The key issue related to the performance of the caching mechanism is caching policy which decides how the documents are brought into and removed from the cache. Between these two types of decisions, the removal policy seems to be more critical to the performance of caching. In order for the caching be beneficial, a caching policy should be, on one hand, efficient in terms of hit rate, i.e. the percentage of documents/bytes transferred from the cache, and on the other hand, of low management overhead. There are already caching mechanisms in operating systems (e.g. Unix) and databases. However, they are based on pages (of the same size) and not on files (with different sizes), and therefore are not appropriate for a Web server which requires entire files. Another problem with the use of caching mechanism of operating systems is that the same main memory cache is shared with other applications, and is read/write oriented, while Web documents are usually read-only.
Most often, removal policies like LRU (Least Recently Used) used in cache software (squid accelerator, or proxy-cache) are those used in traditional operating systems, and do not take into account the specificities of Web traffic. Studies of the performance of LRU and its new variants have been reported in the last two years for Web applications. The reader is referred to [1,8] for server cache and [3,7,9,10,11] for client cache.