Source and Channel Coding Issues for ATM Networks y. ECSE Department, Rensselaer Polytechnic Institute, Troy, NY 12180, U.S.A

Transcription

1 Source and Channel Coding Issues for ATM Neworks y V.Parhasarahy, J.W.Modesino and K.S.Vasola ECSE Deparmen, Rensselaer Polyechnic Insiue, Troy, NY 12180, U.S.A ParhasarahyV@indy.ce.com, fmodesin,vasolag@ecse.rpi.edu Absrac This paper discusses source and channel coding issues as applicable o ATM neworks. Asynchronous Transfer Mode (ATM) has rapidly emerged as he appropriae ranspor echnique for Broadband ISDN. Among he various services oered in fuure ATM neworks, packeized variable bi-rae (VBR) video is likely o be one of he larges users of bandwidh. However, packe loss is virually ineviable in such neworks for VBR video due o he sochasic naure of rac. Imperfecly recovered packes lead o error propagaion in represenaive video compression algorihms, paricularly hose using moion compensaion. As a resul, i would appear highly benecial o use some form of acive recovery scheme, such as forward error-conrol (FEC) coding, which oers he poenial bene of improved recovery in he even of packe loss and/or errors. This paper discusses dieren echniques of applying FEC incorporaing ideas of combined source-channel coding. Furhermore, i inroduces a simple code selecion sraegy which yields codes providing close o opimal performance. Vericaion of is eciency is provided by comparing performance of such seleced codes o informaion-heoreic bounds. 1 Inroducion The developmen of broadband neworks has led o he possibiliy of a wide variey of new and improved service oerings. Packeized video is likely o be one of he mos signican high-bandwidh users of such neworks. The ransmission of variable bi-rae (VBR) video oers he poenial promise of consan video qualiy bu is generally accompanied by packe loss which signicanly diminishes his poenial. In his paper, we sudy a class of error recovery schemes employing forward error-conrol (FEC) coding o recover from such losses. In paricular, we show ha a hybrid error recovery sraegy involving he use of y This work was performed while he rs auhor was a he ECSE Dep., Rensselaer Polyechnic Insiue. He is presenly a Thomson Consumer Elecronics, Indianapolis, USA. This work was suppored in par by ARPA under Conrac No. F C-0030.

2 acive FEC in andem wih simple passive error concealmen schemes oers very robus performance even under high packe losses. We discuss wo dieren mehods of applying FEC o alleviae he problem of packe loss. The convenional mehod [1]-[4] of applying FEC generally allocaes addiional bandwidh for channel coding while mainaining a specied average video coding rae. Such an approach suers performance degradaions a high loads since he bandwidh expansion associaed wih he use of FEC creaes addiional congesion ha negaes he poenial bene in using FEC. In conras, we sudy a more ecien FEC applicaion echnique in our hybrid approach which allocaes bandwidh for channel coding by hroling he source coder rae (i.e., performing higher compression) while mainaining a xed overall ransmission rae. More specically, we consider he performance of he hybrid approach where he bandwidh o accommodae he FEC overhead is made available by hroling he source coder rae sucienly so ha he overall rae afer applicaion of FEC is idenical o ha of he original unproeced sysem. Following his we characerize he sensiiviy of such a scheme o he choice of he paricular code. We devise a simple code selecion sraegy and demonsrae ha i yields codes providing close o opimal performance for a wide range of operaing condiions. 2 Preliminaries Beginning wih a broad sysem framework, we describe in his secion he coding deails as well as some adapaions o sui nework ranspor. We use an enropy-consrained subband coding scheme (ECSBC) employing pyramidbased hierarchical moion-compensaed predicion (HMCP) developed by Kim and Modesino [5, 6] due o is ecien encoding as well as muli-resoluion properies. Figure 1 provides a general block diagram of a video coding and prioriizaion scheme for ransmission over a packe-swiched nework. Alhough he ideas presened in his paper would be applicable o arbirary packe-swiched neworks, his paper focuses on ATM due o is emergence as he appropriae ranspor scheme for supporing B-ISDN. Fig. 1 diagram is generic in he sense ha i is applicable o any chosen source coding scheme (e.g., a subband or a DCT-based sysem such as MPEG) or ranspor coding scheme (e.g., single or muliple prioriies, wih or wihou FEC). The oupu of he video coder, in he form of parallel bi sreams, eners he prioriizaion and ranspor coder. For example, in he case of a subband-based coding scheme hese bi sreams migh resul from coding dieren subbands. In a DCT-based scheme, hey could resul from enropy-coding he DC and AC coeciens. These oupu bi sreams can hen be packeized individually and classied ino separae prioriy classes by he prioriizaion and ranspor

