Demand Response aware cluster resource provisioning for parallel applications

Abstract

Data center energy consumption is significant and accounts for about 2% of total energy use in the U.S. recently. A range of approaches, from cooling techniques to workload consolidation have been taken to improve data center energy efficiency. In contrast to most methods published so far, this paper treats a data centre as a consumer in an electricity market. Our intention is to make data centres more responsive to electricity market conditions with minimal impact on their performance. In electricity markets, Demand Response(DR) is a method for improving grid efficiency by encouraging consumers to adjust their demand during price peaks or network stress. Significant consumption and cost savings can potentially be made via implementing DR programs involving a large set of consumers. Traditionally, DR has been a largely manual process, however, automated DR is becoming increasingly prevalent due to the deployment of smart grid technologies. In this paper, we treat the server cluster in a data centre as an energy consumer that participates DR activities. We give two algorithms to enable the cluster to automatically adjust the number of active servers to respond to DR requests while maintaining acceptable system performance. We evaluate our algorithms using real traces.