Stochastic approximation control of power and tardiness in a three-tier web-hosting cluster

Large-scale web-hosting and data centers are increasingly challenged to reduce power consumption while maintaining a minimum quality of service. Dynamic voltage and frequency scaling provides one technique to curb power consumption by limiting the power supply and/or frequency of the CPU at the expense of lower execution speed. Model-based approaches often require tedious offline profiling, and generating an accurate model under all conditions may be infeasible. This paper develops a stochastic feedback-control algorithm, and couples it with a method of stochastic optimization to minimize power consumption while maintaining tardiness in a three-tier system. Our approach assumes nothing about the system and the application, treating each as a 'black box.' The scheme is effective under limited dynamic workload conditions that can alter the response times and power consumption to be approximated. With little overhead, the control scheme is able to maintain a specified quantile of tardiness under a desired threshold, while suppressing power consumption to within 1% of its theoretical minima.