Design and evaluation of load control in Web server systems

Nonlinear discrete-time modeling of a Web server system is investigated. Server systems typically contain nonlinearities such as saturations and bounded queue lengths. The incoming traffic and service rates are best modeled by stochastic processes, well described and analyzed by queuing theory. Here, we develop and validate a control theoretic model of a general single server queue, a so-called G/G/1-system. Based on the nonlinear system model, design of admission controllers are presented and the closed loop stability is analyzed. The behavior of the server model is verified with respect to queue theoretic models. Finally, experimental evaluation is performed on an Apache Web server in a laboratory network. A traffic generator is used to represent client requests. The control of the Apache server has been re-written to implement our algorithms. We show that the control theoretic model aligns well with the experiments on the Web-server. Measurements in the laboratory setup show the robustness of the implemented controller, and how it corresponds to the results from the theoretical analysis and the simulations.