A Novel Scheme of Loss Differentiation and Adaptive Segmentation to Enhance TCP Performance over Wireless Networks

Transcription

1 A Novel Scheme of Loss Differentiation and Adaptive Segmentation to Enhance Performance over Wireless Networks Marco Mancuso Politecnico di Milano, Dept. of Electronics and Information, P.zza L. Da Vinci 32, Milano, ITALY Abstract was designed to cope with packet loss due to congestion. Thus, it performs poorly in lossy environments, such as heterogeneous networks including wireless links. Standard is unable to differentiate loss due to packet corruption from that due to network congestion. In this paper, a sender-side method of end-to-end loss differentiation and adaptive segmentation (Robin) is proposed, for enhancing performance in heterogeneous networks. The loss differentiation algorithm enables the sender to distinguish congestion losses from corruption losses. Moreover, the segmentation algorithm, according to the perceived congestion level of the network, improves the error recovery phase during a non-congestive period. The Robin scheme has been added to -Westwood, a wellknown variant recently proposed for wireless networks. Based on simulations of networks including a wireless link (Bluetooth), we show that Robin yields higher throughput in presence of errors on the radio link, whilst being friendly with other concurrent connections in wired networks. In addition, since Robin involves changes only at the sender side without requiring changes of the receiver protocol stack or interception at intermediate nodes, it could be easily adopted in current Internet. Index Terms Bluetooth, Internet, mobile computing, protocols,, wireless networks. I. INTRODUCTION he Transmission Control Protocol () is a popular TInternet protocol for reliable data delivery, which adapts very well to disparate environments. However, it was designed optimizing its performance to cope with packet loss due to network congestion: thus, it performs poorly in lossy environments. The assumption that packet loss is an indicator of network congestion may not apply to heterogeneous networks, such as those including wireless links, in which packet loss may be induced by noise or any other reason than congestion [1]. Packet loss due to bit corruption can be misinterpreted, triggering unnecessarily congestion control actions. In order to integrate wireless networks in the Internet, numerous variants have been proposed to improve its performance in heterogeneous networks. Some of them hide the lossy parts of network [2] or improve link transmission qual- sender Fixed Host network Base Station wireless link receiver Mobile Host Fig.1: connection over heterogeneous network with wireless link. ity [3] or are based on explicit loss notification [4][5]. Alternatively, end-to-end approaches consist of extending basic Reno [6] to cope with specific problems of wireless networks: e.g., Westwood (W) [7], Eifel [8], SACK [9][10]. In this paper, we propose an end-to-end loss differentiating criterion together with a novel segmentation algorithm, suitable for enhancing performance in heterogeneous networks including wireless links, as shown in Fig. 1: the Robin scheme. It requires changes only to the sender and is thus interoperable with existing. Robin is particularly useful when incorporating -awareness in intermediate nodes is not wanted or possible, e.g. when IP payload is encrypted. This paper is organized as follows. Sec. II summarizes some background works. In Sec. III, we present the Robin scheme and outline the rationale of the loss differentiation and adaptive segmentation algorithms proposed. The simulation scenario is described in Sec. IV. Sec. V presents simulations results for Robin performance evaluation. Sec. VI draws some conclusions and outlines topics left for future work. II. BACKGROUND WORK In Reno, the Data Sending Rate (DSR) is controlled by means of a congestion window (cwnd), which is the maximum number of outstanding packets that can be transmitted before receiving acknowledgements (ACKs) from the data receiver. Reno interprets each packet loss as congestion indication, thus reducing drastically (and needlessly) the cwnd size over disturbed wireless links. Westwood [7] is a variant based on estimating run-time, at the sender side, the actual available bandwidth by monitoring the ACK reception rate. This estimated connection rate is then used to compute the congestion window cwnd and the slow-start threshold ssthresh after each loss episode. This way, W avoids overly conservative reduction of cwnd and ssthresh and thus ensures a faster recovery. On the other hand, Vegas [11] is based on proactive congestion detection. The expected throughput is defined as Expected = cwnd/rtt min, where RTT min is the minimum of all measured Round Trip Times (RTTs), while the actual throughput is defined as Actual = cwnd/rtt, where RTT is the last measured RTT. Vegas attempts to detect the congestion status by monitoring their difference, that is diff=expected-actual. Then, it increments proactively cwnd when diff RTT min <α (typically, α=1 Sender Maximum Seg-

