A New Modified Split-Connection Approach for Improving TCP Performance Over Wireless Networks

Transcription

1 A New Modified Split-Connection Approach for Improving TCP Performance Over Wireless Networks Yashar Komijani University of Tehran Arash Hooshmand University of Tehran Abstract- One of the most important challenges in providing Internet connectivity to mobile users is TCP performance degradation in wireless environments, due to its inability to distinguish congestion losses from corruptions. One of the most effective solutions to this problem is Split-Connection which has not been practical yet because of suffering from End-to-End semantic violation of TCP acknowledgements. In this paper, different Split-Connection methods are compared and a novel, practical and efficient modification to Split- Connection is proposed to preserve the End-to-End semantic without compromising its performance. I. INTRODUCTION Fast growth of Internet popularity and mobile communication has made it necessary to provide users with Internet connectivity, using wireless technology. However, there are several problems in providing Internet over wireless networks which have been recently challenged by researchers. TCP performance degradation over wireless links is one of these problems that is well considered during last years. Although several solutions have been proposed to the problem, the results are still far from ideal conditions. A. Background on TCP TCP [1] has been used in Internet as a transport protocol to provide a reliable byte-stream service to upper layers by exchange of data and acknowledgement packets (Ack) between sender and receiver. It limits the number of outstanding packets using a sliding window and uses Ack arrivals as a clock to send new data packets to the network The first TCP congestion control mechanism developed by Jacobson [2] is used in Tahoe, the core of which is a variable-size sliding window which adapts the number of outstanding packets to instant network state using the information carried by acknowledgment packets. Tahoe has a cyclic window evolution which includes Slow Start and Congestion Avoidance phases. During Slow Start phase, window is rapidly grown up to the estimated capacity of the network and then Congestion Avoidance is employed to probe for further network capacity. This growth is continued until an overflow occurs in one of the intermediate buffers and one of the sent packets is lost. Packet loss is detected by timer expiry and the window size is reduced to one packet and Slow Start is then employed again to fill the connection pipe and start the self-clocking mechanism. Half of maximum window size just before timer expiry is selected to be an estimation of network capacity for the next cycle. In Reno version [3] developed by Jacobson, Fast Retransmission and Fast Recovery mechanisms were added and resulted in better performance. Fast Retransmission enables sender to detect losses long before coarse-grained TCP timer expiry by inferring the reception of three duplicate Acks as a sign of packet loss. Fast Recovery is used after Fast Retransmission to reduce the window size to half of its value and invoke Congestion Avoidance without return to Slow Start. Interested readers are referred to [2, 3, 4] for a detailed description of Tahoe and Reno, to [5] for an analytical analysis and to [7] for a simulation study. B. Previous Works The problem of improving TCP performance over wireless links has been investigated during last few years by research community and lots of methods are proposed. Based on these works, the main causes of TCP performance degradation over wireless links can be classified in following three groups [11]: 1) Congestion and corruption loss differentiation: TCP infers packet loss as a sign of congestion and reduces its window size. In the presence of wireless links, packets may also be lost due to corruptions caused by fading effects and hand offs. Hence TCP performance is degraded in wireless links because of inability to distinguish between congestion and these corruption losses. When packet loss exceeds a minimum threshold, consecutive window size reduction results in very poor performance [5, 6]. The classic solution to this problem is employing Explicit Loss Notification [12] in which receiver informs sender of packet loss. Knowing that the corruption is the cause of packet loss, sender does not reduce its window and just retransmits the lost packet. ELN Implementation has several practical issues which are not solved yet. The most important problem is how receiver can find out whether packet loss is occurred due to congestion or corruption. Another problem is that ELN needs End-to-End TCP modification. Others have proposed other solutions to congestion and corruption differentiation, most of them based on using RTT interarrival times [13]. 2) Multiple packet loss in a window: Another well-known problem of TCP in wireless environments is that TCP can not recover from multiple packet loss in a window without resorting to a coarse-grained timer expiry [7]. The reason is that TCP uses a cumulative acknowledgement that one Ack acknowledges reception of all previous data segments. Multiple packet losses are common especially in wireless links due to burst errors caused by fading and hand offs.

