Reducing data center energy consumption through peak shaving and locked-in energy avoidance

Abstract

The grid company enforces high penalties for the peak power demands of cloud data centers. These penalties result in high electricity bills that can be avoided by relying on the servers' uninterruptible power supply (UPS) batteries as a source of energy during high-demand periods. This paper proposes a resource management framework that leverages the availability of UPS batteries to reduce the data center's peak demands. Our framework consists of: 1) a scheduler that accounts for both the stored energy and the resource slacks when making workload placement decisions and 2) a power distributor that decides which UPS battery should store energy and by how much in order to increase the amount of energy accessible for peak shaving during high-demand periods. Our framework considers two different distributed UPS topologies: server-level topology, where each server in the data center is dedicated a single UPS battery, and rack-level topology, where each group of servers is dedicated a set of UPS batteries to be used in a shared manner. Experiments based on real Google traces show that our framework reduces significantly the data center's electricity bill and these reductions can be up to 20% when compared to existing management techniques. - 2017 IEEE.