TSB: A DVS algorithm with quick response for general purpose operating systems

  • Authors:
  • Euiseong Seo;Seonyeong Park;Jinsoo Kim;Joonwon Lee

  • Affiliations:
  • Department of EECS, Korea Advanced Institute of Science and Technology, Gusongdong Yusonggu, Daejeon 305-701, Republic of Korea;Department of EECS, Korea Advanced Institute of Science and Technology, Gusongdong Yusonggu, Daejeon 305-701, Republic of Korea;Department of EECS, Korea Advanced Institute of Science and Technology, Gusongdong Yusonggu, Daejeon 305-701, Republic of Korea;Department of EECS, Korea Advanced Institute of Science and Technology, Gusongdong Yusonggu, Daejeon 305-701, Republic of Korea

  • Venue:
  • Journal of Systems Architecture: the EUROMICRO Journal
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

DVS is becoming an essential feature of state-of-the-art mobile processors. Interval-based DVS algorithms are widely employed in general purpose operating systems thanks to their simplicity and transparency. Such algorithms have a few problems, however, such as delayed response, prediction inaccuracies, and underestimation of the performance demand. In this paper we propose TSB (time slice based), a new DVS algorithm that takes advantage of the high transition speeds available in state-of-the-art processors. TSB adjusts processor performance at every context switch in order to match the performance demand of the next scheduled task. The performance demand of a task is predicted by analyzing its usage pattern in the previous time slice. TSB was evaluated and compared to other interval-based power management algorithms on the Linux kernel. The results show that TSB achieved similar or better energy efficiency compared to existing interval-based algorithms. In addition, TSB dramatically reduced the side effect of prolonging short-term execution times. For a task requiring 50ms to run without a DVS algorithm, TSB prolonged the execution time by only 6% compared to results of 136% for CPUSpeed and 20% for Ondemand.