Explicit Loss Notification and Wireless Web Performance

Hari Balakrishnan and Randy H. Katz

Abstract

This paper describes our experiences with improving TCP and Web performance over a WaveLAN-based wireless network. In previous work, we analyzed the problems to TCP performance in error-prone wireless networks and designed the Berkeley Snoop protocol to significantly improve performance over error-prone wireless links [7, 5].

Most efforts to date have focused on improving performance for transfers to a mobile host. We present a novel protocol based on Explicit Loss Notification (ELN) to improve performance when the mobile host is the TCP sender, a situation that is becoming increasingly common. Then, we use experimental packet traces of wireless errors from a production wireless network to derive an empirical model of channel errors. We use this to evaluate the performance of TCP Reno, TCP Selective Acknowledgments [17] and the Snoop protocol for Web workloads to mobile hosts. We also discuss the scaling behavior of the Snoop protocol and reflect on some general lessons we have learned about efficient protocol design for reliable wireless transport.

1. Introduction

Wireless technologies are playing an increasingly prominent role in the global Internet infrastructure. They are ideal as Internet access technologies, providing a convenient and cheap solution to the "last mile" problem. However, reliable transport protocols such as the Transmission Control Protocol (TCP) [22, 25], used by popular applications like the World Wide Web, file transfer, electronic mail, and interactive remote terminal applications, show greatly degraded performance over wireless networks. Over the past many years, TCP has been tuned for traditional networks comprising wired links and stationary hosts. It assumes congestion in the network to be the primary cause for packet losses and unusual delays, and adapts to it. The TCP receiver sends cumulative acknowledgments (ACKs) for successfully received segments, which the sender uses to determine which segments have been successfully received. The sender identifies the loss of a packet either by the arrival of several duplicate cumulative ACKs, triggering a fast retransmission, or by the absence of an ACK for a timeout interval equal to the sum of the smoothed round-trip delay and four times its mean deviation. TCP reacts to packet losses by retransmitting missing data, and simultaneously invoking congestion control by reducing its transmission (congestion) window size and backing off its retransmission timer. These measures reduce the level of congestion on the intermediate links.

Unfortunately, when packets are lost for reasons other than congestion, these measures result in an unnecessary reduction in end-to-end throughput and hence, in sub-optimal performance. Communication over wireless links is often characterized by high bit-error rates due to channel fading, noise or interference, and intermittent connectivity due to handoffs. TCP performance in such networks suffers from significant throughput degradation and very high interactive delays because the sender misinterprets corruption for congestion [see e.g., 8].

Recently, several schemes have been proposed to alleviate the effects of non-congestion-related losses on TCP performance over error-prone networks [3, 7, 26]. In previous work [5], we compared many existing schemes and argued for transport-aware link protocols to improve performance. The end result of this work was a more adaptive TCP/IP protocol architecture that adapted not only to network congestion, but also to error-prone wireless links.

However, most design and measurement efforts to date have focused on the (common) case of data transfers to a mobile host. The case of a mobile host transmitting data over a first-hop wireless link to hosts on the wired Internet has largely been ignored. Such access scenarios are becoming increasingly common in the Internet; for example, researchers at Daimler Benz are working on a prototype of a Mercedes car that runs a Web server, which disseminates information over wireless and wired links about the status of the vehicle to users on the Internet [13].

In this paper, we present a novel protocol based on Explicit Loss Notification (ELN) to improve transport performance when the mobile host is the TCP sender, a situation that is becoming increasingly common in many wireless access networks ( See . Transfers from a Mobile Host ). Then, we obtain experimental packet traces of wireless errors from a production outdoor wireless network deployed as part of the Reinas environmental monitoring project at UC Santa Cruz [9] and derive an empirical model of channel errors based on this data ( See . Modeling Wireless Errors ). We use this to evaluate the performance of TCP Reno, TCP SACK [17] and the Snoop protocol for an empirically derived Web workload to mobile hosts ( See . Performance: Fixed to Mobile Host ). Finally, we reflect on some lessons learned about improving wireless transport performance in the face of wireless bit-errors ( See . Discussion ) and conclude with a summary and directions for future work ( See . Summary ).

