Source and Channel Coding Issues for ATM Networksy

Source and Channel Coding Issues for ATM Networksy V.Parthasarathy, J.W.Modestino and K.S.Vastola ECSE Department, Rensselaer Polytechnic Institute, Troy, NY 12180, U.S.A Email: [email protected], fmodestin,[email protected]

Abstract This paper discusses source and channel coding issues as applicable to ATM networks. Asynchronous Transfer Mode (ATM) has rapidly emerged as the appropriate transport technique for Broadband ISDN. Among the various services oered in future ATM networks, packetized variable bit-rate (VBR) video is likely to be one of the largest users of bandwidth. However, packet loss is virtually inevitable in such networks for VBR video due to the stochastic nature of trac. Imperfectly recovered packets lead to error propagation in representative video compression algorithms, particularly those using motion compensation. As a result, it would appear highly bene cial to use some form of active recovery scheme, such as forward error-control (FEC) coding, which oers the potential bene t of improved recovery in the event of packet loss and/or errors. This paper discusses dierent techniques of applying FEC incorporating ideas of combined source-channel coding. Furthermore, it introduces a simple code selection strategy which yields codes providing close to optimal performance. Veri cation of its eciency is provided by comparing performance of such selected codes to information-theoretic bounds.

1 Introduction The development of broadband networks has led to the possibility of a wide variety of new and improved service oerings. Packetized video is likely to be one of the most signi cant high-bandwidth users of such networks. The transmission of variable bit-rate (VBR) video oers the potential promise of constant video quality but is generally accompanied by packet loss which signi cantly diminishes this potential. In this paper, we study a class of error recovery schemes employing forward error-control (FEC) coding to recover from such losses. In particular, we show that a hybrid error recovery strategy involving the use of y This work was performed while the rst author was at the ECSE Dept., Rensselaer Polytechnic Institute. He is presently at Thomson Consumer Electronics, Indianapolis, USA. This work was supported in part by ARPA under Contract No. F30602-92-C-0030.

active FEC in tandem with simple passive error concealment schemes oers very robust performance even under high packet losses. We discuss two dierent methods of applying FEC to alleviate the problem of packet loss. The conventional method [1]-[4] of applying FEC generally allocates additional bandwidth for channel coding while maintaining a speci ed average video coding rate. Such an approach suers performance degradations at high loads since the bandwidth expansion associated with the use of FEC creates additional congestion that negates the potential bene t in using FEC. In contrast, we study a more ecient FEC application technique in our hybrid approach which allocates bandwidth for channel coding by throttling the source coder rate (i.e., performing higher compression) while maintaining a xed overall transmission rate. More speci cally, we consider the performance of the hybrid approach where the bandwidth to accommodate the FEC overhead is made available by throttling the source coder rate suciently so that the overall rate after application of FEC is identical to that of the original unprotected system. Following this we characterize the sensitivity of such a scheme to the choice of the particular code. We devise a simple code selection strategy and demonstrate that it yields codes providing close to optimal performance for a wide range of operating conditions.

2 Preliminaries Beginning with a broad system framework, we describe in this section the coding details as well as some adaptations to suit network transport. We use an entropy-constrained subband coding scheme (ECSBC) employing pyramidbased hierarchical motion-compensated prediction (HMCP) developed by Kim and Modestino [5, 6] due to its ecient encoding as well as multi-resolution properties. Figure 1 provides a general block diagram of a video coding and prioritization scheme for transmission over a packet-switched network. Although the ideas presented in this paper would be applicable to arbitrary packet-switched networks, this paper focuses on ATM due to its emergence as the appropriate transport scheme for supporting B-ISDN. Fig. 1 diagram is generic in the sense that it is applicable to any chosen source coding scheme (e.g., a subband or a DCT-based system such as MPEG) or transport coding scheme (e.g., single or multiple priorities, with or without FEC). The output of the video coder, in the form of parallel bit streams, enters the prioritization and transport coder. For example, in the case of a subband-based coding scheme these bit streams might result from coding dierent subbands. In a DCT-based scheme, they could result from entropy-coding the DC and AC coecients. These output bit streams can then be packetized individually and classi ed into separate priority classes by the prioritization and transport

VARIABLE RATE

PRIORITY

BIT STREAM

CLASSES

{b(1)

}

{b(2)

