A Feasibility Analysis of Power Awareness in Commodity-Based High-Performance Clusters

C. Hsu and W. Feng

IEEE International Conference on Cluster Computing (Cluster 2005)
Boston, Massachusetts, USA, September 27 - 30, 2005

---

Abstract

Steep power demands and their subsequent energy costs are becoming serious design issues for operating commodity-based high-performance clusters. Furthermore, the excessive power demands of high-performance clusters can often lead to thermally-induced failures, and hence, reduced reliability. While massive cooling helps to reduce the frequency of such failures, researchers ought to consider a complementary approach that proactively reduces intrinsic power demands, such as hibernating a subset of cluster nodes when the workload is light.

In this paper, we propose and study a new power-reduction scheme that reduces the thermal power of processors by lowering frequency and voltage in the context of high-performance computing. Our feasibility analysis on a 16-processor Opteron-based Beowulf cluster shows that one can easily reduce a significant amount of CPU and system power dissipation and its associated energy costs while still maintaining high performance. Specifically, our study shows that a 5\% performance slowdown can be traded off for an average of 19\% system energy savings and 24\% system power reduction. This preliminary empirical result, via real measurements, is encouraging because hardware failures often occur when the cluster is running hot, i.e, when the workload is heavy, and the new power-reduction scheme can effectively reduce a cluster's power demands during these busy periods.

---

  

START Conference Manager (V2.49.7)

Maintainer: rrgerber@softconf.com