3 VARIABLE RATE BIT STREAM {b(1) } PRIORITY CLASSES HP MUX {b(2) } VIDEO ENCODER PRIORITIZATION AND TRANSPORT ENCODER TO NETWORK {b(s) } LP DEMUX ^(1) {b } HP {b ^(2) } FROM NETWORK VIDEO DECODER TRANSPORT DECODER {b ^(s) } LP Figure 1: A Generic Block Diagram of a Sysem for Video Transmission over a Packe-Swiched Nework. coding block in he gure. The applicaion of FEC would also be performed in his block. The allocaion of prioriy levels can be performed in a hierarchical muliresoluion manner o provide scalable video a dieren resoluions. As an example, he encoded bis from subbands 1-4 could be allocaed he highes-prioriy (HP) level (prioriy level 1) and subbands 5-16 allocaed he lowes-prioriy (LP) level (prioriy level 2). Then wo prioriy packe sreams would be he resuling oupu of he prioriizaion and ranspor encoder block. Correc recepion of subbands 1-4 would guaranee a low-resoluion video sequence while correc recepion of all 16 subbands would provide he highes-level video resoluion. The backward moion-compensaion scheme encodes/decodes he residual frame dierence wihou requiring explici ranspor of moion vecors. The frame difference afer enropy coding is prioriized, packeized and ranspored over he nework. Furher deails can be found in [7] and [9]. We employ RS codes for FEC as hey make use of he generaed overheads ecienly and have aracive minimum disance properies. Accordingly, hey can be used eecively for burs erasure recovery which will prove valuable in he face of correlaed cell loss. FEC is applied hrough inerlaced coding across packes by grouping he informaion bis in he packe ino q-bi symbols. The echnique used here is he same as he approach in [1]-[4]. The packe size we consider o illusrae our approach is 48 byes which corresponds o a sandard ATM cell payload. Deails including he delay implicaions and informaion heoreic argumens in suppor of inerlacing can be found in [7] and [8].

4 3 Performance Evaluaion We begin by describing a rule yielding FEC codes which provide excellen performance. Eciency of his policy has been conrmed in [7] by using more rigorous rae-disorion argumens. This is followed by a brief descripion of he channel used o model he end-o-end packe loss behavior. Finally, we provide descripions of he compuaion of informaion-heoreic bounds on performance. 3.1 Code selecion principle As we will show, he mos ecien FEC applicaion is performed by hroling he coding rae o accommodae he FEC overheads. Consequenly, he hroling operaion is o be minimized o preven sacrice in qualiy under ligh loads resuling in small packe loss raes. In oher words, a criical parameer is he code rae R = K=N which deermines he fracion of he overall rae allocaed o he source coding operaion. Therefore, i is desired o make K=N as close o 1 as possible, which is he ideal code rae under lossless condiions. A he same ime, o provide good proecion wih reasonable delay, i is desired o mainain he FEC coding delay and he decoded loss probabiliy below appropriae hresholds. Therefore, we choose he code which solves he consrained opimizaion problem: maximize: subjec o: R = K N Decoded loss prob L hreshold FEC coding delay D hreshold : (1) Here, L hreshold and D hreshold are he qualiy-of-service (QOS) consrains which refer o he hresholds below which he decoded loss probabiliy P dec and FEC coding delay are o be mainained. P dec refers o he packe loss probabiliy afer he FEC decoding operaion is performed. Numerical mehods for heir compuaion are provided in [7] and [9]. In such a formulaion, we assume he environmen o be jier oleran. We adop a brue force mehod of enumeraion in order o nd he required code. The maximizaion of K=N is over all RS codes including he shorened and exended codes. Alhough i is somewha edious, such an enumeraion needs o be performed only once for a given se of parameers. In he nex secion, we presen a Markov model used o capure he behavior of he packe loss process. 3.2 Modeling he end-o-end packe loss behavior Modeling packe loss in high speed inegraed neworks is a challenging problem. While a number of sophisicaed echniques and models have been developed, our ineres in modeling packe loss is only one componen in our

