EECS technical reports

Oskicat catalog record
Conditions of Use

Archive Home Page

TCP Behavior of a Busy Internet Server: Analysis and Improvements

Balakrishnan, Hari
Padmanabhan, Venkata N.
Seshan, Srinivasan
Stemm, Mark
Katz, Randy H.
Technical Report Identifier: CSD-97-966
August 1997

Abstract: The rapid growth of the World Wide Web in recent years has caused a significant shift in the composition of Internet traffic. Although past work has studied the behavior of TCP dynamics in the context of bulk-transfer applications and some studies have begun to investigate the interactions of TCP and HTTP, few have used extensive real-world traffic traces to examine the problem. This interaction is interesting because of the way in which current Web browsers use TCP connections: multiple parallel short connections from a single host.

In this paper, we analyze the way in which Web browsers use TCP connections based on extensive traffic traces obtained from a busy Web server (the official Web server of the 1996 Atlanta Olympic games). At the time of operation, this Web server was one of the busiest on the Internet, handling tens of millions of requests per day from hundreds of thousands of clients. We first describe the techniques used to gather these traces and reconstruct the behavior of TCP on the server. We then present a detailed analysis of TCP's loss recovery and congestion control behavior from the recorded transfers. The two most important results are that: (1) short Web transfers lead to poor loss recovery performance for TCP, and (2) concurrent connections cause an overly aggressive use of the network. We then discuss techniques designed to solve these problems. To improve the loss recovery performance of short transfers, we present a new technique for TCP loss recovery. To improve the congestion control and loss recovery performance of multiple parallel connections, we present a new transport-level integrated connection approach to congestion control and loss recovery. Simulation and trace analysis results show that our enhanced loss recovery scheme could have eliminated 25% of all timeout events, and our integrated connection approach provides greater fairness and improved startup performance for parallel connections. Our solutions are more general than a specific protocol enhancement such as the use of persistent connections in P-HTTP and HTTP/1.1, and include techniques, such as improved TCP loss recovery, that are not addressed by these protocols.