}

t

t

MUX

HP

TO NETWORK

PRIORITIZATION AND TRANSPORT ENCODER

VIDEO ENCODER

{b(s) } t

LP

DEMUX ^ {b(1) } t

HP

^ {b(2) } t

FROM NETWORK TRANSPORT DECODER

VIDEO DECODER

^ {b(s) } t

LP

Figure 1: A Generic Block Diagram of a System for Video Transmission over a Packet-Switched Network. coding block in the gure. The application of FEC would also be performed in this block. The allocation of priority levels can be performed in a hierarchical multiresolution manner to provide scalable video at dierent resolutions. As an example, the encoded bits from subbands 1-4 could be allocated the highest-priority (HP) level (priority level 1) and subbands 5-16 allocated the lowest-priority (LP) level (priority level 2). Then two priority packet streams would be the resulting output of the prioritization and transport encoder block. Correct reception of subbands 1-4 would guarantee a low-resolution video sequence while correct reception of all 16 subbands would provide the highest-level video resolution. The backward motion-compensation scheme encodes/decodes the residual frame dierence without requiring explicit transport of motion vectors. The frame difference after entropy coding is prioritized, packetized and transported over the network. Further details can be found in [7] and [9]. We employ RS codes for FEC as they make use of the generated overheads eciently and have attractive minimum distance properties. Accordingly, they can be used eectively for burst erasure recovery which will prove valuable in the face of correlated cell loss. FEC is applied through interlaced coding across packets by grouping the information bits in the packet into q-bit symbols. The technique used here is the same as the approach in [1]-[4]. The packet size we consider to illustrate our approach is 48 bytes which corresponds to a standard ATM cell payload. Details including the delay implications and information theoretic arguments in support of interlacing can be found in [7] and [8].

3 Performance Evaluation We begin by describing a rule yielding FEC codes which provide excellent performance. Eciency of this policy has been con rmed in [7] by using more rigorous rate-distortion arguments. This is followed by a brief description of the channel used to model the end-to-end packet loss behavior. Finally, we provide descriptions of the computation of information-theoretic bounds on performance.

3.1 Code selection principle As we will show, the most ecient FEC application is performed by throttling the coding rate to accommodate the FEC overheads. Consequently, the throttling operation is to be minimized to prevent sacri ce in quality under light loads resulting in small packet loss rates. In other words, a critical parameter is the code rate R = K=N which determines the fraction of the overall rate allocated to the source coding operation. Therefore, it is desired to make K=N as close to 1 as possible, which is the ideal code rate under lossless conditions. At the same time, to provide good protection with reasonable delay, it is desired to maintain the FEC coding delay and the decoded loss probability below appropriate thresholds. Therefore, we choose the code which solves the constrained optimization problem: maximize: R = KN (1) subject to: Decoded loss prob  Lthreshold FEC coding delay  Dthreshold : Here, Lthreshold and Dthreshold are the quality-of-service (QOS) constraints which refer to the thresholds below which the decoded loss probability Pdec and FEC coding delay are to be maintained. Pdec refers to the packet loss probability after the FEC decoding operation is performed. Numerical methods for their computation are provided in [7] and [9]. In such a formulation, we assume the environment to be jitter tolerant. We adopt a brute force method of enumeration in order to nd the required code. The maximization of K=N is over all RS codes including the shortened and extended codes. Although it is somewhat tedious, such an enumeration needs to be performed only once for a given set of parameters. In the next section, we present a Markov model used to capture the behavior of the packet loss process.

3.2 Modeling the end-to-end packet loss behavior Modeling packet loss in high speed integrated networks is a challenging problem. While a number of sophisticated techniques and models have been developed, our interest in modeling packet loss is only one component in our

1- ρ LL

ρ

RECEPTION

LOSS

ρ

LL

NN

1- ρ

NN

Figure 2: Markov model representation of the packet loss behavior. end-to-end view of packet video. Thus, it is essential that our model be simple and tractable. The most important characteristic of packet loss in networks which will challenge recovery schemes for video is the correlation between losses caused by buer over ow and cell dropping by the network (as a form of congestion control). Our Markovian loss representation emulates the basic end-to-end loss characteristics in the entire network without requiring detailed knowledge of the particular topology. In particular, we have modeled the packet loss behavior by a simple two-state Markov chain as illustrated in Fig. 2. Given a desired value of the steady-state loss probability, PL , once LL is chosen the other parameter NN can be readily calculated.

