Challenges to reliable data transport over heterogeneous wireless networks

The Transmission Control Protocol (TCP) is the de facto standard for reliable data transmission in the Internet today.;While TCP adapts well to network congestion, it does not adequately handle the vagaries of wireless media. In this thesis, we address these challenges in detail and design solutions to them. These solutions incorporate link-layer techniques as well as enhancements to TCP at the sender and receiver.;The adverse effects of wireless bit-errors on TCP performance arise primarily because TCP misinterprets wireless losses as being due to congestion. We present the design and implementation of a novel protocol, called the Berkeley Snoop Protocol, that exploits cross-layer protocol optimizations to improve performance. In addition, we design a mechanism called Explicit Loss Notification that can successfully distinguish between congestion and channel error-induced packet losses to substantially enhance end-to-end performance. These mechanisms have been demonstrated to provide performance improvements between 100% and 2000% across a range of bit-error rates in various cellular topologies in a Lucent WaveLAN-based wireless LAN, for data transfer to and from mobile hosts and when wireless transit links are present. Asymmetric networks pose a challenge to feedback-based protocols like TCP, because the characteristics of the reverse path used for acknowledgments (ACKs) have the potential to seriously affect forward performance. We classify asymmetry with respect to TCP performance into several categories, such as bandwidth asymmetry, latency asymmetry, loss-rate asymmetry, etc. We design several end-to-end and transport-aware link-layer solutions, such as ACK filtering, TCP sender adaptation and ACK reconstruction to obtain significantly better performance over such networks. These techniques lead to performance improvements for Internet access over packet radio networks such as Metricom's Ricochet network and for bandwidth-asymmetric networks such as Hybrid's wireless cable network.;Low channel bandwidths are common in many wireless networks as well as on many Internet paths today. This, and the short connections common in many Web transfers, lead to TCP windows being very small. This greatly affects its data-driven loss recovery mechanism and causes a large number of timeouts. In addition to implementing and evaluating the performance of TCP Selective Acknowledgments, we analyze data from packet-level traces of the Web server of the Atlanta Olympic Games to motivate enhancements to TCP's loss recovery mechanism. Our enhancements lead to an Enhanced Recovery scheme for TCP that greatly reduces the number of timeouts when windows are small.;In this research, we learn some general lessons about protocol design for heterogeneous networks. Specifically, we demonstrate the significant performance benefits of cross-layer protocol optimizations, of soft-state agents in the network infrastructure, and the advantages of data-driven loss recovery over timer-driven ones for better performance. (Abstract shortened by UMI.).