Discriminating Congestion Losses from Wireless Losses using. Inter-Arrival Times at the Receiver. Texas A&M University.

Transcription

1 Discriminating Congestion Losses from Wireless Losses using Inter-Arrival Times at the Receiver Saad Biaz y Nitin H. Vaidya Department of Computer Science Texas A&M University College Station, TX 7784-, USA saadb@cs.tamu.edu Phone : (49) Fax : (49) Technical Report #98-4 Abstract TCP has been designed and tuned to perform well under the assumption that all losses are an indication of congestion. When a TCP connection traverses a wireless link, packets may be lost due to wireless transmission errors, in addition to congestion losses. TCP implicitly assumes that all packet losses are due to congestion, and triggers congestion control mechanism when a packet loss is detected. It has been previously demonstrated that this feature of TCP aects performance adversely when packets are lost due to transmission errors. To avoid the performance degradation, techniques to distinguish between corruption and congestion losses without any explicit information from the network (routers or switches) are of interest. The objective of this paper is to present a simple scheme which can be implemented at the receiver to distinguish congestion losses from corruption losses. The scheme works in the case where the last hop to the receiver is a wireless link and has the smallest bandwidth among all links on the connection path. With such mechanism, the receiver can attempt to detect the real cause of the packet loss and inform the TCP sender to take the appropriate actions. Specically, if the loss is a transmission error, the receiver can speed up the recovery and avoid shrinkage of sender's congestion window. We added our mechanism to TCP-Reno to evaluate the performance improvement. We compared our scheme against Ideal TCP-Reno which is TCP-Reno that can perfectly (but articially) distinguish between congestion losses and wireless transmission losses. Under favorable conditions, our scheme performs similarly to Ideal TCP-Reno and can lead to signicative throughput improvement. Key Words: TCP { Packets Inter-Arrival Time { Transmission Error Losses { Congestion Losses Research reported is supported in part by the Fulbright Program and the National Science Foundation. y Principal contact author

2 Introduction TCP is a popular protocol for reliable data delivery in the Internet. TCP is robust in that it can adapt to disparate network conditions []. Specically, TCP performs well if most of the packet losses are due to congestion: TCP uses congestion control mechanisms to recover from all packet losses that may occur in the network. Therefore, if a fair portion of packet losses are due to transmission errors, TCP's congestion control mechanisms degrade its performance. In recent years, wireless environments with transmission errors are becoming more common. This trend motivates our work on distinguishing between the two kinds of packet losses. Motivation: In recent years, with the advent of mobile computing, there has been signicant interest in using TCP over wireless links [4, 5, 8, 6, 5, 4, 9]. Previous work has shown that, unless the protocol is modied, TCP performs poorly on paths that include a wireless link subject to transmission errors. The reason for this is the implicit assumption in TCP that all packet losses are due to congestion. Whenever a TCP sender detects a packet loss, it activates congestion control mechanisms [] (these mechanisms reduce sender's window in response to the packet loss, reducing throughput temporarily). Taking congestion control actions may be appropriate when a packet loss is due to congestion, however, it can unnecessarily reduce throughput if packet losses happen to be due to wireless transmission errors. Past proposals for improving performance of TCP over wireless require some cooperation from an intermediate node on the path from the sender to the receiver [4, 5, 6, 5]. An IETF work group dedicated to TCP over satellite link is undergoing an intense eort to identify specic mechanisms [] and issues [] to adapt TCP to this environment. Among these issues is the problem discussed by the present paper. Our interest is in mechanisms that impose minimal demands (if any) on any host other than the sender or the receiver. Ideally, it would help if the sender could dierentiate between packet losses due to congestion from the packet losses due to wireless transmission errors, using some end-to-end technique (that does not get any help from any intermediate host). Once a sender knows that the packet loss is due to congestion or corruption, it can respond appropriately. We previously [7] tried to adapt some well-known congestion avoidance schemes to enable the sender to distinguish between the two types of packet losses. That approach did not yield good results. With the cumulative acknowledgement used by TCP, the sender does not know exactly which packets are lost. The receiver has a better view of the losses : it knows exactly which packets are lost. This observation led us to consider schemes which can enable the receiver to distinguish between congestion losses and transmission error losses. In this paper, we present our scheme and measure its ability to distinguish congestion error The TCP layer at receiver may not know the cause of a packet loss even if it is a transmission error on the last hop, as a lower layer may discard the packet before it reaches the TCP layer.

