| Abstract: |
Measurements over the last few years have shown HTTP traffic to be nearly 75% of overall Internet traffic. Virtually all implementations of HTTP use TCP as the transport layer given TCP?s
ubiquity. Several of the performance problems related to the Web were traced to HTTP?s reliance on TCP. TCP was optimized for long transfers but most Web responses still consist of an average of
8 Kilobytes, i.e., a few packets. The setting up and tearing down cost of TCP dominate the overall cost of a HTTP transfer. Persistent connections were introduced in HTTP/1.1 to reduce overhead
and latency by reducing the number of separate connections needed to download resources. Given the variety of TCP variants in use (NewReno, Reno, Tahoe, etc.), it is a natural question to examine
which specific aspects of TCP affect Web performance and do certain variants help provide better Web performance. By Web performance, we mean the end to end latency experienced by Web clients and
the impact on the Web server in handling the requests. All the popular variants of TCP predated the popularity of HTTP and to the best of our knowledge there has not been a comprehensive study on
evaluating the performance of HTTP over the various TCP variants. In this paper, we explore the various features of the TCP variants, their motivation, and the actual impact on the network. |
