Power and performance management via model predictive control for virtualized cluster computing

There is growing research interest in the power and performance management problem for the virtualized computing systems (most data centers) today. In such systems, multiple control actions are feasible, including turning on or off the machine, creating, destroying or migrating the virtual machines, adjusting the frequency of processor etc. In this paper, a novel discrete system state model is built, in which almost all the control actions as well as the time delay effect can be reflected by the state transition. Based on the model, the performance and power can be defined on each state. Then, the model predictive controller is implemented and validated by dynamically minimizing the quadratic cost function which integrate the quality of performance tracking and power saving. Experiment results show that by choosing a moderate weight between performance and power, performance can be well tracked (the requirement being satisfied for 99% times) and power consumption can be saved significantly (by 33%) on average compared with the system without the controller.