Disk layout optimization for reducing energy consumption

Excessive power consumption is becoming a major barrier to extracting the maximum performance from high-performance parallel systems. Therefore, techniques oriented towards reducing power consumption of such systems are expected to become increasingly important in the future. Since disk systems of high-performance architectures are known to constitute a large fraction of the overall power budget, they form an important optimization target. Previous work on disk power management focuses primarily on hardware based schemes. However, since disk access pattern, i.e., the order in which disks on a system are accessed, is mainly shaped by the program code access pattern and disk layout of data, software techniques can also play a critical role in disk power management. Motivated by this observation, this paper proposes and evaluates a profile-driven disk layout optimization scheme for reducing energy consumption. The proposed scheme analyzes the array access traces obtained through profiling and determines, for each disk-resident data structure, the start disk from which the data is striped, the number of disks over which the data is striped, and the stripe unit. This paper discusses implementation details of our approach and presents an experimental evaluation of it. Our experiments with the entire suite of Spec95 floating-point benchmarks that are modified to operate on disk-resident data show that the proposed approach is very effective in reducing disk energy consumption. The results also show that the performance degradation caused by our approach is very small. This paper also compares our approach to a code restructuring based optimization mechanism and discusses how the two techniques can be combined for achieving the best results.