The Case for Energy-Proportional Computing - Google, 2007

  • [Paper] [Mirror]

  • Historically, energy efficiency has been mostly for longer battery life for mobile devices
  • Computers have used embedded techniques clock gating (so much fun and pain), different voltage planes and dynamic frequency adjustments

“Servers are rarely completely idle and seldom operate near their maximum utilization, instead operating most of the time at between 10 and 50 percent of their maximum utilization levels.”

  • This isn’t because of poor usage, but actually good engineering. Consider:
    • fail-over
    • maintenance tasks (which require additional resources)
    • load spikes
    • distributing data across machines (not just computation) that is hard to move

“Even an energy-efficient server still consumes about half its full power when doing virtually no work.”

  • The authors advocate for designing servers so that their energy usage is proportional to their utilization
    • Surprisingly, the CPU is no longer the main source of power
    • They use a wide dynamic power range and active low-power modes
      • Disk can’t really do this now because of the high cost to spin up a disk
        • Both in latency and energy
      • Maybe we could use a memory hierarchy and the five minute rule for spun-down disks