2 ment Size, SMSS), decrements it when diff RTT min >β (typically, β=3 SMSS), or does not change it otherwise. Based on such congestion avoidance approach, loss predictors have been defined in [12][13]. The loss predictor f VEGAS = diff RTT min guesses if next packet transmitted will be lost due to congestion or not. Unfortunately, this technique has not proven very useful in practice, because the Actual throughput computed as above, especially in asymmetric networks, is not the forward data rate, but rather the ACK rate [14]. In this work, thus, we aimed first at discriminating congestion more effectively. III. THE ROBIN METHOD The loss predictor f VEGAS provides a simple means to differentiate congestive and non-congestive states. In addition, we adopted the bandwidth estimation technique of W to estimate the actual Data Receiving Rate (DRR), thus yielding a more sophisticated heuristic method. In summary, the Robin scheme is composed of three algorithms: a heuristic criterion to distinguish between congestive and non-congestive states; a novel Adaptive Segmentation technique that enables the sender, if a non-congestive packet loss is detected, to retransmit smaller packets, having aggregate payload equal to the payload of the lost packet; a modified Additive Increase (AI) algorithm, performing adaptive window opening based on the perceived state. A. Distinguishing Congestive and Non-Congestive States In Robin, blind reduction of send window as in Reno is avoided. Referring to Fig. 2, the basic idea is to perform continuously sender-side independent estimates of the average flight-size length L and of the average sending time interval T. Following the Double Filtering Technique (DFT) approach, proposed in [15], this is done by measuring and low-pass filtering packet lengths and intervals between sending times. Hence, the average DSR can be estimated as DSR = L T (1). On the reverse path, the DRR is estimated by measuring the rate of received ACKs, as in Westwood [7]. DRR represents the Actual rate allowed by the network. Note that the bidirectional flow of a connection obeys a conservation principle [16], stating that in a steady-state connection the DRR tracks the DSR. Nevertheless, when the offered load is large enough, packets start queuing at some bottleneck node and increasing further the DSR results in a small increase of DRR. Now, we define the threshold th_robin as the ratio of current available sending window size (swnd flightsize) to RTT min. It represents the extra rate allowed in the short term by the connection. Thus, similarly to Vegas, we can assume Expected=DSR+th_robin. The difference Expected-Actual follows the bottleneck queue trend. Thus, Robin monitors the state differentiation function, defined as sender data segments forward path DSR FILTER FILTER DRR ACKs reverse path network data segments ACKs receiver Fig.2: DSR and DRR estimation in a connection by the Robin method. frobin = [( DSR + th _ robin) DRR] RTTmin (2) that somehow represents the data backlog in the network [11]. After each packet loss, Robin assumes that the connection is most likely in non-congestive state if f ROBIN <0. Otherwise, Robin assumes conservatively that packet losses are induced by congestion [14]. B. Adaptive Segmentation Decreasing segment size reduces Packet Error Rate (PER) [17]. Therefore, similarly to what proposed in [18], adaptive segmentation is enabled after a packet loss if non-congestive state is detected. The following rules are applied: when packet loss is detected by fast retransmit if (f_robin < 0) /* random loss most likely due to bit errors */ resend missing packet in smaller segments; else /* congestive loss is most likely */ classic Fast Retransmit action is performed; when packet loss is detected by retransmit timeout timer if (f_robin < 0) /* random loss most likely due to bit errors */ timeout action is performed with new segmentation; else /* congestive loss is most likely */ classic timeout action is performed; After each packet loss, if non-congestive state is detected, Robin source reduces its SMSS and retransmits smaller segments, having aggregate payload equal to the payload of the lost packet. The resulting network load is the same of a classic connection, except for higher segmentation overhead. The SMSS default size is restored when the retransmission is completed. Further segmentation is not allowed if the overhead/data ratio exceeds 30%. The main observation, as pointed out for Smart Framing [18], is that congestion control is only marginally driven by the rate at which bytes leave the source. Rather, it is driven by the rate at which segments (and their respective ACKs) are sent (received) at the source. Note that, even if non-congestive state is detected wrongly, reducing packet size does not impair data transfer performance. Buffers in modern routers are partitioned in variablesize chunks; thus, smaller-size packets experience lower dropping probability.

