Dynamically Tracking Performance-based Working Set Size and Its Application to Energy-Efficient Memory Management.

Pin Zhou, Anand Raghuraman, Yuanyuan Zhou, Sanjeev Kumar.

As University of Illinois at Urbana Champaign Technical Report 2315, January 2003.

Memory can be efficiently utilized if an application's dynamic memory demands can be analyzed at run time. This paper proposes a method that uses Mattson's Stack algorithm to accurately track an application's dynamically changing working set size at run time. Tracking working set sizes at run time has several applications. For example, it can be used to save memory energy consumption when extra active memory provides little performance gain. Moreover, in a multi-programmed setting, it can also be used to efficiently allocate memory to different applications based on their dynamic memory needs.

This paper also demonstrates how tracking dynamic working set sizes can be used for energy-efficient memory power management. We investigate a run-time, adaptive energy-efficient memory power management scheme to reduce DRAM energy consumption while still providing acceptable performance. Our execution-driven simulation results using enhanced SimpleScalar with SPEC2000 benchmarks show that our method can save DRAM energy consumption by up to 74% while slowing down applications by only 0.4-4.3%.