3 losses from wireless transmission error losses. We use simulation to study this ability of discrimination. Then, we modied a TCP-Reno sender and a TCP receiver to integrate our scheme and study the throughput enhancement induced. The TCP-Reno modied with our scheme will be called TCP- Aware. We compare the performance of our scheme TCP-Aware against an Ideal TCP-Reno sender which distinguish perfectly congestion losses from wireless transmission losses. The Ideal TCP-Reno sender is implemented articially such that the TCP sender has a perfect knowledge of the real cause of a packet loss. This paper is organized as follows. Section. presents the system model and describes the proposed scheme. Simulation model and simulation results on the ability of our scheme to distinguish losses are discussed in Section. Section 4 is dedicated to the performance study of TCP using our scheme. Conclusions and further work are presented in Section 5. The Proposed Scheme In this section, we present rst the model of environment in which our scheme works well. We then describe our scheme and show why such environment is so favorable.. The Model TCP connections may traverse wireless links in several scenarios, as illustrated in Figure - a solid line in this gure depicts a wired link, and a dashed line depicts a wireless link. In Figure (a), only the last hop is wireless; this scenario typically occurs when a mobile host communicates with a xed host. In this case, the mobile host is connected with a base station via a wireless link. As shown in Figure (b), intermediate links may also be wireless. This situation can occur, for instance, when the path includes a satellite hop. Other scenarios shown in Figure also occur in practice. In this paper, we consider only the scenario shown in Figure (a) where the sender is on the wired network and the receiver is connected via the wireless link. Moreover, we assume that the wireless link is the bottleneck for the connection. This assumption that the wireless link is the bottleneck may be valid (a) (b) base station (c) (d) LEGEND node wired link wireless link for environments such as the following : Figure : Typical scenarios illustrating use of wireless links

4 mobile hosts connected through cellular phones an Ethernet LAN with wireless access over WAVELAN. If the wireless link is the bottleneck for the connection, then the packets sent tend to queue up at the base station. Therefore, most of the packets are sent back to back on the wireless link. This last characteristic is the key to the simple heuristic we developed to distinguish packet losses due to transmission errors from congestion losses on the wired network. We describe in Section. our heuristic.. Packet Inter-Arrival Time as the Discriminator We describe in this section a scheme which allows the receiver to discriminate between transmission losses and congestion losses. For the sake of simplicity in the description, initially, we assume that the processing time at the base station and at the receiver is negligible. We also assume that : only the last link on the path from the sender to the receiver is wireless. the wireless link is the bottleneck for the connection the sender performs a bulk data transfer In this situation, because the wireless link is the bottleneck, the base station will typically buer more than one packet that is destined to the receiver. Now consider three situations, as illustrated in Figure. The gures show three scenarios that may occur when the sender sends packets, and to the receiver. In Figure (a), none of the packets are lost. Therefore, the receiver receives all of the packets. In this case, the \packet inter-arrival gap", or the time between arrival of consecutive packets, is approximately equal to the time T required to transmit one packet on the wireless link - in practice, the inter-arrival gap can vary due to other factors, including queueing delay. Note that the time of arrival of a packet is the time when all bits belonging to the packet have been received by the receiver. Now, consider Figure (b) which shows that packet is lost when transmitted over the wireless link. In this case, the time between the arrival of the two packets received by the receiver (i.e., packets and ) is T, since packet (lost due to transmission) uses the wireless link for T time units. Now consider the third possibility shown in Figure (c). Here, assume that packet is lost due to queue overow (congestion) at the intermediate router. In this case, the inter-arrival gap between packets and will be comparable to T (this holds true if packet arrives at the base station either before, or just after, the base station has transmitted packet ). When packet is lost and packet Note that the sender sends at irregular intervals packets, and. At the router, the inter delay between packets may be modied due to the cross trac and the queueing policy of the router. But, at the base station, the packets are sent back to back because the wireless link is the slowest link along the path for the connection. 4

