IEEE TRANSACTIONS ON NETWORKING, VOL. 10, NO.

2,202 _ 208, JUNE 2011

A Model-Based Approach to Evaluation of the Efficacy of FEC Coding in Combating Network Packet Losses
VENKATA GIRI.MADDIPATI(*) GIET,RAJAHMUNDRY DEPT.OF.COMPUTERSCINCE &ENGG, maddipati.venkatagiri@gmail.com RAJASEKHAR.SWARNA GIET,RAJAHMUNDRY DEPT.OF.COMPUTERSCINCE &ENGG, swarnarajasekhar@gmail.com

AbstractWe propose a model-based analytic approach for evaluating the overall efficacy of FEC coding combined with interleaving in combating packet losses in IP networks. In particular, by modeling the network path in terms of a single bottleneck node, described as a queue, we develop a recursive procedure for the exact evaluation of the packet-loss statistics for general arrival processes, based on the framework originally introduced by Cidon et al., 1993. To include the effects of interleaving, we incorporate a discretetime Markov chain (DTMC) into our analytic framework. We study both single-session and multiple-session scenarios, and provide a simple algorithm for the more complicated multiple-session scenario. We show that the unified approach provides an integrated framework for exploring the tradeoffs between the key coding parameters; specifically, interleaving depths, channel coding rates and block lengths. The approach facilitates the selection of optimal coding strategies for different multimedia applications with various user quality-ofservice (QoS) requirements and system constraints. We also provide an information-theoretic bound on the performance achievable with FEC coding in IP networks. Index TermsAutocorrelation function, FEC coding, interleaving, packet-loss processes, residual packet-loss rates, single-multiplexer model.

I. INTRODUCTION THE packet transport service provided by representative packet-switched networks, including IP networks, is not reliable and the quality-of-service (QoS) cannot be guaranteed. Packets may be lost due to buffer overflow in switching nodes, be discarded due to excessive bit errors and failure to pass the cyclic redundancy check (CRC) at the link layer, or be discarded by network control mechanisms as a response to congestion somewhere in the network. Forward error correction (FEC) coding has often been proposed for end-to-end recovery from such packet losses. However, the use of FEC in this application provides a double-edged sword. From an end users perspective, FEC can help recover the lost packets in a timely fashion through the use of redundant packets, and generally adding more redundancy can be expected to improve performance provided this added redundancy does not adversely affect the network packet loss characteristics. On the other hand, from the networks perspective, the widespread use of FEC schemes by end nodes will increase the raw packet-loss rate in a network

because of the additional loads resulting from transmission of redundant packets. Therefore, in order to optimize the end-to-end performance, the appropriate tradeoff, in terms of the amount of redundancy added, and its effect on network packetloss processes, needs to be investigated under specific and realistic modeling assumptions. In this paper,we provide a study of the overall effectiveness of packet-level FEC coding, employing interlaced Reed-Solomon codes, in combating network packet losses and provide an information- theoretic methodology for determining the optimum compromise between end-to-end performance and the associated increase in raw packet loss rates using a realistic model-based analytic approach. Intuitively, for a given choice of block length we expect that there is an optimum choice of redundancy, or channel coding rate, since a rate too high (low redundancy) is simply not powerful enough to effectively recover packet losses while a rate too low (high redundancy) results in excessive raw packet losses due to the increased overhead which overwhelms the packet recovery

capabilities of the FEC code. The optimum channel coding rate results in an optimum compromise between these two effects. In order to analytically investigate this tradeoff, we consider a simplified network scenario described in terms of a single bottleneck node, modeled as a multiplexer. For networks with reliable transmission media, where transmission errors are negligible, buffer overflows due to congestion in routers are the major cause of packet losses. In a packet-switched network, a flow of packets crosses a chain of routers before it reaches the destination node. Most of the packet losses from a flow occur in the router which has the smallest bandwidth. Therefore, we can model the whole chain of routers in terms of this single bottleneck node. Both theoretical analysis and experimental results justify this assumption [1][3].Asingle-multiplexer model for this bottleneck node is widely used to analyze the associated queueing-related packet losses, e.g., losses due to buffer overflows and excessive delays. Since the correlation level of the packet-loss process has great impact on the FEC efficacy, we investigate this dependence using the autocorrelation function of the packet-loss process. The performance of FEC in recovering network packet losses has been studied in many papers [4][13]. In [4], the use of redundant parity packets was proposed to reconstruct lost data packets and the corresponding performance evaluation indicated that residual packet-loss rates can be reduced up to three orders of magnitude. However, in [4] for analysis purposes the packet-loss process resulting from the single-multiplexer model was assumed to be independent and, consequently, the simulation results provided show that this simplified analysis considerablly overestimates the performance of FEC. By modeling the single-multiplexer as an or queue, Cidon et al. [5] proposed a recursive algorithm to compute the packet-loss statistics (block errror density), through which the exact residual packet-loss rate after decoding was computed. Surprisingly, all numerical results given in [5] indicate that the resulting residual packet-loss rates with coding are always greater than without coding, i.e., FEC is ineffective in this application. However, in [5] only a single parity packet is used1 and the block length was constrained to the range . As we show later, these conclusions are somewhat misleading and result from inappropriate parameter choices. In [6][8], more general arrival processes were considered and coding tradeoffs assessed but performance results were obtained using large deviation bounds to characterize the packet-loss processes and were not exact. In [9] [12],

