Energy-aware server with SRPT scheduling: Analysis and optimization

Abstract

We consider the optimal energy-aware control of a single server in a server farm. The server is modeled as an M/G/1 queue with a particular control policy that allows to put the server to a sleep mode to save energy with an additional delay cost, the setup delay, after the server is turned on again. Our main result is the derivation of mean response time for such a system under SRPT scheduling. In particular, we show that the mean response time can be decomposed into two parts: the mean response time of an ordinary M/G/1-SRPT, and an additional penalty term for switching the server to a sleep state. Furthermore, we study the energy-performance optimization of the system and prove that, for the Energy Response time Weighted Sum (ERWS) and Energy Response time Product (ERP) cost metrics, the optimal control either puts the server into a sleep state immediately when it becomes idle or keeps it idling until the next job arrives.