Power-aware QoS Management in Web Servers

  • Authors:
  • Vivek Sharma;Arun Thomas;Tarek Abdelzaher;Kevin Skadron;Zhijian Lu

  • Affiliations:
  • -;-;-;-;-

  • Venue:
  • RTSS '03 Proceedings of the 24th IEEE International Real-Time Systems Symposium
  • Year:
  • 2003

Quantified Score

Hi-index 0.01

Visualization

Abstract

Power management in data centers has become an increasinglyimportant concern. Large server installations aredesigned to handle peak load, which may be significantlylarger than in off-peak conditions. The increasing cost ofenergy consumption and cooling incurred in farms of high-performanceweb servers make low-power operation duringoff-peak hours desirable. This paper investigates adaptivealgorithms for dynamic voltage scaling in QoS-enabled webservers to minimize energy consumption subject to servicedelay constraints. We implement these algorithms inside theLinux kernel. The instrumented kernel supports multipleclient classes with per-class deadlines. Energy consumptionis minimized by using a feedback loop that regulatesfrequency and voltage levels to keep the synthetic utilization1 around the aperiodic schedulability bound derived inan earlier publication. Enforcing the bound ensures thatdeadlines are met. Our evaluation of an Apache server runningon the modified Linux kernel shows that non-trivial off-peakenergy savings are possible without sacrificing timeliness.