Performance Analysis of Cache Policies for Web Servers

Abstract Many cache policies have been proposed for Web document caching However, the unique characteristics of Web workloads and the "right" approach to Web caching are still not well understood In this paper we attempt to analyze some of the important issues by proposing a universal, mathematically sound, approach to Web caching and showing that cache policies based on that approach outperform other polices We use traces from five Web servers to compare and analyze performance of various cache policies We also discuss implementation problems and identify memory fragmentation as a potential problem in any dynamic cache policy We show how the proposed universal approach can be modified to account for memory fragmentation without significant performance degradation Unfortunately, memory fragmentation further increases CPU overhead and contention between threads in a multithreaded Web server We identify static caching as an easy way to avoid these problems Although the paper focuses on primary Web servers, its results should also be useful to the designers of Web proxy caches