2. Background

In this section, we summarize some protocols and transport enhancements that have been proposed to improve the performance of TCP over wireless links.

• Link-layer protocols: There have been several proposals for reliable link-layer protocols to improve wireless performance [1, 14, 19]. These typically use forward error correction (FEC) or retransmissions (ARQ) to improve performance.

The main advantage of employing a link-layer protocol for loss recovery is that it fits naturally into the layered structure of network protocols and achieve local reliability. However, there are several occasions when transport performance does not improve when such protocols are used. There are three chief modes of adverse interaction that could arise between independent reliable transport and link layers:

1. Timer interactions: Independently set timers at both layers could trigger at the same time, leading to redundant retransmissions at both layers and degraded performance [10]. While this is a concern in general, TCP does not suffer from this problem because of the coarse and conservative timeout intervals in practice.
2. Fast retransmission interactions: This arises when a link layer protocol achieves reliability by local retransmissions, but does not preserve the in-order sequential delivery of TCP segments to the receiver. Then, although local recovery occurs, the receipt of later segments causes duplicate ACKs from the receiver as well, leading to redundant sender fast retransmissions, sender window reduction, and reduced throughput [5].
3. Large round-trip variations: If a link layer protocol is designed to be overly robust and attempt, for example, to provide a TCP-like service, it often result in long latencies and large round-trip time deviations at the TCP sender. This leads to long and conservative timeouts when congestion-related losses occur on the path. Other problems that arise include a large number of redundant retransmissions if a sender timeout occurs for a wireless loss [15], a problem observed in GSM networks running the RLP protocol [18].
4. Split connection protocols [3, 26]: Split connection protocols split each TCP connection between a sender and receiver into two separate connections at the base station -- one TCP connection between the sender and the base station, and the other between the base station and the receiver. Over the wireless hop, a specialized protocol tuned to the wireless environment may be used. In [26], the authors propose two protocols -- one in which the wireless hop uses TCP, and another in which the wireless hop uses a selective repeat protocol (SRP) on top of UDP. They study the impact of handoffs on performance and conclude that they obtain no significant advantage by using SRP instead of TCP over the wireless connection in their experiments. However, our experiments demonstrate benefits in using a simple SACK scheme with TCP over the wireless connection.

Indirect-TCP [3] is a split-connection solution that uses standard TCP for its connection over the wireless link. Like other split-connection proposals, it attempts to separate loss recovery over the wireless link from that across the wireline network, thereby shielding the original TCP sender from the wireless link. However, as our experiments indicate, the choice of TCP over the wireless link results in several performance problems. Since TCP is not well-tuned for the lossy link, the TCP sender of the wireless connection often times out, causing the original sender to stall. In addition, every packet incurs the overhead of going through TCP protocol processing twice at the base station (as compared to zero times for a non-split-connection approach), although extra copies are avoided by an efficient kernel implementation. Another disadvantage of split connections is that the end-to-end semantics of TCP ACKs is violated, since ACKs to packets can now reach the source even before the packets actually reach the mobile host. This makes the system vulnerable to base station crashes. Also, since split-connection protocols maintain a significant amount of state at the base station per TCP connection, handoff procedures tend to be complicated and slow [2].

• Snoop Protocol [7]: The snoop protocol introduces a module, called the snoop agent , at the base station. The agent monitors every packet that passes through the TCP connection in both directions and maintains a cache of TCP segments sent across the link that have not yet been acknowledged by the receiver. A packet loss is detected by the arrival of a small number of duplicate ACKs from the receiver or by a local timeout. The snoop agent retransmits the lost packet if it has it cached and suppresses the duplicate ACKs. Thus, the snoop protocol is a transport-aware link protocol.

