Modeling and control design for performance management of web servers via an LPV approach

This paper presents a control-theoretic approach to the performance management of Internet Web servers to meet service-level agreements. In particular, a CPU frequency management problem is studied to provide response time guarantees with minimal energy cost. It is argued that linear time-invariant modeling and control may not be sufficient for the system to adapt to dynamically varying load conditions. Instead, a Linear-parameter-varying (LPV) approach is presented in this paper, where workload arrival and service parameters are chosen as scheduling parameters to characterize time-varying operating conditions. Modeling the performance management of a Web server as an LPV system has been extensively discussed in this paper; we have derived first-principles models based on analyzing transient and steady-state queueing dynamics as well as empirical models using system identification algorithms. LPV H_∞ controllers are then designed for the derived LPV system models. Using real Web server workloads, the performance of LPV control compares favorably to various linear control designs and a design based on the conventional queueing theory. The proposed LPV modeling and control framework can be generalized to incorporate more sophisticated workload models and more complicated server environments. In addition, due to the LPV nature of Web systems with respect to load conditions, the proposed approach can be applied to a variety of resource management problems and used for middleware designs.