3.3 Computing an upper bound on achievable performance In this section, a performance bound is determined for operating at a given average source coding rate over a channel with a certain loss probability. Our approach to calculating the upper bound begins with descriptions of the bitlevel and the packet-level behavior of the system. At the bit-level, we use the notion of a block interference channel, rst developed in [10]. We then model the packet loss behavior as an independent process which typically holds true under ideal interleaving.

3.3.1 Modeling the bit-level transmission as a block interference channel

The channel model we use for this purpose is a special case of a block interference channel introduced by Mcliece and Stark [10]. In [10], successive blocks of length m bits are serially transmitted over one of a nite number of distinct channels with the choice made independently, according to some speci ed distribution, for each block. We consider only two possible channels; the binary symmetric channel (BSC) and the binary erasure channel (BEC). The particular channel in

use for any block is represented by the channel state s, with s = 0 corresponding to the BSC while s = 1 corresponds to the BEC. Generally, the probability of selecting either of these channels (the state selection process) would depend on the behavior of the block loss process. For simplicity and tractability in our analysis, we assume in this section that the packet losses occur independently with probability PL;k for the kth priority class (i.e., LL;k = PL;k ; NN;k = 1 ? PL;k ). As noted previously, this would generally be true under ideal interleaving and provides a theoretical upper bound on performance. The capacity under perfect CSI of the block interference channel, specialized to represent the packet loss process, can be determined for the kth priority class as [7] Ck (m) = Ck = (1 ? PL;k )(1 ? H (p)); bits/channel use, (2) where PL;k represents the probability of a packet being lost.

3.3.2 Upper bound on the reception quality Our interest in this section is in calculating an upper bound on the video reception quality for transmission over a block interference channel by relating the capacity in (2) to a delity metric. Consider the ECSBC scheme where one could distribute the coding rate among the various subbands in many ways. The issue then is to determine the optimal distribution of the coding rate among the various subbands so as to minimize the overall distortion under loss. The minimum distortion, Dmin , is given by solving the following rate-allocation problem. K X Dmin = min Dk (Rs;k ) R

s;k

subject to

Rs =

(3)

k=1

XK Rs;k k=1

(4)

Ck

In the above equation, Ck denotes the capacity computed for the kth priority class, K represents the total number of priority classes and Rs;k , the source coding rate allocated to the kth priority class. Observe then that RC denotes the overall transmission rate. Solving (3) requires the computation of the distortion rate characteristics Dk (Rs;k ) for each of the k priority classes. This distortionrate characteristics can easily be obtained from the individual distortion-rate characteristics computed for each of the subbands belonging to that priority class by setting up a similar optimization problem as in (3). There are standard methods of solving constrained optimization problems as in (3) (for example, the BFOS algorithm). s;k k

4 Results and Discussion In this section, we use the performance evaluation technique described earlier in our results and discussion. In our FEC-based schemes, for those dropped packets which are not recovered by the RS code, the missing region is then obtained by using a passive error concealment scheme. In our ECSBC coder, the speci c passive error concealment scheme employed is that of temporal interpolation. We now compare the throttled, unthrottled and unprotected scheme performance. To do so we perform the following two-step experiment. We base the choice of the parameters of the two-state Markov loss model on simulations of a multiplexer with the video sources modeled as a discrete-time autoregressive process [11]. The model parameters matched the bit-rate statistics of the coder [11] operating on the Football sequence. The parameters of the two-state Markov model, namely PL and LL were chosen to match those obtained from the simulation. Further details may be obtained from [7] and [8]. Both the throttled and unthrottled scheme employed the RS(15,13) code at an interleaving depth of 1. Results are illustrated in Fig. 3 in the high-load region as the congestion 31.6