The main advantage of this approach is that it suppresses duplicate ACKs for TCP segments lost and retransmitted locally, thereby avoiding unnecessary fast retransmissions and congestion control invocations by the sender. The per-connection state maintained by the snoop agent at the base station is soft , and is not essential for correctness. A detailed description of the Snoop protocol appears in [7].

• Selective Acknowledgments : Since TCP Reno uses a cumulative ACK scheme, it often does not provide the sender with sufficient information to recover quickly from multiple packet losses within a single transmission window. The SACK RFC [17] proposes that each ACK contain information about up to three non-contiguous blocks of data that have been received successfully by the receiver. Each block of data is described by its starting and ending sequence number. Due to the limited number of blocks, it is best to inform the sender about the most recent blocks received. The RFC does not specify the sender behavior, except to require that standard TCP congestion control actions be performed when losses occur. SACKs help greatly when window sizes are large and multiple losses occur in a single transmission window [11].

3. Transfers from a Mobile Host

We now proceed to discuss the case of data transfer from a mobile host. While this has conventionally been a less common scenario for wireless information access, it is rapidly gaining in importance. There is an increasing number of Web and other information servers that serve data across first-hop wireless links [13, 9]. A representative topology of this scenario is shown in See Topology for data transfer from a mobile host. .

It might be tempting to come to the conclusion that a protocol very similar to the snoop protocol described in See . Background and [7] for data transfer to the mobile host can be used in this case as well. In particular, one of the important features of that case was that it required changes only to the base station and not to any other entities in the network. However, it is unlikely, if not impossible, that a protocol with modifications made only at the base station can substantially improve the end-to-end performance of reliable bulk data transfers from the mobile host to other hosts on the network, while preserving the precise semantics of TCP ACKs. For example, caching packets at the base station and retransmitting them as necessary is not useful, since most of the packet losses will be from the mobile host to the base station. Running a Snoop agent at the mobile host will be inappropriate because the same duplicate ACKs signify both corruption and congestion losses. There is no way for the mobile sender to know if the loss of a packet happened on the wireless link or elsewhere in the network due to congestion.

There are at least two ways in which the basic Snoop scheme can be enhanced to achieve good performance for this case -- the first involves the use of negative ACKs, and the second the use of Explicit Loss Notifications (ELN). We first describe the first approach, and then the second, which we believe to be an elegant and simple solution to the problem. We also show how the algorithm naturally generalizes to the case when a cellular wireless link is the transit link bridging two or more wired networks.

4. Modeling Wireless Errors

Wireless testbed: We used the experimental network deployed in the Reinas Environmental Monitoring project at the University of California, Santa Cruz to obtain experimental data. This network has a number of point-to-point long-haul WaveLAN links as well as lower-bandwidth RF links. The Snoop protocol has been in daily production use in this network since 1996, during which time we have been able to perform numerous performance experiments and obtain error traces of real network behavior. We use ns (http://www-mash.cs.berkeley.edu/ns) as our simulator.

Analytic state models: The simplest error models are one-state models that are characterized by a single rate parameter and a statistical distribution that determines the error rate. For example, a one-state exponential error model generates errors at a certain rate based on an exponential distribution in the appropriate domain (packet, byte, or time). More complex multi-state error models are implemented by composing one-state models. Such models have been extensively studied in the literature and are characterized by a number of states organized as a Markov chain. Each state is a one-state error model governed by a certain rate and distribution, with transitions occurring between states according to a transition matrix of probabilities.

Empirical distribution-based models: While analytic state models are sound in theory, it is hard to determine in practice what the parameters of these models should be. The process of fitting experimental channel error data in this framework is not straightforward. While using simple one-state models enables us to quantify the relative performance as a function of a base error rate, it does not give us any insight into what "typical" performance might be like in a real wireless environment under realistic error conditions. This motivates us to develop an error model based on actual traces from a wireless environment and use that model to evaluate our protocols. This section describes how we do this.

