Optimistic Deltas for WWW Latency Reduction

Summary by Michael Stok

An approach to reducing latency in WWW transactions using a cooperating client and server proxy system was presented at this session. Although caching is widely used at both the client and server end of a WWW transaction, having a proxy at both the client and server end can allow quite complex interactions to be performed without any change to the client or the server.

The client proxy intercepts the user requests and keeps a cache of pages delivered to the client. It also manages the delivery of pages to the client and allows the client to remain ignorant of aborted transfers between proxies by delivering only complete pages.

The optimistic delta strategy can increase the amount of data sent back in response to a client, but it takes advantage of times that are usually "wasted" to transfer possibly out-of-date deltas or data between the client proxy and the server proxy with the ability to abort the transfer if fresh data become available. The server proxy can be sending a delta or potentially stale data to the client proxy while it's waiting to see if the data on the server have been changed. The difference between the communication speeds between the server and the server proxy (fast) and the server proxy and the client proxy (slow) allows time to compute and compress deltas and abort the transfer of stale data without wasting too much time.

The speaker described system mechanics and the trade-offs made in the optimistic delta system and presented results showing that the deltas were frequently small enough for the scheme to be successful. He also showed that optimistic deltas can be used on dynamically generated WWW content because there is often so much boilerplate that deltas are effective in delivering the pages.

Questions covered ground such as whether deltas between images could be effectively computed. Some of the outlying points on the graphs illustrating the gains achieved using the optimistic delta system were explained as the result of the daily change of the GIF image on the Dilbert page.

The need for techniques to effectively reduce the load on the slowest part of the Web and to make effective use of what the user might see as "dead time" when the modem is idle will get greater as the volume of data and information on the Web gets ever larger.

Originally published in ;login: Vol. 22, No.2, April 1997.