3 Adaptive segmentation introduces a few advantages. When BER is high, sending smaller packets is better, because if one packet is lost, classic resends it as a whole, while Robin resends only one smaller-size segment. Robin segmentation algorithm helps the sender to receive enough ACKs and possibly to detect loss by Fast Retransmit (FR). FR is a much better way to detect packet loss, especially when latency is the major performance indicator. The increased number of ACKs on the reverse path refines DRR estimation. Robin sender transmits the same load to the network as a classic sender. With no random loss, it mimics behavior in terms of window growth, congestion control, aggressiveness and friendliness. When Delayed-ACKs are used and the connection is trapped in non-congestive state, the receiver has not to wait for 200 ms, as in standard, before generating an ACK. C. The Adaptive Increase Algorithm Robin refines also the AI phase of basic. As stated in Sec. III.A, Robin assumes that, when f ROBIN <0, the available bandwidth is under-utilized (non-congestive state). In addition, if -α<f ROBIN <0, the connection is inferred to approach full utilization (we chose α=3, expressed in current [SMSS] units, in case reduced due to adaptive segmentation). In this latter case, the window growth rate is decreased to defer early coming of self-induced congestion losses, as in Reno, forcing connection to stay longer in the equilibrium region where bandwidth is fully utilized. Hence, the rule: if ( (f_robin > -ALPHA) && (f_robin < 0) ); /* connection is approaching full utilization */ set reduced cwnd growth factor; Adaptive increase algorithm can enhance performance significantly, yielding better utilization of bottleneck link and reduced congestion loss. IV. SIMULATION SCENARIO We evaluated the performance of the Robin scheme using the simulation tool ns-2 (version 2.1b8a) [19]. Moreover, we used the public ns code of W [20], appropriately modified according to our Robin scheme. The simulation scenario is outlined in Fig. 3. A connection is established from the fixed host (FTP source) to the mobile host (FTP sink), linked to routers R1 and R2 respectively. This connection shares the link R1-R2 with background traffic generated by four ON-OFF traffic generator objects Traffic/Pareto [19] (shape parameter = 1.5, mean idle period = 1 s, mean busy period = 2 s) and transported by protocol. We denote the congestion loss rate of the connection as R c, i.e. fraction of lost packets in the R1 bottleneck router. For each set of network parameters, the rate of the crosstraffic sources, during the ON time, is adjusted to produce the desired value of R c, in the range from 300 kbit/s to 3 Mbit/s. fixed host FTP source On-Off source #1 On-Off source #N R1 R2 wireless link mobile host FTP sink On-Off sink #1 On-Off sink #N Fig. 3: Network model simulated for performance evaluation. As far as the wireless link is concerned, we chose to simulate a Bluetooth radio link [21], modeled as discrete two-state Markov chain. The radio link was assumed to alternate between two states, Good or Bad, with mean periods ms and 55.8 ms, respectively [22]. In the Bad state, we varied the PER in the range 0% to 10%. Moreover, the propagation delay from R2 to the mobile host was set to 1 ms (all propagation delays do not include transmission and queuing delays). All buffers in the network are Drop-Tail buffers: i.e., when the buffer is full, incoming packets are dropped. Router R1 has an output queue (towards R2) whose size is limited to Q packets (packets of length 1 SMSS, i.e. the default value before possible reduction due to adaptive segmentation) plus headers. All other queues at the two routers are unlimited. In each experiment, the fixed host performs a bulk data transfer to the mobile host over a 600 s time interval. To achieve good confidence of simulation results, 100 independent experiments ( data transfers) were accomplished in all cases studied and results were averaged. V. SIMULATION RESULTS AND PERFORMANCE EVALUATION The performance achieved by Reno, W and W enhanced with Robin, both in lossy and non-lossy networks, was evaluated by simulation varying the following parameters: R c, the PER, the bandwidth of wired links B, the bandwidth of the wireless link B w, the SMSS default value and the round-trip propagation time RTT p. Only a small selection of simulation results is presented in this paper, to compare the performance achieved by the three protocols. A. Impact of the Loss Rate in the Wireless