In just a few years since its inception, the World Wide Web has grown to be the most dominant application in the Internet. In large measure, this rapid growth is due to the Web's convenient point-and-click interface and its appealing graphical content. Since Web browsing is an interactive activity, minimizing user-perceived latency is an important goal. However, layering Web data transport on top of the TCP protocol poses several challenges to achieving this goal.
First, the transmission of a Web page from a server to a client involves the transfer of multiple distinct components, each in itself of some value to the user. To minimize user-perceived latency, it is desirable to transfer the components concurrently. TCP provides an ordered byte-stream abstraction with no mechanism to demarcate sub-streams. If a separate TCP connection is used for each component, as with HTTP/1.0, uncoordinated competition among the connections could exacerbate congestion, packet loss, unfairness, and latency.
Second, Web data transfers happen in relatively short bursts, with intervening idle periods. It is difficult to utilize bandwidth effectively during a short burst because discovering how much bandwidth is available requires time. Latency suffers as a consequence.
To address these problems, we first developed a new connection abstraction for HTTP, called persistent-connection HTTP (P-HTTP). The key ideas are to share a persistent TCP connection for multiple Web page components and to pipeline the transfers of these components to reduce latency. These ideas, developed by us in 1994, have been adopted by the HTTP/1.1 protocol. The main drawback of P-HTTP, though, is that the persistent TCP connection imposes a linear ordering on the Web page components, which are inherently independent.
This drawback of P-HTTP led us to develop a comprehensive solution, which has two components. The first component, TCP session, decouples TCP's ordered byte-stream service abstraction from its congestion control and loss recovery mechanisms. It integrates the latter mechanisms across the set of concurrent connections between a pair of hosts, thereby combining the flexibility of separate connections with the performance efficiency of a shared connection. This integration decreases download time by up to a factor of ten compared to HTTP/1.0 layered on standard TCP. TCP session does not alter TCP's messaging semantics, so deployment only involves local changes at the sender.
The second component of our solution, TCP fast start, improves bandwidth utilization for short transfers by reusing information about the network conditions cached in the recent past. To avoid adverse effects in case the cached information is stale, TCP fast start exploits priority dropping at routers, and augments TCP's loss recovery mechanisms to quickly detect and abort a failed fast start attempt.
In addition to the two challenges we have discussed, a third challenge arises from the increasing deployment of asymmetric access networks. Although Web browsing has asymmetric bandwidth requirements, bandwidth asymmetry could adversely impact Web download performance due to a disruption in the flow of acknowledgement feedback that is critical to sustaining good TCP throughput. To avoid performance degradation, we have developed end-host and router-based techniques, that both reduce disruption in the feedback and reduce TCP's dependence on such feedback. In certain situations, these techniques help decrease download time by a factor of fifteen.
This thesis includes mathematical and trace-based analysis, simulation, implementation and performance evaluation. In addition to the algorithms that we have designed and software that we have developed, our contributions include a set of paradigms for advancing the state-of-the-art in Internet transport protocols. These paradigms include the use of shared state and/or persistent state, and the exploitation of differentiated services mechanisms in routers.
Prelude(Abstract, Acknowledgements, Table of Contents, etc.)
Chapters
1. Introduction
2. Background and Related Work
3. Methodology
4. Analysis of HTTP Performance
5. Persistent-Connection HTTP
6. TCP Session
7. TCP Session: Advanced Issues
8. TCP Fast Start
9. Asymmetric Access Networks
10. Conclusions and Future Work
Appendices
A. Performance Analysis Tools
B. Implementation of TCP Session
Bibliography