5 sender BS T T T receiver router (a) T BS (b) ~T BS drop (c) Figure : Inter-arrival gap arrives at the receiver, packet is called an out-of-order packet. Based on these observations, We have developed the following heuristic : Let T min denote the minimum inter-arrival time observed so far by the receiver during the connection. Let P o denote an out-of-order packet received by the receiver. Let P i denote the last in-sequence packet received before P o. Let T g denote the time between arrivals of packets P o and P i. Finally, let the number of packets missing between P i and P o be n (assuming that all packets are of the same size). if (n + )T min T g < (n + )T min, then the n missing packets are assumed to be lost due to wireless transmission errors. Otherwise, the n missing packets are assumed to be lost due to congestion. Note that we are very selective for wireless losses. The reason is that it is preferable, for the sake of the network, to mistake a wireless loss for a congestion loss. Indeed, if we mistake congestion losses for wireless losses too frequently, TCP's congestion control mechanisms will not be triggered and this may lead to congestion collapse. By mistaking some wireless losses for congestion losses, we will not perform worse than TCP as it is today (TCP considers all wireless losses as congestion losses). Of course, there are many scenarios for which this heuristic will be defeated. Our preliminary measurements show that our heuristic works in the case when the wireless link has the lowest 5

6 bandwidth. To evaluate our scheme, we will measure the accuracy of discrimination of our heuristic. We use two simple and natural metrics : A c : accuracy of congestion loss discrimination A w : accuracy of transmission loss discrimination. A c is dened as the ratio of the number of congestion losses correctly identied over the total number of congestion losses. For instance, if congestion losses occurred and our heuristic identied them correctly 75 times then the accuracy A c is :75. A w is similarly dened but for transmission error losses. Simulation Model. Simulation Model We evaluate our scheme using the simulation tool ns- (version.b) [] from berkeley. Figure depicts the topology used. The topology is simple and yet serves our purpose. We have a TCP connection from the xed host to the mobile host. We use the Reno agent from ns- for the TCP connection. This connection shares the link R! R with a cross trac issued by four T raffic=expoo [] agents. The trac ows from the sources CBR I to the sinks SINK i (i = ; ; ; ). The T raffic=expoo agent from ns- [] is a constant-bit rate (CBR) source with idle time and busy time exponentially distributed with mean. s. UDP is the transport protocol used for this type of source. All the links in Figure are labeled with a (bandwidth, propagation delay) pair. Note that propagation delay does not include transmission time or queueing delays. The link between the router R and the mobile host is wireless. This link has a bandwidth bw which takes values 64 Kbits/s, 8 Kbits/s, 56 Kbits/s, 5 Kbits/s, Mbits/s and Mbits/s. This link has a transmission error rate r w which will take the values from % to 5%. r w is measured as a fraction (or percentage) of packets lost due to wireless transmission errors. In this paper, we set the propagation time from R to the mobile host to ms. Similarly, the four links from R?! SINK i have a propagation time of ms. All others links are wired with a bandwidth bw and a propagation delay. The bandwidth bw takes values 64 Kbits/s, 8 Kbits/s, 56 Kbits/s, 5 Kbits/s, Mbits/s and Mbits/s. The propagation delay takes the values ms, 8 ms and 8 ms such that the round trip propagation time varies from 6 ms to 74 ms. Note that propagation delay does not include queueing and transmission delays. ns- does not support mobility. A wireless link here is a link with transmission error packet losses 6

