Dynamic Control of Running Servers

Abstract

Motivated by a data center setting, we study the problem of joint dispatching and server sleep state control in a system consisting of two queues in parallel. Using the theory of Markov decision processes and a novel lookahead approach, we explicitly determine near-optimal control policies that minimize a combination of QoE costs, energy costs, and wear and tear costs due to switching. Guidelines are provided as to when these combined policies are most effective.