Our methodology for constructing a realistic wireless error model is motivated by the approach presented in [20], where the authors analyze in-building traces of wireless traffic and construct empirical models for them. Rather than use the traces they collected, we decided to obtain fresh ones. This is because we could validate the performance observed in simulation using these error traces by comparing these results to actual implementation performance.

To develop an empirical error model of the channel, we proceed as follows. We transmit a constant rate UDP stream at rate R bps, with a fixed packet size, S bits. We ensure that the transmitted rate is smaller than the sustainable link rate, and that the size of the packets is large enough that the number of packets per second does not overwhelm the receiving processor, which needs to take an interrupt for every reception. Then, we send a large number of packets, N , with an application-level sequence number on each packet. Thus, the receiver can now determine which packets have arrived correctly, and which have not. UDP checksum processing at the receiver is enabled, so any packet whose lower-layer headers are correct but whose payload is corrupted will be discarded. Of course, no packets are retransmitted.

Using this data, the receiver constructs histograms of the lengths of error and error-free sequences. We obtain the distribution of the number of successive packets that were successfully received, as well as the number of missing ones once there was a loss. At this point, this histogram is in terms of packets, rather than time. However, the underlying characteristics of the channel's error properties are temporal, related to time-dependent factors like the average fade duration and duration of inter-symbol interference due to multi-path effects. Hence, we need to convert our packet-based distribution into a time-based one, which we achieve by multiplying the number of packets in any histogram bin by the packet transmission time, S/R .

We used the experimental network deployed in the Reinas Environmental Monitoring project at the University of California, Santa Cruz to obtain our data. This network has a number of point-to-point long-haul WaveLAN links as well as lower-bandwidth RF links.

We performed eight separate experiments using this approach, with N = 100,000 packets in each experiment, packet size S = 8000 bits and R = 1 Mbps. The resulting error and error-free temporal distributions are shown in See Cumulative Distribution Functions (CDF) of empirically measured error and error-free durations. The CDF was computed over a duration of 100,000 packets sent over 800 seconds. . In the experiments that follow, we picked one of these data sets. In this data set, 10,591 packets were lost due to wireless errors, a gross loss rate of 10.6%. While this might seem high, it is in fact quite representative of the quality of the outdoor wireless links in the Reinas network. The typical loss rates in the network vary from 1% to over 30%, depending on external factors and environmental conditions.

The cumulative distribution functions (CDF) of error and error-free durations are used to generate bit errors in the following manner. There are two linked states, corresponding to the error-free and error durations, that the error model transitions between. It starts in the error-free state where no errors occur, and stays there for an amount of time determined by a lookup into the corresponding CDF. After this time has elapsed, there is an automatic transition to the error state, where all incoming packets are corrupted. The amount of time spent in this state is, as before, determined by the corresponding CDF. Thus, this empirical error model shown in See Conventional multi-state analytic error models (left) and empirical distribution-based models (right). In analytic models, the different states correspond to different base error rates, and the arc probabilities (p) govern state transitions. Our empirical models are easy to fit to observed experimental data and contain two states -- error-free and error, which have CDF's associated with them that determine state occupancy. State transitions are automatic when this time elapses. is a two-state Markov chain, but is quite different from the analytic models. Here, the states themselves correspond to time, with automatic state transitions. The analytic models are richer in that each state has a different base error rate, with explicit transition probabilities between states. However, it is hard to reconcile experimental data to meaningful parameters that can be used in these models.

5. Performance: Fixed to Mobile Host

6. Discussion

This section reflects on some key lessons we have learnt from our experience with reliable transport over error-prone wireless networks.