7 Fixed Host ( TCP Sender) Mobile Host (TCP Receivers) bw, δ bw, δ Router R qs bw, δ Router R bw, ms bw, ms CBR - CBR (Cross traffic) SINK - SINK Figure : ns Simulation Model. Methodology The objective is to set a TCP connection between a xed host and a mobile host. This connection has to contend with a random cross trac to share the link R?! R. This contention will generate congestion losses. Let r c denote the congestion loss rate for the TCP connection. r c is measured as a fraction (or percentage) of packets lost due to congestion. For each set of parameters r w, r c, bw, bw, and, the simulation has two phases. The rst phase is used to determine the right rate pkr for the T raffic=expoo agent to produce a given congestion loss rate r c. When the rate pkr is known, the second phase begins : we set the rate for the T raf f ic=expoo agent to pkr and run times one long-lived TCP connection ( to seconds). Note that each simulation starts with a warm-up period of s during which only the CBR sources are active. After the warm up period, the TCP connection starts after a random period between and ms. For each run, we measure the accuracies of discrimination A c and A w.. Simulation Results We conducted simulations varying the following parameters : the congestion loss rate r c takes values %, %, %, 4% and 5% the transmission error loss rate r w takes values %, %, %, 4% and 5% the wired bandwidth bw takes values 64 Kbits/s, 8 Kbits/s, 56 Kbits/s, 5 Kbits/s, Mbits/s and Mbits/s the wireless bandwidth bw takes values 64 Kbits/s, 8 Kbits/s, 56 Kbits/s, 5 Kbits/s, Mbits/s and Mbits/s. the round trip propagation T p time takes values 6ms, 4 ms and 74 ms. 7

8 For each set of parameters r c, r w, bw, bw, and T p, we measured the accuracy A c for congestion losses and the accuracy A w forwireless transmission errors. Figure 4 presents respectively the accuracies A c and A w with r c = % and r w = %. Figure 4(a) plots the accuracy A c of congestion losses for 4 dierents values (64 Kbits/s, 56 Kbits/s, 5 Kbits/s,and 48 Kbits/s) for the bandwidth bw on the wireless link. The four curves in Figure 4(a)corresponds to the four values of bw. The x-axis, with a logarithmic scale (base ), represents the bandwidth bw on the wired links. Figure 4(b) plots similarly the accuracy A w of wireless transmission errors. Two factors de-.8 56Kb 5Kb 48Kb.8 56Kb 5Kb 48Kb.6.6 Ac Aw wired bandwith (Kbits/s) (rw=% rc=% rtt= ms qs= pkts) wired bandwith (Kbits/s) (rw=% rc=% rtt= ms qs= pkts) (a) Accuracy A c for congestion losses (b) Accuracy A w for transmission error losses termine the accuracies A c and A w : Figure 4: Accuracies A c and A w with r c = % and r w = % the ratio bw bw of the wired bandwidth over the wireless bandwidth the overall loss rate r c + r w when bw bw < bw bw > : When bw bw >, i.e the bandwidth on the wireless link bw is smaller than the bandwidth on the wired link, the accuracy of transmission losses A w increases as the ratio bw bw noted earlier, the larger is the ratio bw bw basestation. Therefore the scheme is more ecient as the ratio bw bw increases. As we and larger will be the likelihood of packets buering at the increases. bw ' : Now, if the ratio bw is close to one then the packets can be equally likely at the sender, bw bw at the router R, or at the base station R output buers. Therefore, it is rare that packets are sent back to back from the base station. In this case, the accuracy for wireless transmission losses is small. 8

