Scalability Issues in Cluster Web Servers

Abstract

In this dissertation a Web cluster architecture is designed, implemented and evaluated. We believe that the use of COTS hardware and integrated front end and back end software better scalability and flexibility can be achieved for Web clusters. Furthermore, the clustering software is all implemented in Java, with the goal of evaluating Java as a platform for the implementation of high performance network servers. The Web cluster prototype described here is built out of a set of distributed components that collaborate to achieve good scalability as a Web server. As all of the components have been designed from scratch to integrate with each other and work well together, the scalability problems associated with state of the art content aware commercial clusters are avoided.

The prototype Web cluster described in this dissertation will be evaluated using publicly available Web benchmarking tools and Web access log traces. Different workloads will be applied to the cluster as well as to a standalone Java Web server to compare their performance and scalability. Web clustering research has frequently assumed that the Web traffic is mostly static, that is, mainly composed of requests for static files. However, more and more Web sites implement dynamic Web applications and there is a tendency towards completely dynamic sites. Thus, apart from standard static traffic dynamic Web traffic is also simulated and evaluated in our tests.

This dissertion is organised as follows: chapter two contains a state of the art overview of Web clustering, chapter three discusses the design and implementation of an scalable Web cluster in Java, chapter four evaluates the performance and scalability of the cluster prototype and chapter five presents conclusions and future work.