2 To alleviate this problem, TCP Sack is designed [8] that using sender and receiver modification, enables sender to have explicit information about receiver window gaps and which packet must be retransmitted. Another sender-side modification of TCP is NewReno [9] which is designed to use partial Acks to decide which packet must be retransmitted. Sack and NewReno require End-to-End TCP modification. 3) Reaction Delay: Even in the best case, if TCP was able to detect corruptions and recover well from multiple packet losses, it had to wait until reception of multiple duplicate Acks and retransmit the lost packet(s). This takes multiple round-trip times and results in reaction delay which degrades performance. Dependency of TCP throughput to the average round-trip-time in first approximation [10] is 1.22L Th( τ, q) (1) τ q in which L is packet length, τ is round-trip-time and q is packet loss probability. Equation 1 shows that round-trip time influence is greater than packet loss probability. Reducing this reaction time is the solution to the mentioned problem. Link-layer and Split-Connection methods solve this problem by reducing reaction time. Link-layer solutions use FEC/ARQ over wireless link to hide the channel loss from sender view [15, 16]. The channel appears to sender as a high-quality but longer channel with a lower capacity. Link-layer solutions suffer from out-of-order delivery which causes duplicate Ack and sender window reduction [14]. To cure this problem TCP-aware link-layer solutions are proposed but they need inter-layer communication and difficult implementation [17]. Layer 2 and 4 timers competition is another problem of link-layer solutions [18]. Split-connection approach is another solution to the reaction delay problem which splits TCP connection in BS and employs an optimal TCP over wireless part. So, the round-trip time is reduced in the second connection, because sender and receiver are closer to each other and this allows faster reaction to the corruption losses. Split- Connection details are deferred to following sections. Foreign Host Wired Network Base Station Fig 1. Basic model of a wireless LAN connection. II. SPLIT-CONNECTION APPROACH Mobile Host Split-Connection (Indirect-TCP) method [19] is based on breaking TCP connection at the base station into two distinct connections, a standard (Reno) connection from foreign host to the base station and another well-tuned TCP protocol connection over wireless part (figure 2). FH TCP-Reno BS Optimal TCP Fig. 2. Split-Connection method. MH Packets arrived at BS in the first connection are immediately acknowledged and their payload is then transferred to the second connection. This is called indirect- Ack and is illustrated in figure 3. C. Outline Although a thorough solution to the problem is still far from achievement. As stated in previous section Split- Connection is especially useful in reducing reaction delay in long paths. Therefore we limit our study to the connections with a long wired wan path and a short wireless one, as is typical of WLAN mobile users connections to Internet. This paper includes a comparison among several proposed Split-Connection variants and a novel method which is shown to have considerable improvement on previous solutions. Remainder of paper is organized as follows. Split-Connection details are explained in the next section. The proposed solution for preserving End-to-End semantic is introduced in third section. Forth section is devoted to a simulation-based comparison of explained methods and our conclusion is brought in section five. Fig. 3. Indirect-Acknowledgement diagram. In addition to reaction delay reduction, Split-Connection has several other advantages: 1) The wireless link is fully isolated so that losses due to corruption and hand off are hidden from TCP sender. Arriving packets in BS are always immediately acknowledged except when the BS buffer is full which causes the arrived packets to be dropped. This helps the wireless link to be treated by End-to-End TCP as a bottleneck link with average capacity of Eq. 1 and queue size of buffer length in BS. Because of flow control and corruption loss recovery separation, losses seen by sender are whether caused by congestion in wired network or BS buffer overflow.