• Cross-layer protocol optimizations: For data transfer to a mobile host, the Snoop protocol uses a transport-aware link protocol between the base station and mobile host to achieve good end-to-end performance for bulk and Web-like transfers. For data transfer from a mobile host, it uses ELN to make the transport protocol link-aware. This approach violates the conventional layered model used in network protocol design, where there is no explicit control information transferred between higher and lower layers of the protocol stack. The benefits of our solutions in both simulation and real deployments show that better protocol adaptation is enabled by making different layers of the protocol stack more aware of what the other layers are doing and also of what the underlying channel is doing. Such an approach achieves superior end-to-end performance because it explicitly prevents adverse interactions that could occur between independent protocol
• The end-to-end argument: The design of the snoop agent and protocol has been strongly influenced and guided by the end-to-end argument and design principle propounded by Saltzer, Reed and Clark [23]. The end-to-end argument is a design principle that guides the placement of services in a system. It states that a service should be implemented in a subsystem only if the service can be completely implemented in that subsystem, or if by partially implementing the service in the subsystem, it substantially improves end-to-end performance. This is precisely the philosophy that guided the design of the Snoop protocol and its approach toward local recovery. In contrast to split-connection approaches, the Snoop protocol does not violate the end-to-end semantics of TCP and is not vulnerable to base station failures. The good performance of some of the split connection approaches derives not from their spoofing acks and breaking the connection, but from their running a wireless-optimized protocol over the noisy link as shown in [5]. And, it is possible to obtain the same or higher benefit using the Snoop approach, without any change in end-to-end semantics.
• Soft-state agents in the network infrastructure: The performance benefits of deploying a soft-state snoop agent at the base station are significant. In general, we believe that this is a way in which network heterogeneity and associated performance problems can be solved, using agents in the network infrastructure. The snoop agent state is soft, because it is not crucial to the correct end-to-end functioning of the system and because it is periodically and automatically refreshed upon the arrival of data and ACKs. Any mistake in the state does not compromise the correctness of the connection and is automatically rectified when new messages arrive. This soft-state approach also enables low handoff latencies with little connection disruption, since the contents of the snoop cache can be mirrored using multicast from the home agent [24].
• Data-driven retransmissions: Reliable protocols can recover from losses in two ways -- using data-driven retransmissions, which infer from the receipt of later data that an earlier packet is missing, and timeouts, which are triggered when the sender does not receive an ACK within a certain period of time. Data-driven retransmissions recover from losses with little disruption in the stream, while timeouts necessarily have to be conservative to ensure that data that is delayed but still in transit does not get prematurely retransmitted [27]. In the Snoop protocol, we explicitly minimize the dependence on local timers by performing data-driven retransmissions as far as possible.
• Higher priority local retransmissions: Retransmitting packets at a higher priority improves performance at all error rates. This enables retransmitted packets to reach the mobile host sooner, reducing the number of duplicate ACKs and leading to improved throughput. In the common case when the wireless link is the connection's bandwidth bottleneck, the first new ACK generated after a higher-priority retransmission is usually less than what it would be without such prioritization. This implies that the number of segments transmitted by the sender in a burst is smaller in the former case, because the sliding window shifts by a smaller amount. This in turn reduces the degree of congestion in the network by reducing the degree of burstiness, leading to improved overall performance.

Higher priority local retransmissions also allow us to avoid unnecessary local timeouts in the following way. When a retransmission is scheduled, the snoop agent precisely knows what the last in-sequence transmission was. Hence, it knows what the next expected new ACK should be, assuming no further losses occur. Now, if the next new ACK is smaller than the next expected one, a retransmission occurs without a timeout. This helps in recovering from multiple losses in a single window without a timeout.