9 bw bw < : Now the wireless bandwidth is larger than the wired bandwidth. Then the packets are buered at the sender and the router R output buers. The link R?! R is the bottleneck. In this case, the minimum inter-arrival time is equal to the service time at router R. A wireless loss cannot be diagnosed reliably and accurately because the inter-arrival does not depend on the transmission time over the wireless link. Moreover, congestion losses may be mistakenly diagnosed as wireless losses. We show in the following that the smaller is the overall loss rate r c + r w number of congestion losses mistakenly diagnosed as wireless loss. and larger will the In Figure 5, we consider an overow. Cross trac packets are black and TCP packets are white and numbered,, and. The TCP packet number is dropped because of queue overow. Figures 5(a) and 5(b) dier only by the number of cross trac packets between the two TCP packets and. In Figure 5(a), there is only one cross trac packet between TCP packets and. In this case, our criteria may mistake this congestion loss for a wireless loss because the interarrival between packets and is more than twice the transmission time. If packets and are sent back to back, the loss will denitely be mistaken as a wireless loss. In Figure 5(b), there are two cross trac packets between TCP packets and. The interarrival time between packets and at the receiver is larger than thrice the service time at router R. In this case, the loss of TCP packet will be diagnosed correctly (This remains valid for or more cross trac packets between TCP packets and ). The case depicted in Figure 5(a) is more likely to happen than the second case depicted in Figure 5(b) if the cross trac is lighter than the TCP trac, i.e if the congestion loss rate r c is low or the overall loss rate r c + r w is low. This is clearly illustrated by the plots in Figure 4. In Figure 4 (r c = % and r w = %), we observe that, when the ratio bw bw gets smaller, A w increases and A c decreases. This is due to the fact that most congestion losses are mistakenly diagnosed as wireless losses. Now, if a packet p is lost on the wireless link, this loss will be diagnosed as wireless provided that packets, and are sent back to back on the link R?! R. Therefore, most of the losses (due to congestion or transmission) are diagnosed as transmission losses. When the congestion loss rate r c increases or if the overall loss rate r c + r w increases, the situation in Figure 5(b) is more likely to happen than the one in Figure 5(a). Therefore, most losses are diagnosed as congestion losses when the ratio bw bw <. This is illustrated in Figures 6, 7, and 8. x Cross Traffic TCP packets (a) One cross traffic packet on top of queue (b) Two consecutive cross traffic packet at top of queue Figure 5: Congestion losses due to buer overow 9

10 .8 56Kb 5Kb 48Kb.8 56Kb 5Kb 48Kb.6.6 Ac Aw wired bandwith (Kbits/s) (rw=5% rc=% rtt= ms qs= pkts) wired bandwith (Kbits/s) (rw=5% rc=% rtt= ms qs= pkts) (a) Accuracy A c for congestion losses (b) Accuracy A w for transmission error losses Figure 6: Accuracies A c and A w with r c = 5% and r w = % With r c = 5% and r w = %, in Figure 6, the overall error loss rate is 6%. When the ratio bw <, most losses are diagnosed as congestion losses. Therefore, A c is high and A w is low. When bw the ratio bw >, the accuracy A w for the wireless losses increases as the ratio increases. The bw accuracy A c is also high because the scheme in these conditions is ecient and very selective for wireless losses..8 56Kb 5Kb 48Kb.8 56Kb 5Kb 48Kb.6.6 Ac Aw wired bandwith (Kbits/s) (rw=% rc=5% rtt= ms qs= pkts) wired bandwith (Kbits/s) (rw=% rc=5% rtt= ms qs= pkts) (a) Accuracy A c for congestion losses (b) Accuracy A w for transmission error losses Figure 7: Accuracies A c and A w with r c = % and r w = 5% When r c = % and r w = 5%, in Figure 7, the plots are similar to the plots in gure 6. This show that the ratio bw bw and the overall loss rate are determinant. When r c = 5% and r w = 5%, the

11 .8 56Kb 5Kb 48Kb.8 56Kb 5Kb 48Kb.6.6 Ac Aw wired bandwith (Kbits/s) (rw=5% rc=5% rtt= ms qs= pkts) wired bandwith (Kbits/s) (rw=5% rc=5% rtt= ms qs= pkts) (a) Accuracy A c for congestion losses (b) Accuracy A w for transmission error losses Figure 8: Accuracies A c and A w with r c = 5% and r w = 5% plots are similar to the plots in gure 6. We observe a slight decrease in A w which can be explained by the small average congestion window size. The average congestion window size has a direct impact on the number of packets buered at the basestation..8 56Kb 5Kb 48Kb.8 56Kb 5Kb 48Kb.6.6 Ac Aw wired bandwith (Kbits/s) (rw=% rc=% rtt= ms qs= pkts) wired bandwith (Kbits/s) (rw=% rc=% rtt= ms qs= pkts) (a) Accuracy A c for congestion losses (b) Accuracy A w for transmission error losses Figure 9: Accuracies A c and A w with r c = % and r w = % Finally, when r c = % and r w = %, the overall loss rate is the same as the overall loss rate in gure 6 and gure 7. Note that the plots are very similar, conrming that the overall loss rate is determinant.