Altman et al. derived analytical formulas for the frame-loss probabilities for the single-multiplexer model using multi-dimensional probability generating functions and show that the frame loss probabilities can be reduced if a sufficiently large amount of redundancy is added. However, contrary to [4], [5], in these works the authors used the frame-loss probability as the evaluation metric for FEC performance, because it was assumed that the failure to recover any lost data packet will lead to the loss of all the data packets in that block. For ATM networks, this assumption is valid since the loss of a single cell does result in the discarding of the whole message. However, for other networks, like IP networks, this is not the case. In IP networks, packet-level FEC coding can be performed across several IP packets. Even if any lost data packets cannot be recovered, the correctly received data packets in the same coding block may still be useful. For example, in the application of video over IP, when some lost video packets cannot be recovered, the correctly received video packets in the same coding block need not be discarded and can even be used to estimate the information in lost video packets using an appropriate error concealment scheme. As shown in Section II, in many cases, the FEC performance predicted by using the frame-loss probability is not only quantitatively different, but also qualitatively different, from that reflected by the packet-loss probability. Although the asymptotic analysis (when block size goes to infinity) for the frame-loss rates in [9][12] can provide some insight into FEC performance on IP networks, the methodology and specific conclusions developed in these works are not useful for a comprehensive evaluation of FEC effectiveness on IP networks since they are based on an inappropriate evaluation metric. II. PRELIMINARIES A. Single-Multiplexer Network Model If a networks performance is limited by a single bottleneck node, the network can often be modeled in terms of a single multiplexer. As illustrated in Fig. 1, the single-multiplexer model is a queueing system which consists of three components: 1) an arrival process for packets from different sources with corresponding packet arrival rates , ; 2) a buffer which can hold up to packets, which are assumed served in first-come-first-served (FCFS) order; and 3) an output link with average packet service rate . We assume if a packet finds a full buffer upon arrival, it will be discarded. For analytical convenience, we assume the packet service times are independent and identically distributed (i.i.d.) with an exponential distribution and average packet service time .With denoting the overall packet arrival rate, the normalized load to the system is (1) Although the assumption of exponential service time may not be

accurate in some situations, it enables us to obtain some analytical results and gain some insight on the effects of packet-loss processes due to buffer overflows.

can represent constant bit rate (CBR) sources. For the hyperexponential distribution, the average arrival rate C. System Model Evaluation for FEC Performance

B. Source Model

Consider the communication system model illustrated in Fig. 2. We suppose there are homogeneous and independent sources sharing the single-multiplexer and each source generates packets with average rate . The FEC coder for each source applies an interlaced Reed-Solomon code [6], [8], [15] to the packets from the source, which means for every block of source information packets it creates an additional parity packets to the network. The channel coding rate is given by . As a result of the channel coding, the packet arrival rate into the network will increase to . Let the random variable denote the number of lost packets within a block. If , we assume all the lost packets within that block can be recovered by the channel decoder. Assume denotes the block-error distribution, i.e., the probability that packets out of are lost. Therefore, the expected number of lost packets within a block is (6) and the expected number of lost information packets within a block is (7) Finally, the effective information packet loss rate after channel decoding is (8) D. Evaluation Metric: Packet-Loss Probability or Frame-Loss Probability? In some networks, such as ATM networks, the failure to recover a single packet results in the loss of the entire frame (block). In this case, the frame-loss probability is a more suitable metric than packet-loss probability for evaluation of FEC performance, as used in [9][12]. The frame-loss probability is given by (9) Figs. 3 and 4 illustrate comparisons of the FEC performance predicted by packet-loss probability with that predicted by

