Hyper Text Transfer Protocol (HTTP) traffic dominates Internet traffic. The
exchange of HTTP messages is implemented using the connection-oriented TCP
. HTTP/1.0 establishes a new TCP connection for each HTTP request, resultin
g in many consecutive short-lived TCP connections. The emerging HTTP/1.1 re
duces latencies and overhead from closing and re-establishing connections b
y supporting persistent connections as a default. A TCP connection which is
kept open and reused for the next HTTP request reduces overhead and latenc
y. Open connections, however, consume sockets and memory for socket-buffers
. This trade-off establishes a need for connection-management policies. We
propose policies that exploit embedded information in the HTTP request mess
ages, e.g., senders' identities and requested URLs, and compare them to the
fixed-timeout policy used in the current implementation of the Apache Web
server. An experimental evaluation of connection management policies at Web
servers, conducted using Web server logs, shows that our URL-based policy
consistently outperforms other policies, and achieves significant 15-25% re
duction in cost with respect to the fixed-timeout policy. Hence, allowing W
eb sewers and clients to more fully reap the benefits of persistent HTTP. (
C) 1999 Published by Elsevier Science B.V. All rights reserved.