Throttled source coder/FEC Unprotected Unthrottled source coder/FEC

Average received SNR

28.2

25.1

22.4

20.0 38.0

39.0 40.0 41.0 Number of multiplexed sources

42.0

Figure 3: Comparison of the 3 schemes. Multiplexer speed is 100 Mbps and buer size is 500. Source coding rate = 0.80 bits/pixel for the throttled scheme and 0.93 bits/pixel for the unthrottled, unprotected scheme. RS(15,13) code is used at interleaving depth = 1. is normally more severe in such cases. We use only a single priority structure in this example as the primary purpose is to demonstrate the superiority of the throttled approach without having to get into the network priority and buer management issues. Furthermore, as the simulations are time-consuming, they are restricted to a region of maximum interest. The multiplexer operating rate is 100 Mbps (FDDI speed) while the buer size chosen is 500 packets. As the

gure indicates, with increasing load, there is as much as a 6 dB dierence between the throttled and unthrottled scheme and up to a 4 dB dierence between the throttled and unprotected scheme. This gure also con rms that the unthrottled scheme leads to poor performance under heavy loads due to added congestion. In fact, it behaves much worse than the unprotected system. We now employ the code selection strategy described earlier and determine its performance. Figure 4 shows the optimized code rates for a particular example of the Markov channel model. In the example, the loss probability PL is 1  10?2 while Dthreshold = 5 msec and Lthreshold = 10?4 . As mentioned earlier, we do not include the jitter as a constraint to simplify our evaluation. Although 20 msec typically represents the tolerable end-to-end delay for interactive applications [8], a value of 5 msec is chosen for Dthreshold to allow for queueing delays which typically dominate end-to-end delay. We also assume in these examples a CCIR 601 video resolution. Notice that low optimized FEC code rates are obtained when the operating rate is quite small. This means that substantial throttling of the source coding rate, and hence reduced transmission quality, is required to accommodate FEC overheads. Furthermore, observe that the selected code depends on the operating rate. This occurs because the FEC coding delay depends on the operating rate. Hence, a code which satis es the delay threshold at a given operating rate may not do so at a lower operating rate. Consequently, the philosophy of using only one code (which was originally proposed for AAL 1)ndependent of the operating rate is questionable. A better approach would be to pre-determine a number of codes, one for each operating rate for xed channel conditions. Figure 5 demonstrates the performance of dierent codes at various code rates. In this example, a dierent code is used for each priority class optimized independently for that class in accordance with the respective values of the loss parameters. The codelengths illustrated (N = 7; 15; 63) represent the performance when the corresponding code is applied to both priority classes. The optimized code for the high-priority class is RS(61,56) and for the low-priority class is RS(56,50). Though it seems that the code rate which maximizes the performance for the codelength of 63 does better than the optimizing code, it should be noted that the particular code does not satisfy the delay threshold of 5 msec in this particular example. Notice that the performance of the code of codelength 63 is very close to the optimizing code rate as well as the information-theoretic bound on performance. Consequently, as long as the code rate is properly chosen, a codelength of 63 is sucient in providing good performance. This is particularly interesting since it indicates that codes of relatively small codelength (the FEC decoding complexity as well as the net FEC delay introduced depend on the codelength) are sucient to yield good performance. Also shown in the gure is the performance of the RS(128,124) code which has been proposed for the AAL 1 layer [12]. The code optimized according to the selection policy performs much better than the RS(128,124) code. Though the optimized code was selected here for a particular value of the

1.10

Code rate, R=K/N

0.90

0.70

Ideal case (no losses) Independent losses ρLL = 0.1 ρLL = 0.4

0.50

0.30 0.0

0.5 1.0 Average operating rate (bits/pixel)

Figure 4: Typical pro le of optimized code rate (vs.) transmission rate. In this example, the loss probability PL = 1  10?2 and Dthreshold = 5 msec while Lthreshold = 10?4 .

0.50

Normalized mean-squared distortion

0.40

Code length = 7 Code length = 15 Code length = 63 RS(128,124) Optimized code: 20 msec Performance bound Mismatch: ρLL = 0.10, Dthresh = 5 Mismatch: ρLL = 0.10, Dthresh = 20

0.30

0.20

0.10

0.00 0.60

0.70

0.80 Code rate, R=K/N

0.90

1.00

Figure 5: Comparative performance of a hybrid error concealment scheme at dierent code rates. Loss probability PL = 5  10?3 for the high-priority class, 1  10?2 for low-priority class, LL = 0:40, sequence is the Football sequence, Nf = 12 and the overall average transmission rate is 0.85 bits/pixel.