• Ineffectiveness of end-to-end TCP SACK: An interesting observation concerns the lack of effectiveness of purely end-to-end TCP SACK in the wireless environment, despite the sender's ability to recover quickly from multiple losses in a window when SACK information is available. Initially, this was surprising to us because of SACK's benefits in burst loss situations. However, upon closer analysis [5], we found that even with SACK, the sender ended up incurring numerous coarse timeouts. With a large enough window, burst losses over the wireless link are indeed recovered quickly, but the sender still performs congestion control and reduces its window. The end result is a small average window size, with only a small number of outstanding segments. As a result, there is not enough information for the sender to distinguish between reordered and lost packets, necessitating coarse timeouts. A solution to this problem, based on enhancing TCP's loss recovery algorithm in [6, 4].
• Sharing loss recovery information: When there are multiple connections outstanding to the same mobile host, the snoop agent can perform loss recovery in an integrated way by effectively sharing ACK information between them. We can do this by treating the different connections as one integrated stream and by keeping track of the relative transmission order between packets belonging to different connections. Then, when an ACK arrives on a connection that acknowledged a data item that was transmitted after a currently unacknowledged segment belonging to another connection, we treat this as a later ACK indication for the latter connection. When the number of later ACKs reaches a threshold, we retransmit the earliest missing segment, independent of the connection(s) that generated this later ACK information. Results of this integrated loss recovery scheme in the context of TCP are presented in [6, 21].
• End-to-end security: An important issue concerning the viability of approaches like the Snoop protocol and split-connection approaches is their interaction with end-to-end IP security. If a connection has end-to-end security enabled, its entire IP payload, including the transport header, would be encrypted and therefore unreadable by network agents like the snoop agent. One possible solution is for the source to not encrypt part of the TCP header -- this would not compromise the security model provided by IP Security and would overcome the problem. Another (less attractive) possibility is to include the base station as part of the mobile host's trust infrastructure and allow it to decrypt packets destined for the mobile host. Developing a standardized way to solve this problem without compromising end-to-end security is currently gaining momentum in the IETF.

7. Summary

In this paper, we presented a novel protocol based on Explicit Loss Notification (ELN) to improve transport performance when the mobile host is the TCP sender, a situation that is becoming increasingly common [13]. Then, we obtained experimental packet traces of wireless errors from a production wireless network and derived an empirical model of channel errors based on this data. We used this to evaluate the performance of TCP Reno, TCP SACK and the Snoop protocol for Web workloads to mobile hosts. We discussed the scaling behavior of the Snoop protocol and reflected on some general lessons we have learned over the past two years about efficient protocol design for reliable wireless transport based on real-world deployments of the Snoop protocol.

Acknowledgments

We are grateful to Darrell Long, Patrick Mantey and Eric Rosen of U.C. Santa Cruz for continuous access to their WaveLAN-based Reinas wireless network. They were also very tolerant users of the Snoop protocol and suffered through crashes and downtimes during the early stages of deployment. We cannot thank them enough for their generous support that led to the Snoop implementation and deployment becoming extremely robust and tuned to perform well. We hope that our solutions continue to serve them well in the future.

We thank Suchitra Raman for several comments and suggestions that significantly improved the quality and presentation of this paper.

This work was supported by DARPA contract DAAB07-95-C-D154, a Segasoft research grant from the Okawa Foundation, the California MICRO program, Metricom Inc., Hybrid Networks Inc., and Hughes Aircraft Corp.

References

[1] E. Ayanoglu, S. Paul, T. F. LaPorta, K. K. Sabnani, and R. D. Gitlin. AIRMAIL: A Link-Layer Protocol for Wireless Networks. ACM ACM/Baltzer Wireless Networks Journal, 1:47-60, February 1995.

[2] A. Bakre and B. R. Badrinath. Handoff and System Support for Indirect TCP/IP. In Proc. Second Usenix Symp. on Mobile and Location-Independent Computing, April 1995.

[3] A. Bakre and B. R. Badrinath. I-TCP: Indirect TCP for Mobile Hosts. In Proc. 15th International Conf. on Distributed Computing Systems (ICDCS), May 1995.

[4] H. Balakrishnan. Reliable Data Transport over Heterogeneous Wireless Networks. PhD thesis, Univ. of California, Berkeley, 1998. In preparation.