We assume the packet arrival process for each source is a renewal process, i.e., for any source the packet interarrival times are i.i.d. with arbitrary probability density function . In particular, we consider the case of the Erlang interarrival time distribution When (assuming is fixed), the variance of converges to 0 and the interarrival time becomes deterministic with period . With appropriate parameter choice, this case

provided the block length is chosen large enough. As described above, in this paper we focus on IP networks and use packet-loss probability to evaluate the FEC performance. E. Autocorrelation Processes Function of Packet-Loss

Fig. 3. Performance difference between coding and without coding as a function of the number of information packets in each frame, single session Other system parameters are as follows: system load buffer size , 1 parity packet used

For a packet-loss process we use the autocorrelation function to characterize the dependence between the packet-loss events over time. Let the random sequence represent the packet- loss process, with 1 denoting loss and 0 denoting reception. If is stationary, then the autocorrelation function of is given as (10) where is the lag and is the expectation of the sequence III. FEC PERFORMANCE WITH A SINGLE SOURCE A. FEC Without Interleaving We begin our analysis with the simplest case: there is only one user for the multiplexer . As (8) illustrates, the key quantity in evaluating the residual packet-loss rate after FEC decoding is , the blockerror distribution for an arbitrary number of consecutive packets. In [5], Cidon et al. propose a recursive algorithm to compute for the finite buffer queue with Poisson arrivals and exponential service times, denoted as the queue. In order to analyze the packet losses for more general arrival patterns, in what follows we first describe the extension of the algorithm to the queue, i.e., the finite buffer queue with general i.i.d. interarrival times and exponential service times. 1) Analysis of Block-Error Distribution: Suppose there is only one source sharing the multiplexer , and the . packet interarrival times are i.i.d. with arbitrary probability density function . Since the packet service times are assumed i.i.d. and exponentially distributed, the singlemultiplexer can be modeled as a standard queueing system. Let be the number of packets in the buffer just before the th packet arrives at the system. Because of the memoryless property of the exponential service time, is a discrete-time Markov chain (DTMC) [16, pg. 249], illustrated in Fig. 5, with the state space and state-transition matrix

Fig. 4. Performance difference between coding and without coding as a function of the number of information packets in each frame, multiple session _ __. Other system parameters are as follows: system load _ _ ___, buffer size _ ___, 1 parity packet used. frame-loss probability. The curves show the differences of the residual packet-loss probabilities and frame-loss probabilities with coding and without coding, denoted by and , respectively. Fig. 3 shows the case of a single session and Fig. 4 shows the case of multiple sessions . Other system parameters are indicated in the corresponding figure captions. When one of the differences is positive (i.e., or ), this implies that using redundant packets does not improve the corresponding performance metric ( or ); that is, use of FEC is ineffective. From Figs. 3 and 4 it should be clear, at least for the cases considered, these two performance metrics can lead to totally different conclusions, both quantitatively and qualitatively, concerning the effectiveness of FEC. In particular, in either case the results indicate that, for the corresponding choice of system parameters, the use of FEC is ineffective for all block lengths based on but can provide improved performance based on

into an matrix row-wise and then read-out from the matrix column wise. Therefore, the originally consecutive packets will be packets apart from each other after interleaving. is called the interleaving depth (ILDP). At the receiver the packets will be reordered in the deinterleaver before decoding. In what follows we analyze the corresponding blockerror distribution incorporating the interleaving operation. 2) Analysis of Block-Error Distribution: As shown in Fig. 12, we again assume the packet arrival process at the single multiplexer (after interleaving) is a renewal process, i.e., the packet interarrival time at the multiplexer is i.i.d. with some common probability density function . Let denote the number of packets in the buffer just before the th packet arrives at the multiplexer (Point in Fig. 12). Again, we assume the packet service times are i.i.d. with an exponential distribution and the average service rate is . Therefore, the random sequence is a discrete-time Markov chain with the state space , transitionprobability matrix and stationary probability as expressed in (11) and (13), respectively. Define another random sequence , where is the number of packets in the buffer just before the th packet, seen by the interleaver (Point in Fig. 12), arrives at the singlemultiplexer. Since after interleaving the originally consecutive packets will be packets apart from each other, is formed by selecting every th element of . Therefore,is also a discrete-time Markov chain with the state space , and the corresponding transitionprobability matrix , is the step transition matrix of , (19) Let be the stationary distribution of , which can be obtained as the solution to (20) subject to a normalization constraint. Then, proceeding with the same procedure as in Section III-A1, we can obtain recursive expressions for computing . If an ideal interleaver is applied, then the packet-loss process is independent. Suppose the average packet-loss rate is , then the packet-loss statistics with ideal interleaving are given by the binomial distribution (21) Once we obtain the block-error distribution incorporating the interleaving operation, the effective packet-loss rate after FEC decoding can be obtained from (8). In Appendix A, we determine the autocorrelation function for packet-loss processes associated with the single-multiplexer model (modeled as a queue) with interleaving. Next we show some numerical examples. 3) Numerical Examples: Fig. 14 illustrates the efficacy of interleaving in reducing packet-loss correlation associated with the single-multiplexer model. In particular, we illustrate the packet-loss autocorrelation function as a functionof lag for the case in which the arrivals are Poisson with the average loadand the buffer size is . It shows, as expected,that interleaving can effectively reduce the correlation of the packet-lo ss process, which means