Markov chain parameters, such an accurate description of the loss behavior in the network is seldom available. As a result, it is important to consider the behavior of the optimized codes for slightly dierent values of the Markov chain parameters so that their behavior can be studied under mismatched conditions. This behavior is illustrated in the gure for two cases, one of which was selected with a relatively strict 5 msec constraint on the delay and the other a more relaxed 20 msec constraint. The 5 msec constraint limits the number of available codes that meet the threshold on the loss. As a result, much lower code rates are required to achieve good performance. Even under mismatched conditions, observe that the optimized codes perform very well indicating the robustness of the code selection strategy.

5 Conclusions In this paper, a generic transmission scheme for robust transmission of VBR video employing a class of hybrid FEC-based error recovery schemes was studied. Such schemes are particularly useful when there is high motion or rapid scene changes in the encoded video. In such a case, the eect of error propagation due to imperfect packet recovery is greatly reduced by FEC. In using the active error concealment technique, the performance of two approaches for applying FEC was studied. In the rst approach, additional bandwidth was allocated for the channel coding operation. Such a scheme was shown to suer performance degradation under higher loads in comparison with an unprotected system. A more judicious approach to applying FEC was investigated in this paper where the bandwidth for FEC application was allocated by throttling the source coder output. Under moderate-to-high packet losses (characteristic of high network loads), employing a throttled source-coder FEC application was shown to oer signi cantly better performance compared to an unprotected system. The performance of the scheme is closely related to the code selected. As a result, it is important to devise a clever code selection strategy. A simple strategy of selecting codes based on a constrained optimization technique was outlined and its performance studied. The results indicate that the selection of a single code for all operating rates is questionable. For most cases, codes of smallto-moderate codelength ( 63) performed very well as long as the code rates were properly chosen. The code selection strategy provided robust performance even under conditions of mismatch in choice of the channel parameters for which the codes were selected. These results were subsequently evaluated and veri ed using the MPEG-2 coder. The only noticeable dierence was that the gains while employing MPEG-2 were lower [7] due to the relative robustness of the scheme to loss.

References [1] N. Shachum and P. McKenney, \Packet Recovery in High-Speed Networks using Coding and Buer Management," Proc. IEEE INFOCOM, San Fran., CA, pp. 124{131 , June 1990. [2] H. Ohta and T. Kitami, \A Cell Loss Recovery Method using FEC in ATM Networks," IEEE Trans. Commun., vol. 39, no. 9, pp. 1471{1483, Dec. 1991. [3] E.W. Biersack, \Performance Evaluation of Forward Error Correction in ATM Networks," Proc. ACM SIGCOMM, Baltimore, MD, pp. 248{257, Aug. 1992. [4] A.M. McAuley, \Reliable Broadband Communication using a Burst Erasure Correcting Code," Proc. ACM SIGCOMM , Philadelphia, PA, pp. 297{306, Sept. 1990. [5] Y.H.Kim, \Adaptive Entropy Coded Predictive Vector Quantization of Images," Ph.D Dissertation, Dept. of Electrical, Computer and Systems Engg., Rensselaer Polytechnic Institute, Troy, NY, 1990. [6] Y.H.Kim and J.W.Modestino, \Adaptive Entropy-Coded Subband Coding of Image Sequences," IEEE Trans. Commun., vol. 41, no. 6, pp. 975-987, June 1993. [7] V.Parthasarathy, \Transport Coding Schemes for Digital Video Transmission over ATM Networks," Ph.D Dissertation, Dept. of Electrical, Computer and Systems Engg., Rensselaer Polytechnic Institute, Troy, NY, June 1995. [8] V. Parthasarathy, J. W. Modestino and K. S. Vastola, \Reliable Transmission of High-Quality Video over ATM Networks," submitted to IEEE Trans. Image Processing. [9] V.Parthasarathy, J.W.Modestino and K.S.Vastola, \Design of a Transport Coding Scheme for Variable-rate Video Transmission over ATM Networks," accepted for publication to IEEE Trans. Systems and Circuits for Video Tech. [10] R.J.McEliece and W.E.Stark, \Channels with Block Interference," IEEE Trans. on Inform. Theory, Vol. IT-30, pp. 44-53, Jan 1984. [11] B.Maglaris, D.Anastassiou, P.Sen, G.Karlsson, J.D.Robbins, \Performance Models of Statistical Multiplexing in Packet Video Communications," IEEE Trans. Commun., Vol. 36, No. 7, pp.834-844, July 1988. [12] C.Partridge, Gigabit Networking, Addison Wesley Publishing Company, 1993.