Multi-tier Energy Management Strategy for HPC Clusters

Power consumption of HPC cluster is increasingly concerned by HPC designers and users. This paper proposes a multi-tier cluster energy management for reducing energy consumption of the cluster system with minimal effect on performance. The proposed management combines cluster-level and node-level strategies. Cluster-level strategy uses a self-learning load estimation algorithm to predict new-coming task´s load and presents a novel PI control theory based node allocation mechanism to decide how many nodes should be selected to execute parallel tasks. The cluster-level strategy also uses an on-demand on/off strategy to decide how the node scales its CPU frequency and whether to be turned off. Node-level strategy uses an enhanced-conservative governor algorithm to improve the sensitivity of the frequency adjustment when load drops. Experiments show that the proposed multi-tier power management is more efficient than other traditional strategies in reducing overall system power consumption.