where , , is the probability that packets are served out of the system during an arbitrary interarrival time, provided that there are at least packets in the system at the beginning of that interarrival time. It can be shown [16, pg. 248]n that (12) subject to a normalization constraint. Let be the stationary distribution of the DTMC, which can be obtained as the solution to (13) Define , , , , to be the probability of losses in a block of packets, given that there are packets in the system just before the first packet of the block arrives to the system. So we have (14) where is determined from (13). B. FEC With Block Interleaving FEC performance is often limited by the bursty nature of typical packet-loss processes, and block interleaving techniques are frequently used to reduce the burstiness of the packet-loss processes in networks [4], thereby improving FEC performance. In this section, we analyze the efficacy of interleaving in reducing the burstiness of network packet-loss processes and in improving the FEC performance. 1) Interleaving Operation: The operation of block interleaving is illustrated in Figs. 12 and 13. Before being transmitted into the network, packets are filled

that interleaving can make the packet losses more independent and isolated. POTENTIAL OF FEC AND AN INFORMATIONTHEORETIC BOUND In Fig. 17, we demonstrated that, with the same packet-loss rate requirement, FEC coding with a larger block size can support increased source traffic. However, the source traffic that can be supported is not unlimited because of the channel capacity limitation imposed by the single-multiplexer transport channel model. In what follows we develop an information-theoretic upper bound on the FEC performance based on the single-multiplexer network model. In this section, we only consider the case of a single source , although the approach can be extended to arbitrary . A. Channel Model for Packet Transmission Over Networks Consider a channel model for packet transmission over a general packet-switched network. Assume a packet has bits. It is either transmitted and received by the receiver, or is lost due to network congestion or buffer overflow. For a received packet, bit errors may be introduced. Then packet transmission over networks can be modeled for coding purpose in terms of serial bit-by-bit transmission of -bit symbols either over a binary symmetrical channel (BSC) with crossover probability Fig. 23. Component channels of BIC
corresponding to packet delivery and loss. Fig. 24. Simplified communication system model. (state 0) or over a binary

independent state transitions, as ill ustrated in Fig. 24. Here we consider only the packet losses caused by the buffer overflows, and assume no bit errors, i.e., the BSC crossover probability . Let be the packetloss rate of thesingle-multiplexer, so . Then, from (27), the capacity of the BIC is given by (29) Assume the source creates packets at rate and the packetservice rate is . Then the normalized system load before coding is . The channel encoder applies channel coding (notnecessarily RS codes) with coding rate to thesource traffic. CONCLUSIONS AND FUTURE WORK We have analyzed the efficacy of FEC in combating network packet losses based on a single-multiplexer network model and demonstrated that FEC has great potential in recovering the packet losses caused by congestion at a bottleneck node of a packet-switched network, provided that the coding rate and other coding parameters are appropriately chosen. We developed a discrete-time Markov chain model to analyze the efficacy of interleaving in improving the FEC performance and determined how much interleaving depth is required for FEC to approach the optimum performance. We derived an upper bound on the end-to-end performance using FEC based on an information-theoretic methodology, which is useful in predicting source rates that can be supported with arbitrarily high reliability. Despite the great potential of FEC coding in recovering network packet losses, the implementation complexity of FEC coding and the corresponding coding/decoding delay also need to be considered, which is an issue particularly important for real-time applications. One objective for future work is the analysis of the additional delay caused by the FEC coding, perhaps combined with interleaving/deinterleaving. Likewise, the application of FEC for network transport is limited by the timevarying and often uncertain error characteristics of the channel, which makes the appropriate choice of FEC coding rate difficult to determine. In real world applications, FEC cod ers are required which can adapt the channel code rate to the time-varying channel conditions. This issue is also a topic for future work. REFERENCES [1] O. J. Boxma, Sojourn times in cyclic queues The influence of the slowest server, in Proc. 2nd Int. MCPR Workshop on Computer Performanceand Reliability, Rome, Italy, May 1988. YU et al.: A MODEL-BASED APPROACH TO EVALUATION