12 4 TCP Performance Using our Scheme In this section, we evaluate through simulation the throughput improvement when using our scheme with TCP-Reno for long lived TCP connections. Thereafter, TCP-Reno modied by our scheme will be designated as TCP-Aware. We compare the performance of our scheme against the performance of an ideal TCP-Reno sender. The ideal TCP-Reno sender is a TCP sender which has articially a perfect knowledge of the real cause of a packet loss. 4. Simulation Model We use a topology similar to the one in Figure. The only dierence is that we have 4 xed hosts and 4 mobile hosts as depicted in Figure. There is one TCP connection between each pair of xed and mobile hosts. The 4 TCP connections share the link R! R with a cross trac issued by four T raffic=expoo [] agents which are exactly the same as in Section. The links between the router R and the mobile hosts are wireless. These links have a bandwidth bw which takes values 64 Kbits/s, 56 Kbits/s, Mbits/s, Mbits/s, and 8 Mbits/s. This link has a transmission error rate r w which will take the values from %, %, and 5%. The bandwidth bw takes values 64 Kbits/s, 56 Kbits/s, Mbits/s, Mbits/s, and 8 Mbits/s. The propagation delay takes the value ms such that the round trip propagation time takes a value of 5 ms. This represents a typical value of round trip propagation time on WANs. Fixed Hosts ( TCP Senders) bw, δ bw, δ CBR - CBR (Cross traffic) Router R qs bw, δ Router R bw, ms Mobile Hosts (TCP Receivers) bw, ms SINK - SINK Figure : ns Simulation Model 4. Methodology The objective is to set 4 TCP connections between respectively 4 xed hosts and 4 mobile hosts. These connections have to contend against each other and against the random cross trac to share the link R?! R. This contention will generate congestion losses. We study the performance of our scheme under light (r c =? %), mild (r c =? 4%) and heavy (r c 5%) congestion. We set the level of congestion by choosing the rate pkr of the random

13 sources simulated by the T raf f ic=expoo agents. We set the rate pkr of the random sources such that the overall random cross trac represents % (light congestion), 6% (mild congestion), or % (heavy congestion) of the bandwith of the bottleneck link R?! R. For each set of parameters r w, bw, and bw, we start the random sources for a warm-up period of at least seconds before opening the TCP connections. Each TCP connection starts independently and randomly between to milliseconds after the end of the warm-up period. Each TCP connection lasts between to seconds depending on the values of the bandwidths bw and bw. We perform such simulations for TCP-Reno, TCP-Aware, and Ideal TCP-Reno. For each version of TCP (TCP-Reno, TCP-Aware and Ideal TCP-Reno), we compute the average throughput of all connections over all simulations. In this paper, we report the ratio R A of the average throughput obtained using TCP-Aware over the throughtput obtained with TCP Reno. This ratio denotes the improvement of our scheme to TCP performance. In order to evaluate this improvement, we compute also the ratio R I of the average throughput obtained using Ideal TCP-Reno over the throughtput obtained using TCP Reno. The ideal TCP-Reno sender has perfect knowledge of the cause of a loss. Therefore, the ratio R I represents the highest improvement we can bring to TCP-Reno by distinguishing perfectly congestion losses from wireless transmission losses. The next section presents our results. 4. Simulation Results Objective of our simulation experiments is two-fold: (a) determine the magnitudes of improvement achieved using TCP-Aware and Ideal TCP-Reno through R A and R I, and (b) determine the variations in these metrics R A and R I as a function of network parameters (such as bw, bw, r w and the level of congestion). We present three sets of results where the wireless transmission error rate r w is held constant and takes respectively the values %, %, and 5%. We present for each set where r w is held constant three plots. For each plot, we held the congestion level constant. The congestion level may be light (r c =? %), mild (r c =? 4%), or heavy (r c 5%). 4.. Wireless Transmission Error Loss r w = % We present in this section the set of results where the wireless transmission error rate r w is held constant at %. For each graph, the x-axis represents the wired bandwidth bw. Each graph displays 4 plots : two plots for R A (TCP-Aware) where bw is respectively held constant at 56 Kbits/s and Mbits/s two plots for R I Mbits/s (Ideal TCP-Reno) where bw is respectively held constant at 56 Kbits/s and

