Thwarting the power-hungry disk

TLDR

Mobile computers must balance power savings from spinning down disks against response‑time penalties, as disks consume a large portion of system power. The study uses trace‑driven simulation to evaluate trade‑offs and compare disk spin‑down algorithms. Simulations model disk accesses from a Macintosh Powerbook Duo 230 and a Hewlett‑Packard 9000/845 workstation on HP Kittyhawk C3014A and Quantum Go•Drive 120 disks. The optimal offline algorithm can cut disk power by 35–50 % without increasing response time, while an online 10‑second threshold algorithm reduces energy by ~40 % but adds 140 delays over four hours.

Abstract

Minimizing power consumption is important for mobile computers, and disks consume a significant portion of system-wide power. There is a large difference in power consumption between a disk that is spinning and one that is not, so systems try to keep the disk spinning only when it must. The system must trade off between the power that can be saved by spinning the disk down quickly after each access and the impact on response time from spinning it up again too often. We use trace-driven simulation to examine these trade-offs, and compare a number of different algorithms for controlling disk spin-down. We simulate disk accesses from a mobile computer (a Macintosh Powerbook Duo 230) and also from a desktop workstation (a Hewlett-Packard 9000/845 personal workstation running HP-UX), running on two disks used on mobile computers, the Hewlett-Packard Kittyhawk C3014A and the Quantum Go•Drive 120. We show that the perfect off-line algorithm--one that consumes minimum power without increasing response time relative to a disk that never spins down--can reduce disk power consumption by 35--50%, compared to the fixed threshold suggested by manufacturers. An on-line algorithm with a threshold of 10 seconds, running on the Powerbook trace and Go•Drive disk, reduces energy consumption by about 40% compared to the 5-minute threshold recommended by manufacturers of comparable disks; however, over a 4-hour trace period it results in 140 additional delays due to disk spin-ups.