3 2) Since the connection is broken at the BS, an optimal TCP can be employed over wireless link without any need for End-to-End TCP modification. This makes it easy to combine all of the proposed solutions together, using Split- Connection in End-to-End and optimal TCP and link-layer solutions over the wireless part. The main problem with Split-Connection is that since two TCP connections are independent, data packets may be acknowledged by base station much before they are actually received by mobile host (Figure 3). The mobile host or base station may go off and acknowledged packets will never be received. Therefore, Split-Connection in its previous form (Indirect-Ack) can not be used in Internet. A. Direct-Acknowledgement To alleviate the problem of End-to-End semantic violation, a simple solution [20] has been suggested which delays Acks from BS to FH until data packets are actually received by MS and corresponding Ack has arrived at BS (fig 4). This method is called direct-ack and is also used in [21]. Ack) while it enjoys very good performance (as like as Indirect-Ack). A. A Hypothetical Virtual-Acknowledgement Approach As a reliable connection-oriented transport protocol, TCP is usually asked to transfer a message which may contain either a large file or a single character. The message is fragmented to several segments and the application is blocked until the message is totally transferred and all of segments are acknowledged. Let s suppose that BS could somehow recognize the last segment in a sequence of segments. Hence it could virtually acknowledge all of the received packets but the last one immediately (opposite to Indirect-Ack in which no recognition is performed and therefore the last one is acknowledged too). The last packet s acknowledgement was postponed to the time when it was really acknowledged by the MS in the second link. The last packet is acknowledged by MS only if all segments of the message are received. So when the last Ack is eventually received by TCP sender in FH, it means that all of segments have been successfully received by MS. So not only almost all of the Acks were sent to the FH sender immediately (sender was not kept waiting as like as in Direct-Ack), but also application layer End-to-End semantic of Acks was preserved because the application layer keeps user waiting until the last Ack is received. Fig. 4. Direct-Acknowledgement diagram. Although Direct-Ack method solves End-to-End semantic problem but it has several disadvantages. The first is that it keeps sender waiting until packets are Acked in the second connection. This reduces isolation between two connections and causes FH timer expiry and window reduction even when BS buffer is empty. When the packet is finally delivered to MS, BS empty buffer results in second link starvation. Another problem is that large and irregular round-trip time of Acks results in large RTT-deviation, long RTO thresholds and hence delayed congestion detection. III. VIRTUAL-ACKNOWLEDGEMENT We believe that Indirect-Ack is too optimistic that all packets will be acknowledged and Direct-Ack is unnecessarily pessimistic. In this section a new method, we will call it Virtual-Ack from now on, is introduced that is between Direct-Ack and Indirect-Ack. Virtual-Ack preserves End-to-End semantic of Acks (as well as Direct- Fig. 5. Virtual-Acknowledgement diagram. B. A Novel Practical Virtual-Acknowledgement Algorithm Unfortunately it is not possible for the BS to recognize which segment is the last one in a sequence of segments (which in general maybe belong to different messages sent back-to-back by FH). However based on the above idea a more complicated yet practical solution is suggested below. Although Bs can not recognize the last segment of a message, in Virtual-Ack solution we modify BS to always treat the last received segment as though it is really the last segment (that must be acknowledged End-to-End). If a new segment with higher sequence number is arrived while the last one has not acknowledged by MH yet, it confirms that

