An evaluation of code and data optimizations in the context of disk power reduction

Abstract

Disk power management is becoming increasingly important in high-end server and cluster type of environments that execute data-intensive applications. While hardware-only approaches (e.g., low-power modes supported by current disks) are successful to a certain extent, one also needs to consider the software side to achieve further energy savings. This paper first demonstrates that conventional data locality oriented code transformations are not sufficient for minimizing disk power consumption. The reason is that these optimizations do not take into account how disk-resident array data are laid out on the disk system, and consequently, fail to increase idle periods of disks, which is the primary metric using which disk power can be reduced. Instead, we propose a disk layout aware application optimization strategy that uses both code restructuring and data layout optimization. Our experimental evaluation with several benchmark codes reveal that the proposed strategy is very successful in reducing disk energy consumption without performing much worse than a pure data locality oriented scheme, as far as execution cycles are concerned. The experiments also show that the benefits coming from our approach increase with the increased number of disks; i.e., it scales very well.