Using stochastic linear-parameter-varying control for CPU management of internet servers

This paper presents a stochastic nonlinear control approach to performance management of Internet Web servers. In particular, a CPU frequency management problem is studied to satisfy response time SLA. The research in this paper is mainly motivated by two reasons: 1) the stochastic nature of workload characterization and performance metrics requires probabilistic approaches to system modeling and control designs, and 2) rather than over-provisioning for the worst-case load, performance management in today's Internet-server environments would limit rather than eliminate the risk of failing to meet service demand, allocating to each application the minimal server resources needed and leaving surplus resources to deploy elsewhere. The presented probabilistic linear-parameter-varying control utilizes an analytical fluid model that is scheduled by randomly distributed workload parameters, and provides control solutions via solving a stochastic semi-definite program. Evaluated through real Web server traces, the presented stochastic robust LPV control not only provides balance between meeting response time and resource efficiency, but also provides computational advantages in the implementation of control designs.