erasure channel (BEC) (state 1), both of which are illustrated in Fig. 23, where is used to indicate the erasure symbol.A lost packet corresponds to the entire codeword symbol of bits being erased, while a received packet means each of the bits is sequentially transmitted over the BSC. This channel model belongs to the class of block interference channels (BIC), introduced by McEliece and Stark [14]. Let represent the state space of the BIC. If the state transitions are independent, then the Shannon capacity of the BIC is given as [14], (26) where is the capacity of the component channel , and the expectation is over the state space . It follows that(27) where is the probability of being in the loss state and isthe binary entropy function, (28) B. Information-Theoretic Performance Bound on FEC

Referring to Fig. 12, suppose the interleaving is ideal, and consequently the packet-loss process seen by the channel decoder is independent. If we consider the interleaver and the deinterleaver as components of the coding channel, then the channel, consisting of the interleaver, the single-multiplexer and the deinterleaver, can be modeled as a BICwith

OF THE EFFICACY OF FEC CODING IN COMBATING NETWORK PACKET LOSSES 641 [2] D. Y. Eun and N. B. Shroff, Network decomposition: Theory and practice, IEEE/ACM Trans. Networking, vol. 13, no. 3, pp. 526539, Jun. 2005. [3] J. Bolot, End-to-end delay and loss behavior in the Internet, in Proc. ACM SIGCOMM 1993, San Francisco, CA, Sep. 1993, pp. 289298. [4] N. Shacham and P. Mckenney, Packet recovery in high-speed networks using coding and buffer management, in Proc. IEEE INFOCOM 1990, San Francisco, CA, Jun. 1990, vol. 1, pp. 124131. [5] I. Cidon, A. Khamisy, and M. Sidi, Analysis of packet loss processes in high-speed networks, IEEE Trans. Inf. Theory, vol. 39, no. 1, pp. 98108, Jan. 1993. [6] R. Kurceren, Joint source-channel coding approach to transport of digital video on lossy networks, Ph.D. dissertation, Rensselaer Polytechnic Inst., Troy, NY, May 2001. [7] R. Kurceren and J. W. Modestino, Optimum FEC coding rate allocation for video transport over ATM networks, in Proc. IEEE Int. Symp. Information Theory (ISIT 1998), Cambridge, MA, Aug. 1998, p. 251. [8] R. Kurceren and J. W. Modestino, joint A source-channel coding approach to scalable delivery of digital video over ATM networks, in Proc. Int. Conf. Image Processing (ICIP 2000), Vancouver, Canada, Sep. 2000, vol. 1, pp. 371374. [9] E. Altman and A. Jean-Marie, Loss probabilities for messages with redundant packets feeding a finite buffer, IEEE J. Sel. Areas Commun., vol. 5, no. 6, pp. 778787, Jun. 1998. [10] O. Ait-Hellal, E. Altman, A. Jean-Marie, and I. A. Kurkova, loss probabilities in presence of On redundant packets and several traffic sources, Perform. Eval., vol. 3637, pp. 485518, 1999. [11] P. Dube, O. Ait-Hellal, and E. Altman, loss On probabilities in presence of redundant packets with random drop, Perform. Eval., vol. 53, pp. 147167, 2003. 12] A. Jean-Marie, P. Dube, D. Artiges, and E. Altman, Decreasing lossprobabilities by redundancy and interleaving: A queueing analysis, in Proc. ITC 18, Berlin, Germany, Sep. 2003. [13] Y. Shen, P. C. Cosman, and L. B. Milstein, Error resilient video communications over CDMAnetworks with a bandwidth constraint, IEEE Trans. Image Process., vol. 15, no. 11, pp. 3241 3252, Nov. 2006. [14] R. McEliece and W. Stark, Channels with block interference, IEEETrans. Inf. Theory, vol. 30, pp. 4453, Jan. 1984. [15] V. Parthasarathy, J. W. Modestino, and K. S. Vasto la, Reliable transmissions of high-quality video over ATM networks, IEEE Trans. Image Process., vol. 8, no. 3, pp. 361374, Mar. 1999.

[16] D. Gross and C. M. Harris, Fundamentals of Queueing Theory, 3rd ed. New York: Wiley, 1998. [17] M. Mushkin and I. Bar-David, Capacity and coding for the Gilbert-Elliott channels, IEEE Trans. Inf. Theory, vol. 35, no. 6, pp. 12771290, Nov. 1989.