14 ..5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s..5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s..5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% light congestion rtt=48 ms).95 /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% mild congestion rtt=48 ms).95 /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% heavy congestion rtt=48 ms) Figure : Performance improvement with r w = % for light, mild, and heavy congestion congestion. From left to right for all gures, the plots correspond respectively to light, mild, and heavy We observe that the improvement is marginal for both TCP-Aware and Ideal TCP-Reno senders when we have light to mild congestion. There is no improvement when congestion is heavy. Two factors are determinant in the performance improvement we can expect from the Ideal TCP-Reno. The rst factor is the ratio rw. The second factor is the bandwidth-delay product. rc An approximation of the long range throughput derived in [, ] can help explaining qualitatively the results and the general trends of our results. A simpler derivation of this approximation can be found in []. The long range throughtput T can be approximated as T = MSS RT T C p p where MSS is the maximum segment size, RT T is the round trip time (assumed constant), p is a constant probability of random loss, and C is a constant. This approximation is based on the fact that the congestion window size is halved after every packet loss, neglecting the details of TCP data recovery and retransmission. If a TCP connection experiences congestion losses with rate r c wireless transmission losses with rate r w, its long range throughtput T T CP with T T CP = M SS RT T C prc+rw. and may be approximated Now, the Ideal TCP-Reno halves its congestion window only when a congestion loss occurs. Therefore, its long range throughput T I deal may be approximated as T I deal = M SS RT T the ratio of improvement may be approximated as T I deal T T CP = r + r w r c C p. Therefore rc This approximation shows that the improvement increases with r w and decreases with r c. The ratio rw is a determinant factor of the improvement we can expect from an Ideal TCP-Reno. Recall rc 4

15 that in Figure, the wireless transmission error loss rate r w is held constant at %. As congestion increases (from left to right on Figure ), the improvement induced by Ideal TCP-Reno decreases. Moreover, when congestion is low (r c =? %), the approximation implies that the improvement should be between. and.4. The plots on Figure for low congestion shows that the improvement remains below these values. This is due to the fact that the bandwidth-delay product is low. The second factor is the bandwidth delay product. When the bandwidth-delay product is small, the Ideal TCP-Reno cannot bring signicative performance improvement. To illustrate this, let us consider the plots for which bw = 56Kbits=s. For all the plots presented, the round trip propagtion delay is 48 ms. Therefore, the bandwidth-delay product is at most 56 bytes (' :5 packet). There is no need of high window size to keep the pipe full. However, a high window size may increase the likelihood of triggering the fast-retransmit mechanism (avoiding time-outs). When bw = 56 Kbits/s, we observe on all plots that the improvement due to Ideal TCP-Reno is low (less that.5). When bw = Mbits/s, the improvement in throughput may be signicative (more than ). Note that our scheme TCP-Aware perfoms similarly to Ideal TCP-Reno. 4.. Wireless Transmission Error Loss r w = % We present in this section the set of results where the wireless transmission error rate r w is held constant at %. The graphs are similar to those presented in Section 4.., except that the y-axis scale is dierent. In these graphs, the scale reaches.6. As expected from the throughput approximation, Figure shows that the improvement is more signicative for r w = % than for r w = %. The improvement remains lower for low bandwidthdelay product (bw = 56 Kbits/s). On Figure, congestion increase from light (left graph) to heavy (right graph). As congestion increases, the improvement due to Ideal TCP-Reno decreases. When congestion is heavy, the improvement is marginal. With r w = %, our scheme TCP-Aware still perfoms similarly to Ideal TCP-Reno..5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s.5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s.5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% light congestion rtt=48 ms) /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% mild congestion rtt=48 ms) /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% heavy congestion rtt=48 ms) Figure : Performance improvement with r w = % for light, mild, and heavy congestion 5

