Improving World-wide Web Latency

Jeffrey C. Mogul & Venkata N. Padmanabhan

The HTTP protocol, as currently used in the World-wide Web, causes excessive latency and overhead. Each client request uses a separate TCP connection, which adds unnecessary overheads for connection set-up, TCP slow-start, and connection tear-down. Use of separate requests for each document and its inlined images means that each retrieval requires at least one network round-trip.

We modified the NCSA HTTP server and the Mosaic client to maintain TCP connections for multiple requests. The server can close inactive connections, if necessary, to manage its resources. We also added two new HTTP primitives, to avoid unnecessary round-trips: GETALL, which retrieves a document and all of its inlined images, and GETLIST, which retrieves a client-specified set of documents. Our implementations interoperate with unmodified HTTP implementations.

With a geographically remote server (round-trip time = 80 ms), our measurements show that the latency seen by the client declines by 15-50%, depending on the number and size of inlined images. Even with a local server, latency improves by 10-40%.

We also modified Mosaic to retain FTP control connections for multiple requests; this also significantly improves response time, typically cutting it down to less than half.

We are currently investigating the performance of a server that prefetches files, using past access patterns to predict future requests. This could mask disk latencies, by performing disk I/O in parallel with network activity, thereby improving response time.

Contact Author: Venkata Padmanabhan (padmanab@cs.berkeley.edu)


Venkata Padmanabhan's Vita

Present

Education

Research Interests

Contact Address

Computer Science Division
University of California at Berkeley
Berkeley, CA 94720
USA
Tel: (510) 643-9435, (510) 548-6234
Email: padmanab@cs.berkeley.edu


Jeffrey Mogul's Vita

7/86 to present

Education

Selected publications

Other