5 1 - ρ LL ρ LL LOSS RECEPTION ρ NN 1 - ρ NN Figure 2: Markov model represenaion of he packe loss behavior. end-o-end view of packe video. Thus, i is essenial ha our model be simple and racable. The mos imporan characerisic of packe loss in neworks which will challenge recovery schemes for video is he correlaion beween losses caused by buer overow and cell dropping by he nework (as a form of congesion conrol). Our Markovian loss represenaion emulaes he basic end-o-end loss characerisics in he enire nework wihou requiring deailed knowledge of he paricular opology. In paricular, we have modeled he packe loss behavior by a simple wo-sae Markov chain as illusraed in Fig. 2. Given a desired value of he seady-sae loss probabiliy, P L, once LL is chosen he oher parameer N N can be readily calculaed. 3.3 Compuing an upper bound on achievable performance In his secion, a performance bound is deermined for operaing a a given average source coding rae over a channel wih a cerain loss probabiliy. Our approach o calculaing he upper bound begins wih descripions of he bilevel and he packe-level behavior of he sysem. A he bi-level, we use he noion of a block inerference channel, rs developed in [10]. We hen model he packe loss behavior as an independen process which ypically holds rue under ideal inerleaving Modeling he bi-level ransmission as a block inerference channel The channel model we use for his purpose is a special case of a block inerference channel inroduced by Mcliece and Sark [10]. In [10], successive blocks of lengh m bis are serially ransmied over one of a nie number of disinc channels wih he choice made independenly, according o some specied disribuion, for each block. We consider only wo possible channels; he binary symmeric channel (BSC) and he binary erasure channel (BEC). The paricular channel in

6 use for any block is represened by he channel sae s, wih s = 0 corresponding o he BSC while s = 1 corresponds o he BEC. Generally, he probabiliy of selecing eiher of hese channels (he sae selecion process) would depend on he behavior of he block loss process. For simpliciy and racabiliy in our analysis, we assume in his secion ha he packe losses occur independenly wih probabiliy P L;k for he k h prioriy class (i.e., LL;k = P L;k ; N N;k = 1? P L;k ). As noed previously, his would generally be rue under ideal inerleaving and provides a heoreical upper bound on performance. The capaciy under perfec CSI of he block inerference channel, specialized o represen he packe loss process, can be deermined for he k h prioriy class as [7] C k (m) = C k = (1? P L;k )(1? H(p)); bis/channel use, (2) where P L;k represens he probabiliy of a packe being los Upper bound on he recepion qualiy Our ineres in his secion is in calculaing an upper bound on he video recepion qualiy for ransmission over a block inerference channel by relaing he capaciy in (2) o a deliy meric. Consider he ECSBC scheme where one could disribue he coding rae among he various subbands in many ways. The issue hen is o deermine he opimal disribuion of he coding rae among he various subbands so as o minimize he overall disorion under loss. The minimum disorion, D min, is given by solving he following rae-allocaion problem. subjec o D min = min R s;k KX R s = k=1 D k (R s;k ) (3) k=1 KX R s;k C k (4) In he above equaion, Ck denoes he capaciy compued for he k h prioriy class, K represens he oal number of prioriy classes and R s;k, he source coding rae allocaed o he k h prioriy class. Observe hen ha R s;k C k denoes he overall ransmission rae. Solving (3) requires he compuaion of he disorion rae characerisics D k (R s;k ) for each of he k prioriy classes. This disorionrae characerisics can easily be obained from he individual disorion-rae characerisics compued for each of he subbands belonging o ha prioriy class by seing up a similar opimizaion problem as in (3). There are sandard mehods of solving consrained opimizaion problems as in (3) (for example, he BFOS algorihm).

7 4 Resuls and Discussion In his secion, we use he performance evaluaion echnique described earlier in our resuls and discussion. In our FEC-based schemes, for hose dropped packes which are no recovered by he RS code, he missing region is hen obained by using a passive error concealmen scheme. In our ECSBC coder, he specic passive error concealmen scheme employed is ha of emporal inerpolaion. We now compare he hroled, unhroled and unproeced scheme performance. To do so we perform he following wo-sep experimen. We base he choice of he parameers of he wo-sae Markov loss model on simulaions of a muliplexer wih he video sources modeled as a discree-ime auoregressive process [11]. The model parameers mached he bi-rae saisics of he coder [11] operaing on he Fooball sequence. The parameers of he wo-sae Markov model, namely P L and LL were chosen o mach hose obained from he simulaion. Furher deails may be obained from [7] and [8]. Boh he hroled and unhroled scheme employed he RS(15,13) code a an inerleaving deph of 1. Resuls are illusraed in Fig. 3 in he high-load region as he congesion 31.6 Throled source coder/fec Unproeced Unhroled source coder/fec 28.2 Average received SNR Number of muliplexed sources Figure 3: Comparison of he 3 schemes. Muliplexer speed is 100 Mbps and buer size is 500. Source coding rae = 0.80 bis/pixel for he hroled scheme and 0.93 bis/pixel for he unhroled, unproeced scheme. RS(15,13) code is used a inerleaving deph = 1. is normally more severe in such cases. We use only a single prioriy srucure in his example as he primary purpose is o demonsrae he superioriy of he hroled approach wihou having o ge ino he nework prioriy and buer managemen issues. Furhermore, as he simulaions are ime-consuming, hey are resriced o a region of maximum ineres. The muliplexer operaing rae is 100 Mbps (FDDI speed) while he buer size chosen is 500 packes. As he

