Web caching systems : design and analysis

Abstract

Web caching is widely viewed as an effective way to improve Web performance and make it scalable. In this thesis, we examine the design and analysis of Web caching systems. Initially, we propose a new basis for making caching decisions; namely, a site-based approach. Unlike the focus of conventional approaches, the essence of this approach is to consider only the Web site that an object belongs to, not the object itself. We show that this approach brings many benefits for several different types of applications in Web caching areas. Based on the idea of the site-based approach, we propose a site-based mapping scheme for parallel proxy servers. The basic idea is to map all Web objects belonging to the same site to the same proxy server. By taking advantage of the persistent connection feature provided by HTTP/1.1, this mapping scheme reduces 40%-70% of the connection setups and teardowns when compared to a traditional URL-based mapping scheme. To overcome the problem of overloaded proxy servers, we propose a site-based dispatching (SBD) technique that, by preventing unnecessary requests from reaching the proxy, reduces the workload. Analytical and simulation results show that the proposed technique reduced the proxy load by up to 48% and hence enabled a system throughput gain of up to 55%. To address the question of limited storage capacity, we study the startup behaviors of caching systems using the least-recently-used (LRU) cache replacement policy. We present exact analyses of cache occupancy and cache miss probability. These measures provide insights into know how long a cache takes to become fully occupied, and how quickly a cache reaches its acceptable cache miss probability. These analyses are proven to be very accurate by computer simulations. Finally, we discuss polling models that can be used to model many areas in Web caching systems. We present a new approximation of mean-waiting time for asymmetric polling systems with nonexhaustive service discipline. Numerical results show that our proposed method gives a more reliable and accurate result than those of another well-known approximation.