16 4.. Wireless Transmission Error Loss r w = 5% We present in this section the set of results where the wireless transmission error rate r w constant at 5%. The graphs are similar to those presented in Section 4... is held The trends observed in Section 4.. and in Section 4.. are conrmed. However, Ideal TCP- Reno performs better than our scheme when congestion is light and bw is high ( Mbits and 8 Mbits). This is due to the fact that the accuracy of our scheme is not perfect and the wireless transmission loss error rate r w is high. As r w increases, it is obvious that the discrepancy between our scheme and Ideal TCP-Reno will increase (under light or mild congestion). However, the improvement due to our scheme remains signicative..5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s.5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s.5 RI (Ideal) bw=mb/s RA bw=mb/s RI (Ideal) bw=56kb/s RA bw=56kb/s /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% light congestion rtt=48 ms) /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% mild congestion rtt=48 ms) /s 56Kb/s Mb/s Mb/s 8Mb/s wired bandwidth (Kbits/s) (rw=% heavy congestion rtt=48 ms) Figure : Performance improvement with r w = 5% for light, mild, and heavy congestion 5 Conclusion and further work We studied a simple heuristic to distinguish between packet losses due to wireless transmission error from packet losses due to congestion. This heuristic works well when, rst, the last hop for the connection is wireless, second, the bandwidth of the wireless link is much smaller than the bandwidth of the wired link and nally, the overall loss rate is small. We added our scheme to TCP-Reno to distinguish the two types of losses. We compared through simulations the performance of our scheme against Ideal TCP-Reno which can articially diagnose perfectly the cause of a loss. The preliminary results with 4 TCP connections sharing a common bottleneck show that our scheme performs similarly to Ideal TCP-Reno except when the wireless error loss is high and congestion is not heavy. Now, we plan to use this heuristic with TCP and study the impact on the TCP performance under dierent scenarios : (dierent round trip time, queue size, etc... ). We aim to rene this heuristic to work with a shared wireless link with many connections competing for the wireless link as in Wavelan. References [] \VINT project U.C. berkeley/lbnl, ns:network simulator." 6

17 [] M. Allman and D. Glover, \Enhancing TCP over satellite channels using standard mechanisms," May 998. INTERNET DRAFT, mallman/papers. [] M. Allman, D. Glover, J. Griner, K. Scott, J. Touch, and D. Tran, \Ongoing TCP research related to satellites," Mar INTERNET DRAFT, mallman/papers. [4] A. Bakre and B. Badrinath, \I-TCP: Indirect TCP for mobile hosts," in Proc. 5th International Conf. on Distributed Computing Systems (ICDCS), May 995. [5] H. Balakrishnan, V. Padmanabhan, S. Seshan, and R. Katz, \A comparison of mechanisms for improving TCP performance over wireless links," in ACM SIGCOMM'96, Aug [6] H. Balakrishnan, S. Seshan, and R. Katz, \Improving reliable transport and hando performance in cellular wireless networks," ACM Wireless Networks, vol., Dec [7] S. Biaz and N. Vaidya, \Sender-based heuristics for distinguishing congestion losses from wireless transmission losses," Tech. Rep. TR98-, CS Dept., Texas A&M University, June 998. [8] R. Caceres and L. Iftode, \Improving the performance of reliable transport protocols in mobile computing environments," IEEE JSAC Special issue on Mobile Computing Networks, vol., June 995. [9] A. DeSimone, M. Chuah, and O. Yue, \Throughput performance of tranport-layer protocols over wireless lans," in Proc. Globecom '9, Dec. 99. [] V. Jacobson, \Congestion avoidance and control," in ACM SIGCOMM'88, pp. 4{9, Aug [] T. Lakshman and U. Madhow, \The performance of tcp/ip for networks with high bandwidthdelay products and random loss," IEEE/ACM Transactions on Networking, vol. 5, June 997. [] M. Mathis, J. Semke, J. Mahdavi, and T. Ott, \The macroscopic behavior of the tcp congestion avoidance algorithm," Computer Communication Review, vol. 7, July 997. [] T. J. Ott, J. Kemperman, and M. Mathis, \The stationary behavior of ideal tcp congestion avoidance," Aug In progress,obtain via pub/tjo/tcpwindow.ps using anonymous ftp to ftp.bellcore.com. [4] J. Postel, \Transmission control protocol," Sept RFC 79. [5] R. Yavatkar and N. Bhagwat, \Improving end-to-end performance of TCP over mobile internetworks," in Workshop on Mobile Computing Systems and Applications, Dec