Link First, we assessed the effectiveness of the heuristic method proposed on treating appropriately packet losses due to congestion or transmission errors, by studying the impact of the wireless channel PER on the average throughput. In all simulations reported in this section, parameters were set as follows: Q=20 packets, B w =2 Mbit/s, B=10 Mbit/s, RTT p =120 ms, SMSS=1400 bytes and R c =4%. Thus, Fig. 4 plots the average throughput of connections in the network of Fig. 3, for different PER values in the Bad state. The throughput of W+Robin resulted always higher than that of Reno and of plain W, for all random loss probabilities. This is mainly due to the adaptive segmentation algorithm, reducing timeout episodes, and to the Adaptive Increase algorithm, adjusting window growth to defer congestion episodes. These two algorithms appear to be prop-

4 erly triggered according to the value of the state differentiation function f ROBIN. Moreover, the refined error recovery mechanism of Robin enhances W: for large PER values, the higher number of segments increases the number of ACKs on the reverse path and then refines DRR estimation. To clarify the role of Adaptive Increase compared to standard Additive Increase, Fig. 5 compares the congestion window evolution in both cases, W+Robin versus Reno, measured in the same simulation run for PER=1%. By increasing the number of returned ACKs in presence of random errors, a refinement of bandwidth estimation and hence a reduction of the number of timeouts are achieved. Less timeouts indicate that W+Robin save the wasting of idle time, improving bandwidth utilization. In conclusion, our simulation experiments confirm that our heuristic method attains improved performance in lossy environments. Being different from trying to explicitly diagnose the reason of loss of each packet, as addressed in [12] and [13], Robin aims at probing the connection state and performing appropriate reaction. B. Impact of Propagation Delay on Throughput The impact of the RTT p on the average throughput is shown in Fig. 6. PER is set to 1%, bottleneck link transmission speed to B w =2 Mbit/s, B=10 Mbit/s, Q=20 packets, SMSS=1400 bytes and R c =2%. From this graph, we notice that Robin improves throughput performance of W only when RTT p is not very large. In other words, for high propagation delays, the gain achieved over Reno is mostly due to the mechanism of W and further improvement yielded by Robin is negligible. Average Throughput [Mbit/s] 2,0 1,8 1,6 1,4 1,2 1,0 0,8 0,6 0,4 0,2 0,0 W+Robin W Reno 0% 1% 2% 3% 4% 5% 6% 7% 8% 9% 10% Packet Error Rate [%] in Bad state Fig. 4: Average throughput for different PER values on the wireless channel. cwnd [SMSS] W+Robin Reno time [s] Average Throughput [Mbit/s] 2,0 1,5 1,0 0,5 0,0 W+Robin W Reno RTTp [ms] Fig. 6: Average throughput for different values of the round-trip propagation delay RTT p. C. Robin Friendliness Throughput and timeout episodes are not the only metrics by which a transport protocol is evaluated. Friendliness is another important aspect: a new protocol should coexist with standard protocols, while providing opportunities for all connections to progress satisfactorily. Again for the network topology of Fig. 3, we run variable mixtures of 20 Reno/Robin connections, sharing a common bottleneck link of B w =5 Mbit/s. Other parameters were set to B=100 Mbit/s, Q=20 packets, RTT p =120 ms, SMSS=1400 bytes and R c =4% (for all flows). Fig. 7 shows the average throughput per connection, with no packet loss on the wireless link, vs. the number of - Reno connections in the mixture of 20 in total. For example, at the point marked 4 on the X-axis, the experiment run a mixture of 4 Reno plus 16 W+Robin connections. By inspection of Fig. 7, it appears that W+Robin competes for resources with Reno rather fairly, without stealing a high fraction of bandwidth: all connections achieve about 1/20 of the shared bandwidth. On the other hand, Fig. 8 shows the average throughput per connection achieved with uniform random loss rate PER=2% (memoryless loss model) on the wireless link. In this case, it is again evident that W+Robin is more effective in exploiting the available bandwidth compared to Reno. Average Throughput [kbit/s] W+Robin Reno No. of Reno Connections Fig. 7: Average throughput per connection achieved by 20 Reno and W+Robin connections with no packet loss on the wireless link. Fig. 5: Evolution of the congestion window with Additive Increase ( Reno) vs. with Adaptive Increase (W+Robin) for PER=1%.

