Abstract
HPC clusters are widely used to execute parallel tasks. With the increasing number of nodes and frequency of processors, they consume huge amount of energy. The heat generated by clusters also imposes very heavy load for cooling infrastructures. The utilization of some clusters is not always high, indicating that there is a huge space to conserve energy consumption with more intelligent energy management scheme. Although there has been some energy conservation schemes proposed for web clusters, they are not applicable to HPC clusters. In this paper we propose a dynamic energy conservation scheme for HPC clusters. The scheme is to turn some cluster nodes on and off dynamically according to the current and historical workload. The goal is to reduce the energy consumption of clusters with minimal performance loss. We evaluate our scheme by simulation and show that it can effectively conserve energy consumption.
