We have proposed an optimistic deltas approach to reduce the latency of accessing WWW pages. This approach involves sending the differences between versions of a page, or deltas, to the client, instead of sending entire pages. It also permits stale data to be sent during periods of inactivity. Our approach is optimistic because it sends data that may not be needed; instead, it optimizes for the common case when pages change incrementally, at the expense of a slight overhead in the rare cases when a modification drastically changes the content of the page. In other words, we assume that in most cases when a copy cached by the proxy is deemed unusable, it is either still current, or, if it has been modified, the size of the modification is considerably smaller than the page itself.
Our study of an AT&T multi-version archive of WWW pages confirmed the above assumption. In fact, by examining the extent to which the results of AltaVista queries with slightly different parameters differ, we showed that this assumption may even hold for dynamically generated pages. However, in general we expect that other sorts of data, such as images, should be handled specially rather than processed as deltas.
A study of the latency to obtain WWW pages confirmed that the latency in obtaining data may often be sufficient to send stale data, for the purpose of sending a small delta once the data is available. However, performance may be degraded when latency is low and more sophisticated techniques for deciding when to abort the transfer of stale data are required.
We implemented our approach without changing the browser. Instead, we configure the browser to connect to a client proxy on the same machine, which in turn connects to a server proxy. These proxies have been modified to follow the optimistic deltas approach. We compared the performance of this configuration with the original system. This performance study, based on microbenchmarks, showed a significant latency reduction achieved by our approach: an average of 12-33% improvement across all pages in the study, depending on system parameters, with some transfers improved by an order of magnitude. One particularly surprising result was the effect that transferring potentially stale data had on the TCP slow-start algorithm when a link is otherwise idle, consistently improving end-to-end latency.
While a long-term experiment that would compare the performance of our approach with existing proxy caching systems on real-life workloads is needed, the experiments described in the paper strongly suggest that the optimistic delta mechanism results in a considerable reduction of WWW latency.