5 Average Throughput [kbit/s] W+Robin Reno No. of Reno connections Fig. 8: Average throughput per connection achieved by 20 Reno and W+Robin connections with uniform random loss rate PER=2% on the wireless link. VI. CONCLUSIONS AND TOPICS FOR FURTHER RESEARCH In this paper, a sender-side method for enhancing performance in heterogeneous networks including wireless links was proposed: the Robin scheme. In this method, a loss discrimination algorithm enables the sender to distinguish effectively congestion losses from corruption losses. Moreover, a segmentation algorithm, according to the perceived congestion level of the network, improves the error recovery phase during a non-congestive period. The Robin scheme was added to the -Westwood protocol, a well-known variant recently proposed for wireless networks. Its performance was then compared by simulation to that of Reno and plain W, both in lossy and nonlossy networks. Robin proved effective in enhancing W to achieve higher throughput in presence of packet errors on the wireless link. Moreover, W+Robin proved friendly in sharing available bandwidth with Reno. Further work is currently in progress. For example, we are interested to assess quantitatively the accuracy of the Robin heuristic criterion to distinguish between losses due to congestion or to transmission errors, for different network scenarios and wireless channel models. Moreover, the three Robin algorithms will be tested more thoroughly and in various scenarios for further refinement and parameter tuning. Other important research topics regard friendliness and asymmetric connections. ACNOWLEDGEMENTS The Authors wish to thank Fabio Martignon and Antonio Capone (with the same Dept. at Politecnico di Milano), who have been always lavish of helpful suggestions. REFERENCES [1] V. Tsaoussidis, I. Matta, Open Issues on for Mobile Computing, The Journal of Wireless Communications and Mobile Computing, Wiley Academic Publishers, Issue 2,Vol.2, February [2] A. Bakre, B.R. Badrinath, I- Indirect for Mobile Hosts, Proceedings of 15 th International Conference on Distributed Computing System, Vancouver, BC, Canada, June [3] A. Chockalingam, M. Zorzi, V. Tralli, Wireless Performance with Link Layer FEC/ARQ, Proceedings of IEEE ICC 99, Hong Kong, June [4] Rajesh Krishnan, Mark Allman, Craig Partridge, James P.G. Sterbenz, and William Ivancic, "Explicit Transport Error Notification (ETEN) for Error-Prone Wireless and Satellite Networks - Summary," Earth Science Technology Conference , Pasadena, CA, USA, June 11-13, [5] K.K. Ramakrishnan, S. Floyd, D. Black, The Addition of Explicit Congestion Notification (ECN) to IP, RFC 3168, Sept [6] M. Allman, V. Paxson, W. Stevens, Congestion Control, RFC2581, April [7] S. Mascolo, C. Casetti, M. Gerla, M.Y. Sanadidi, R. Wang, Westwood: Bandwidth Estimation for Enhanced Transport over Wireless Links, Proceedings of ACM SIGMOBILE, Rome, Italy, [8] R. Ludwig, R. H. Katz, The Eifel Algorithm: Making Robust Against Spurious Retransmissions, ACM Computer Communication Review, Vol. 30, No. 1, January [9] M. Mathis, J. Mahdavi, S. Floyd, A. Romanow, Selective Acknowledgement Options, RFC2018, Oct [10] S. Cen, P. Cosman, G. Voelker, End to end differentiation of congestion and wireless losses, Multimedia Computing and Networking 2002, San Jose, California, January [11] L.S. Brakmo S.W. O Malley, L.L. Peterson, Vegas: End to End Congestion Avoidance on a Global Internet, Proceedings of SIGCOMM'94, London, UK, Aug [12] S. Biaz and N. Vaidya, "Sender-Based Heuristics for Distinguishing Congestion Losses from Wireless Transmission Losses", Technical Report TR98-013, CS Dept., Texas A&M University, June [13] S. Biaz and N. H. Vaidya, Discriminating Congestion Losses from Wireless Losses Using Inter-arrival Times at the Receiver, Technical Report , CS Dept., Texas A&M University, June Revised August [14] S. Biaz and N. H. Vaidya, Distinguishing Congestion Losses from Wireless Transmission Losses: a Negative Result. Proceedings of IEEE 7 th International Conference on Computer Communications and Networks, New Orleans, LA, USA, October [15] A. Capone, F. Martignon, "Bandwidth Estimates in the Congestion Control Scheme", Proceedings of Tyrrhenian IWDC 2001, Taormina, Italy, September Included in "Evolutionary Trends of the Internet", Ed. S. Palazzo, Berlin, Germany: Springer-Verlag, [16] V. Jacobson, Congestion Avoidance and Control, ACM SIGCOMM88, Stanford, CA, USA, August [17] G. Xylomenos, G.C: Polyzos, P. Mahonen, M. Saaranen, Performance Issues over Wireless Links, IEEE Communications Magazine, Vol. 39, No. 4, pp , April [18] M. Mellia, M. Meo, C. Casetti, Michela Meo, Smart Framing: Using Smart Segments to Enhance the Performance of, Proceedings of Globecom 2001, San Antonio, TX, USA, Nov [19] VINT project U.C. Berkeley/LBNL, "ns-2 network simulator (ver. 2)". Available at URL: [20] Massimo Valla (UCLA-CSD), " Westwood ns-2 modules". Available at URL: [21] J. C. Haartsen, "The Bluetooth Radio System", IEEE Personal Communications, vol. 7, no. 1, Feb. 2000, pp [22] C. Snow and S. Primak, "Performance Evaluation of /IP in Bluetooth Based Systems", Proceedings of IEEE VTC Spring 2002, Birmingham, Alabama, May 2002.