4 the packet which was being hold up to now is not the last one. So it is acknowledged and the new one is believed to be the last one. Suppose that packet n has been just arrived at BS and has the most seen sequence number. This packet is supposed to be the last packet of the message. Hence its Ack is delayed until either a higher sequence number packet is arrived at BS or it is really acknowledged by MH. are used over the wired and wireless link respectively. Wireless link is assumed to be a memory less channel with a probability of packet loss of PER. Simulation runs of each protocol are performed with wireless packet error rate ranging from 0 to 20 percent. For each PER 4000 seconds simulation is performed 3 times. TCP-Reno Optimal TCP FH 800 Kbps 100 msec BS 1 Mbps 10 msec MH Fig. 5. Virtual-Acknowledgement diagram. According to fig. 5, BS acknowledges packet n when it is really acknowledged by MH. When packet n+1 is received, it is assumed that this is the last packet. So its ACK is deferred until it is Acked in second link, but when packet n+2 arrives, it is evident that packet n+1 was not the last one and so n+1 is Acked. Finally Ack n+2, sent by MH is arrived at BS and hence it is acknowledged by BS. Virtual-Ack method preserves End-to-End semantic in application layer because the last packet is acknowledged whenever it is Acked in the second link. Immediate acknowledgement of segments that are not really received by receiver, not only prevents Acks from excessively being delayed in wireless link, but also keeps intermediate buffer full. Thus prevents wireless connection to be underutilized and provides sender with a more regularly spaced RTT samples and as a more accurate RTO which results in better performance. B. Simulation Results Fig. 6. Simulation network topology Split-Connection (Indirect-TCP) is supposed to have best-performance results due to its immediate Ack. The improvement gained by Indirect-Reno is related to decreased RTT and isolation. In order to examine the good influence of selecting a proper protocol over wireless link, different transport protocols must be simulated. Through several simulations that are not presented here, performance of Tahoe, Reno, NewReno, Sack, Vegas, Westwood and WestwoodNR have been examined in similar conditions and TCP Sack has had best-results over wireless link [30]. So we focus on Reno as the standard protocol and Sack as the optimal-protocol over wireless link (fig 7). IV. COMPARISION A. Simulation Setup The Berkeley NS simulation tool [22] is used for performance evaluation of the different proposed Split- Connection variants. Network topology is sketched in figure 6 in which wired network of figure 1 is modeled by a lowcapacity, long-delay link that is typical character of WAN bottleneck links. Bandwidth and delay values are typical values for a connection between a wireless LAN client to an Internet server. The value of 100 msec is chosen as an average round-trip times encountered in Internet [21]. The segments size is assumed to be 1000 bytes. A buffer of 100 Kbytes is set in the base station. TCP Reno and Reno/Sack Fig. 7. Split-Connection Performance Improvement The slight improvement of Reno-based Split-Connection performance is due to second link starvation caused by empty buffer in zero packet error rates.

5 It may be argued that good results of Virtual-Ack reported above are due to the infinite source assumption so that BS always receives higher sequence number packets to acknowledge previously received ones. Further simulation results are presented here for better comparison. The required time for transferring 100 Kbytes, 1 Mbytes and 10 Mbytes files in different PERs are reported in table 2. From this table, it is obvious that the waiting time for the last Ack to arrive doesn't have considerable influence on performance. The required time for Virtual-Ack is almost always much less than Direct-Ack method. TABLE II THE REQUIRED TIME TO TRANSFER A FILE (Sec) Fig. 8. Reno-based Indirect, Direct and Virtual-Ack Comparison. Virtual-Ack exhibits exact matching with Indirect-Ack. Different Ack methods are compared with each other and with the End-to-End Reno protocol in fig 8. While Direct-Ack compromises performance in order to preserve End-to-End semantic, Virtual-Ack exactly matches Indirect- Ack performance which is equal to the ideal behavior. PER 100 Kbytes 1 Mbytes Virtual Indirect Direct Virtual Indirect Direct Sack-based Virtual-Ack to End-to-End Solutions is compared in Figure 10. Sack and NewReno recover from multiple packet losses in a window. No new data packet is transferred during loss recovery phase which lasts for several RTTs. This lateness and sender silence is the main reason that causes End-to-End Sack and NewReno not to improve performance considerably. In Virtual-Ack much smaller RTT is used and as the result, silence durations are reduced leading to better improvement. Fig 9. Sack-based Indirect, Direct and Virtual-Ack Comparison. Virtual-Ack exhibits exact matching with Indirect-Ack Fig. 10. Sack-based Virtual-Acknowledgement is compared to Endto-End Reno, NewReno and Sack.

