EClass: An execution classification approach to improving the energy-efficiency of software via machine learning

  • Authors:

  • Edward Y. Y. Kan;W. K. Chan;T. H. Tse

  • Affiliations:

  • Department of Computer Science, The University of Hong Kong, Pokfulam, Hong Kong;Department of Computer Science, City University of Hong Kong, Tat Chee Avenue, Hong Kong;Department of Computer Science, The University of Hong Kong, Pokfulam, Hong Kong

  • Venue:
  • Journal of Systems and Software
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Energy efficiency at the software level has gained much attention in the past decade. This paper presents a performance-aware frequency assignment algorithm for reducing processor energy consumption using Dynamic Voltage and Frequency Scaling (DVFS). Existing energy-saving techniques often rely on simplified predictions or domain knowledge to extract energy savings for specialized software (such as multimedia or mobile applications) or hardware (such as NPU or sensor nodes). We present an innovative framework, known as EClass, for general-purpose DVFS processors by recognizing short and repetitive utilization patterns efficiently using machine learning. Our algorithm is lightweight and can save up to 52.9% of the energy consumption compared with the classical PAST algorithm. It achieves an average savings of 9.1% when compared with an existing online learning algorithm that also utilizes the statistics from the current execution only. We have simulated the algorithms on a cycle-accurate power simulator. Experimental results show that EClass can effectively save energy for real life applications that exhibit mixed CPU utilization patterns during executions. Our research challenges an assumption among previous work in the research community that a simple and efficient heuristic should be used to adjust the processor frequency online. Our empirical result shows that the use of an advanced algorithm such as machine learning can not only compensate for the energy needed to run such an algorithm, but also outperforms prior techniques based on the above assumption.