[5] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R.H. Katz. A Comparison of Mechanisms for Improving TCP Performance over Wireless Links. IEEE/ACM Trans. on Networking, 5(6), December 1997.

[6] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, M. Stemm, and R.H. Katz. TCP Behavior of a Busy Web Server: Analysis and Improvements. In Proc. IEEE INFOCOM '98, March 1998. To appear.

[7] H. Balakrishnan, S. Seshan, and R.H. Katz. Improving Reliable Transport and Handoff Performance in Cellular Wireless Networks. ACM Wireless Networks, 1(4), December 1995.

[8] R. Caceres and L. Iftode. Improving the Performance of Reliable Transport Protocols in Mobile Computing Environments. IEEE Journal on Selected Areas in Communications, 13(5), June 1995.

[9] D. E. Long., et al. REINAS: A Real-time System for Managing Environmental Data. In Proc. Conf. on Software Engg. and Knowledge Engg., 1996.

[10] A. DeSimone, M. C. Chuah, and O. C. Yue. Throughput Performance of Transport-Layer Protocols over Wireless LANs. In Proc. Globecom '93, December 1993.

[11] K. Fall and S. Floyd. Simulation-based Comparisons of Tahoe, Reno, and Sack TCP. Computer Communications Review, 1996.

[12] R. Fielding, J. Gettys, J. Mogul, H. Frystyk, and T. Berners-Lee. Hypertext Transfer Protocol - HTTP/1.1, Jan 1997. RFC-2068.

[13] A. Jameel, M. Stuempfle, D. Jiang, and A. Fuchs. Web on Wheels: Toward Internet-Enabled Cars. IEEE Computer, January 1998.

[14] P. Karn. The Qualcomm CDMA Digital Cellular System. In Proc. 1993 USENIX Symp. on Mobile and Location-Independent Computing, pages 35-40, August 1993.

[15] R. Ludwig. Exploiting Local Information for End-to-End Congestion Control, 1998. In preparation.

[16] B. A. Mah. An Empirical Model of HTTP Network Traffic. In Proc. InfoComm '97, April 1997.

[17] Mathis, M. and Mahdavi, J. and Floyd, S. and Romanow, A. TCP Selective Acknowledgment Options, 1996. RFC-2018.

[18] M. Mouly and M.-B. Pautet. The GSM System for Mobile Communications. Cell & Sys, 1992.

[19] S. Nanda, R. Ejzak, and B. T. Doshi. A Retransmission Scheme for Circuit-Mode Data on Wireless Links. IEEE Journal on Selected Areas in Communications, 12(8), October 1994.

[20] G. T. Nguyen, R. H. Katz, B. D. Noble, and M. Satyanarayanan. A Trace-based Approach for Modeling Wireless Channel Behavior. In Proc. Winter Simulation Conference, Dec 1996.

[21] V. N. Padmanabhan. Addressing the Challenges of Web Data Transport. PhD thesis, Univ. of California, Berkeley, 1998. In preparation.

[22] J. B. Postel. Transmission Control Protocol. Information Sciences Institute, Marina del Rey, CA, September 1981. RFC-793.

[23] J. H. Saltzer, D. P. Reed, and D. D. Clark. End-to-end Arguments in System Design. ACM Transactions on Computer Systems, 2:277-288, Nov 1984.

[24] S. Seshan, H. Balakrishnan, and R. H. Katz. Handoffs in Cellular Wireless Networks: The Daedalus Implementation and Experience. Kluwer Journal on Wireless Personal Communications, January 1997.

[25] W. R. Stevens. TCP/IP Illustrated, Volume 1. Addison-Wesley, Reading, MA, Nov 1994.

[26] R. Yavatkar and N. Bhagwat. Improving End-to-End Performance of TCP over Mobile Internetworks. In Mobile 94 Workshop on Mobile Computing Systems and Applications, December 1994.

[27] L. Zhang. Why TCP Timers Don't Work Well. In Proc. ACM SIGCOMM, pages 397-405, August 1986.