8 gure indicaes, wih increasing load, here is as much as a 6 db dierence beween he hroled and unhroled scheme and up o a 4 db dierence beween he hroled and unproeced scheme. This gure also conrms ha he unhroled scheme leads o poor performance under heavy loads due o added congesion. In fac, i behaves much worse han he unproeced sysem. We now employ he code selecion sraegy described earlier and deermine is performance. Figure 4 shows he opimized code raes for a paricular example of he Markov channel model. In he example, he loss probabiliy P L is 110?2 while D hreshold = 5 msec and L hreshold = 10?4. As menioned earlier, we do no include he jier as a consrain o simplify our evaluaion. Alhough 20 msec ypically represens he olerable end-o-end delay for ineracive applicaions [8], a value of 5 msec is chosen for D hreshold o allow for queueing delays which ypically dominae end-o-end delay. We also assume in hese examples a CCIR 601 video resoluion. Noice ha low opimized FEC code raes are obained when he operaing rae is quie small. This means ha subsanial hroling of he source coding rae, and hence reduced ransmission qualiy, is required o accommodae FEC overheads. Furhermore, observe ha he seleced code depends on he operaing rae. This occurs because he FEC coding delay depends on he operaing rae. Hence, a code which saises he delay hreshold a a given operaing rae may no do so a a lower operaing rae. Consequenly, he philosophy of using only one code (which was originally proposed for AAL 1)ndependen of he operaing rae is quesionable. A beer approach would be o pre-deermine a number of codes, one for each operaing rae for xed channel condiions. Figure 5 demonsraes he performance of dieren codes a various code raes. In his example, a dieren code is used for each prioriy class opimized independenly for ha class in accordance wih he respecive values of he loss parameers. The codelenghs illusraed (N = 7; 15; 63) represen he performance when he corresponding code is applied o boh prioriy classes. The opimized code for he high-prioriy class is RS(61,56) and for he low-prioriy class is RS(56,50). Though i seems ha he code rae which maximizes he performance for he codelengh of 63 does beer han he opimizing code, i should be noed ha he paricular code does no saisfy he delay hreshold of 5 msec in his paricular example. Noice ha he performance of he code of codelengh 63 is very close o he opimizing code rae as well as he informaion-heoreic bound on performance. Consequenly, as long as he code rae is properly chosen, a codelengh of 63 is sucien in providing good performance. This is paricularly ineresing since i indicaes ha codes of relaively small codelengh (he FEC decoding complexiy as well as he ne FEC delay inroduced depend on he codelengh) are sucien o yield good performance. Also shown in he gure is he performance of he RS(128,124) code which has been proposed for he AAL 1 layer [12]. The code opimized according o he selecion policy performs much beer han he RS(128,124) code. Though he opimized code was seleced here for a paricular value of he

9 Code rae, R=K/N Ideal case (no losses) Independen losses ρ LL = 0.1 ρ LL = Average operaing rae (bis/pixel) Figure 4: Typical prole of opimized code rae (vs.) ransmission rae. In his example, he loss probabiliy P L = 1 10?2 and D hreshold = 5 msec while L hreshold = 10?4. Normalized mean-squared disorion Code lengh = 7 Code lengh = 15 Code lengh = 63 RS(128,124) Opimized code: 20 msec Performance bound Mismach: ρ LL = 0.10, D hresh = 5 Mismach: ρ LL = 0.10, D hresh = Code rae, R=K/N Figure 5: Comparaive performance of a hybrid error concealmen scheme a dieren code raes. Loss probabiliy P L = 5 10?3 for he high-prioriy class, 1 10?2 for low-prioriy class, LL = 0:40, sequence is he Fooball sequence, N f = 12 and he overall average ransmission rae is 0.85 bis/pixel.

