Bharghavan, and R . Srikant, "Fair Scheduliiig in Wireless Packet

Networks," Proceedings o f A C M SIGCOMM ' Y i , pp. 63-74, 1997.

S. L u , V. Bharghavan, and R . Srikant, "Fair Scheduling in Wireless Packet Networks," IEEE/IICM Tkms. Networking, vol. 7, no. 4 , pp. 473-489, Aug.

1999.