Impact of traffic distribution on web cache performance

Caches are a critical element of web-based information systems. Understanding the expected behaviour of cache policies is especially important for achieving good quality of service. Existing works have suggested that the behaviour of the web demand can be modelled as a Zipf distribution with α≤1. New evidence, which is presented in this paper, shows that today websites are following Zipf distributions with α > 1. This article analyses real logs obtained from the client layer of high traffic websites. The main result of this article is that under these conditions, the cache hit ratio can be extremely high with a very small cache size. This means that a very expensive and high resource demanding cache is not needed for effective implementation: A cache size equal to 0.6% of the working set is enough to reach more than 80% of hit ratio, once the right replacement policy has been chosen. Copyright © 2015 Inderscience Enterprises Ltd.