Self-Learning Hard Disk Power Management for Mobile Devices

A multitude of different hard disk power management algorithms exists—applied to real systems or proposed in the literature. Energy savings can only be achieved if the hard disk is idle for a minimum period of time. These algorithms try to predict the length of each idle interval at runtime and decide whether the disk should be switched to a lowpower mode or not. In this paper, we claim that there is no general-purpose policy that maximizes energy savings for every workload and present system services that dynamically switch between different, specialized power management algorithms. The operating system automatically learns which policy performs best for a specific workload. Therefore, hard disk accesses are monitored and fed into a simulator that estimates the drive’s energy consumption under different lowpower algorithms. In order to recognize workloads at runtime, the system additionally monitors a set of I/O-related parameters. Using techniques from machine learning, a set of rules can be derived automatically which enable a power management daemon to identify the current workload and its optimum low-power algorithm on-line. Furthermore, the user can train the system to consider application-specific performance requirements. A prototype implementation for Linux is presented and evaluated through experiments with two different hard disks.