This paper examines pitfalls that arise in the process of generating synthetic Web server workloads in a testbed consisting of a small number of client machines. It exposes the limitations of the simple request generation scheme that underlies state-of-the-art Web server benchmarks. We propose and evaluate a new strategy that addresses these problems using a set of specially constructed client processes. Initial experience in using this method to evaluate a typical Web server indicates that measuring Web server performance under overload and bursty traffic conditions gives new and important insights in Web server performance. Our new methodology enables the generation of realistic, bursty HTTP traffic and thus the evaluation of an important performance aspect of Web servers.
Source code and additional technical information about S-Clients can be found here