Overload Control and Performance Evaluation of Web Servers

We are increasingly dependent on Internet as a large aspect of our daily life is now handled via online applications. Recent years we have seen several large-scale crises. This has made crisis-related research on Internet applications important. We need the applications to be available at any time, so there is a need to increase availability during heavy overload, both in crises and during normal overload situations.

This thesis is based on both theoretical and empirical research with a focus on investigating overload control mechanisms for web servers. Three different types of mechanisms are considered; one mechanism where queueing theory and control theory is combined into an admission control scheme based on a PI-controller, one where admission control supports service level agreements for a web server and finally a content adaptation system based on optimization. The overload control mechanisms have been analyzed and simulated with respect to performance, and two of the overload control mechanisms have been implemented as modules in the Apache web server.

The thesis further presents a web server model used in one of the overload control mechanisms. The model is validated with measurements on the Apache web server. In addition, a new traffic generator for Internet traffic is presented. The generator uses a general traffic model that can be used to represent most kinds of traffic. The most common traffic models have been implemented, as well as a model of crisis-like traffic.