Optimizing Response Time, Rather than Hit Rates, of WWW Proxy Caches

Abstract

This thesis investigates the possibility of improving World Wide Web (WWW) proxy cache performance. Most published research on proxy caches is concerned only with improving the cache hit rate. Improving only the hit rate, however, ignores the actual retrieval times experienced by WWW browser users. This research investigates removal algorithms that consider the time to download a file as a factor. Our experiments show that a removal algorithm that minimizes only the download time yields poor results. However, a new algorithm is investigated that does provide improved performance over common removal algorithms using three factors --- the speed at which a file is downloaded, the size of the file, and the number of references to the file (the number of hits). Experiments are conducted with a modified version of the Harvest Cache which has been made available on the Internet from the Virginia Tech Network Research Group's (VT-NRG) home page. WWW traffic from the ".edu" domain is used in all of the experiments. Five different removal algorithms are compared: least recently used, least frequently used, document size, and two new algorithms. The results indicate that the new three factor algorithm reduces the average latency experienced by users.