6 V. CONCLUSION Preserving End-to-End Semantic of Acks in application layer (instead of transport layer) is sufficient for a reliable data delivery. Based on this idea, a new modification to Split-Connection is proposed (virtual-ack) that shows considerable improvement on previous solutions. Better Isolation of corruption losses and more regular Ack generation is keys to Virtual-Ack better performance. Virtual-Ack can be easily implemented in base station and does not require End-to-End TCP modification. Therefore it can be gradually employed over wireless links in current Internet. Furthermore it can be combined with previous methods. Remaining works include combining Virtual-Ack with link-layer methods. ACKNOWLEDGEMENT We would like to thank Andrei Gurtov for reading of earlier draft of this paper and his valuable advices. REFRENCES [1] J. B. Postel. "Transmission Control Protocol". RFC, Information Sciences Institute, Marina del Rey, CA, September RFC-793. [2] V. Jacobson, "Congestion Avoidance and Control", Proc. ACM Sigcomm 88, August 1998 [3] V. Jacobson, "Modified TCP congestion avoidance algorithm", end2end-interest mailing list, April. 30, 1990 [4] M. Allman, V. Paxson, W. Stevens, TCP Congestion Control, RFC, 2001, RFC-2581 [5] T. V. Lakshman, U. Madhow. "The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss", IEEE/ACM Trans. on Networking, 5(3): , [6] H. M. Chaskar, T. V. Lakshman, U. Madhow, "TCP over Wireless with Link Level Error Control: Analysis and Design Methodology", IEEE/ACM Transactions on Networking, Vol. 7, No. 4, Oct [7] K. Fall, S. Floyd, "Simulation-based Comparisons of Tahoe, Reno, and Sack TCP", Computer Communications Review, [8] Mathis, M., Mahdavi, J. and Floyd, S. and Romanow, "A. TCP Selective Acknowledgement Options", RFC [9] S. Floyd, T. Henderson, "The NewReno Modification to TCP's Fast Recovery Algorithm", RFC, 1999, RFC [10] M. Mathis, J. Semke, J. Mahdavi, "The Macroscopic behavior of TCP Congestion Avoidance Algorithm", ACM Sigcomm, vol 27, number 3, July [11] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, R. H. Katz, "A Comparison of Mechanisms for Improving TCP Performance over Wireless Links". IEEE/ACM Trans. on Networking, 5(6): , [12] H. Balakrishnan and R. H. Katz, Explicit loss notification and wireless web performance, IEEE Global Telecomm. Conference (Globecom 98), Mini Conference, Sydney, Australia, Novermber [14] A-F. Canton, T. Chaed, "End-to-End Reliability in UMTS: TCP over ARQ", proc. of Globecomm 2001 [11] Mario Gerla, M. Y. Sanadidi, Ren Wang, Andrea Zanella, Claudio Casetti, Saverio Mascolo, "TCP Westwood: Congestion Window Control Using Bandwidth Estimation", In Proceedings of IEEE Globecom 2001, Volume: 3, pp , San Antonio, Texas, USA, November 25-29, [13] H. Koja, T. Ikenaga, Y. Hori, Y. Oie, "Out-of- Sequence in Packet Arrivals due to Layer 2 ARQ and Its Impact on TCP Performance in W-CDMA Networks" Symposium on Applications and the Internet, Jan [21] Y. Bai, A. T. Ogielski, G. Wu, Interactions of TCP and Radio Link ARQ Protocol, in Proc. of IEEE VTC, 1999, pp [23] Karunaharan Ratnam, Ibrahim Mata, Effect of Local Retransmission at Wireless Access Points on the Round Trip Estimation of TCP, The 31s Annaual Simulation Symposiium. April 05-09, 1998, Boston, Massachusetts [14] A. Bakre and B.R. Badrinath. I-TCP: Indirect TCP for Mobile Hosts. In Proceedings of ICDCS, pages , [15] K. Brown and S. Singh. M-TCP: TCP for Mobile Cellular Networks. ACM/SIGCOMM Computer Communications Review, 27 No.5:19 43, October [16] L. S. Brakmo, S. W. O'Malley, L. L. Peterson, "TCP Vegas: New techniques for congestion detection and avoidance", Proc. ACM Sigcomm, August 1994 [17] Padhye, J., Firoiu, V., Towsley, D., and Kurose, J., Modeling TCP Throughput: a Simple Model and its Empirical Validation, SIGCOMM [18] K. Ratnam and I. Matta."WTCP: An efficient transmission control protocol for networks with wireless links". Proc. Third IEEE Symposium on Computers and Communications (ISCC '98), Athens, Greece, 1998 [19] D. E. Comer, D. L. Stevens, Internetworking with TCP/IP, Volume II: Design, Implementation, and Internals. Prentice Hall, [20] V. Chirm, V. C. M. Leung, End-to-End Acknowledgments for Indirect TCP over Wireless Internetworks. IEEE, 1997 [13] Q. Ni, Th. Turletti, W. Fu, Simulation-based Analysis of TCP Behavior over Hybrid Wireless & Wired Networks, Proceedings of Wired Wireless Internet Communications 2002 (WWIC 2002), Las Vegas, June [26] Berkeley NS Simulator, available at [28] V. Paxson, Measurement and analysis of end-to-end internet dynamics, Ph.D. dissertation, University of California, Berkeley, 1997.