Predictive reduction of power and latency (PuRPLe)

Increasing efforts have been aimed towards the management of power as a critical system resource, and the disk can consume approximately a third of the power required for a typical laptop computer. Mechanisms to manage disk power have included spin-down policies and APIs to modify access workloads to be more power-friendly. In this work we present a measurement study of disk power consumption, focusing on the potential impact of successfully optimizing disk layout or predicting future disk accesses with predictive read-ahead. We demonstrate how such strategies can allow the reduction of disk power consumption at least as well as traditional disk spin-down schemes, while avoiding the serious performance degradation that can occur from excessive spin-downs. Experimental results showed that a successful predictive disk management policy could reduce disk power consumption by over 80%, while maintaining the responsiveness of a continuously running disk. In contrast, an aggressive spin-down policy that does not attempt to optimize layout or predictively read-ahead data, would achieve the same results at the expense of increasing average delays by 2 to 4 times. Another contribution of this work involves the accuracy of the measurements, which were conducted at a level precise enough to distinguish the power consumption of drive electronics, spindle-motors, and disk arm movement.