10 Markov chain parameers, such an accurae descripion of he loss behavior in he nework is seldom available. As a resul, i is imporan o consider he behavior of he opimized codes for slighly dieren values of he Markov chain parameers so ha heir behavior can be sudied under mismached condiions. This behavior is illusraed in he gure for wo cases, one of which was seleced wih a relaively sric 5 msec consrain on he delay and he oher a more relaxed 20 msec consrain. The 5 msec consrain limis he number of available codes ha mee he hreshold on he loss. As a resul, much lower code raes are required o achieve good performance. Even under mismached condiions, observe ha he opimized codes perform very well indicaing he robusness of he code selecion sraegy. 5 Conclusions In his paper, a generic ransmission scheme for robus ransmission of VBR video employing a class of hybrid FEC-based error recovery schemes was sudied. Such schemes are paricularly useful when here is high moion or rapid scene changes in he encoded video. In such a case, he eec of error propagaion due o imperfec packe recovery is grealy reduced by FEC. In using he acive error concealmen echnique, he performance of wo approaches for applying FEC was sudied. In he rs approach, addiional bandwidh was allocaed for he channel coding operaion. Such a scheme was shown o suer performance degradaion under higher loads in comparison wih an unproeced sysem. A more judicious approach o applying FEC was invesigaed in his paper where he bandwidh for FEC applicaion was allocaed by hroling he source coder oupu. Under moderae-o-high packe losses (characerisic of high nework loads), employing a hroled source-coder FEC applicaion was shown o oer signicanly beer performance compared o an unproeced sysem. The performance of he scheme is closely relaed o he code seleced. As a resul, i is imporan o devise a clever code selecion sraegy. A simple sraegy of selecing codes based on a consrained opimizaion echnique was oulined and is performance sudied. The resuls indicae ha he selecion of a single code for all operaing raes is quesionable. For mos cases, codes of smallo-moderae codelengh ( 63) performed very well as long as he code raes were properly chosen. The code selecion sraegy provided robus performance even under condiions of mismach in choice of he channel parameers for which he codes were seleced. These resuls were subsequenly evaluaed and veried using he MPEG-2 coder. The only noiceable dierence was ha he gains while employing MPEG-2 were lower [7] due o he relaive robusness of he scheme o loss.

11 References [1] N. Shachum and P. McKenney, \Packe Recovery in High-Speed Neworks using Coding and Buer Managemen," Proc. IEEE INFOCOM, San Fran., CA, pp. 124{131, June [2] H. Oha and T. Kiami, \A Cell Loss Recovery Mehod using FEC in ATM Neworks," IEEE Trans. Commun., vol. 39, no. 9, pp. 1471{1483, Dec [3] E.W. Biersack, \Performance Evaluaion of Forward Error Correcion in ATM Neworks," Proc. ACM SIGCOMM, Balimore, MD, pp. 248{257, Aug [4] A.M. McAuley, \Reliable Broadband Communicaion using a Burs Erasure Correcing Code," Proc. ACM SIGCOMM, Philadelphia, PA, pp. 297{306, Sep [5] Y.H.Kim, \Adapive Enropy Coded Predicive Vecor Quanizaion of Images," Ph.D Disseraion, Dep. of Elecrical, Compuer and Sysems Engg., Rensselaer Polyechnic Insiue, Troy, NY, [6] Y.H.Kim and J.W.Modesino, \Adapive Enropy-Coded Subband Coding of Image Sequences," IEEE Trans. Commun., vol. 41, no. 6, pp , June [7] V.Parhasarahy, \Transpor Coding Schemes for Digial Video Transmission over ATM Neworks," Ph.D Disseraion, Dep. of Elecrical, Compuer and Sysems Engg., Rensselaer Polyechnic Insiue, Troy, NY, June [8] V. Parhasarahy, J. W. Modesino and K. S. Vasola, \Reliable Transmission of High-Qualiy Video over ATM Neworks," submied o IEEE Trans. Image Processing. [9] V.Parhasarahy, J.W.Modesino and K.S.Vasola, \Design of a Transpor Coding Scheme for Variable-rae Video Transmission over ATM Neworks," acceped for publicaion o IEEE Trans. Sysems and Circuis for Video Tech. [10] R.J.McEliece and W.E.Sark, \Channels wih Block Inerference," IEEE Trans. on Inform. Theory, Vol. IT-30, pp , Jan [11] B.Maglaris, D.Anasassiou, P.Sen, G.Karlsson, J.D.Robbins, \Performance Models of Saisical Muliplexing in Packe Video Communicaions," IEEE Trans. Commun., Vol. 36, No. 7, pp , July [12] C.Parridge, Gigabi Neworking, Addison Wesley Publishing Company, 1993.