Power efficiency through i/o access pattern reshaping

Power consumption has become a critical problem in both the mobile and server system markets. The need for power management has led to the development of hardware devices with low-level power management support. I/O devices like hard disks and wireless network cards support non-operational low power modes that can save energy during periods of inactivity. Unfortunately, several workloads lead to access patterns that low-power modes cannot exploit for power efficiency. This dissertation proposes high-level I/O access pattern reshaping for power efficiency. Specifically, it presents two methods to reduce power without compromising performance: increasing I/O burstiness and exploiting hardware concurrency. Bursty access patterns increase opportunities for power management by increasing the length of periods during which a device remains inactive. The dissertation proposes a set of prefetching and caching enhancements that lead to a significant increase in idle interval lengths. The proposed enhancements have been implemented in the Linux operating system and evaluated experimentally. They lead to disk energy savings of up to 80% and wireless network card energy savings of up to 30% with little or no reduction in throughput or interactive responsiveness. The dissertation also presents and validates a generic model that predicts the impact of prefetching on I/O device energy consumption under the assumption of a perfect predictor. In several environments, increasing I/O burstiness may not lead to significant power savings. In such cases, hardware concurrency can be used to improve power efficiency. One can exploit the performance benefit achieved through hardware parallelism to reduce the power consumption of server-class disk arrays. Based on state of the art disk specifications, one could replace each server-class disk with a mirrored disk array of three power-efficient laptop-class disks. A new scheduling policy for disk array I/O requests can then exploit the variability of server workloads by diverting accesses to a small subset of the disks during periods of low or moderate load, allowing disks that represent excess capacity to power down. The proposed system has been evaluated through simulation. Potential savings range from 50% to 80% of the total disk energy consumption.