Statistical Power and Performance Modeling for Optimizing the Energy Efficiency of Scientific Computing

High-performance computing (HPC) has become an indispensable resource in science and engineering, and it has oftentimes been referred to as the "thirdpillar" of science, along with theory and experimentation. Performance tuning is a key aspect in utilizing HPC resources to the fullest extent. However, recent exascale studies suggest that power and energy consumption will be a major impediment to HPC in this coming decade. Therefore, performance tuning should evolve and take energy consumption into account. Unfortunately, the increase in system complexity and the number of tunable parameters in applications makes the performance tuning of an application cumbersome. To address these issues, we propose energy-efficient tuning via statistical regression techniques. Such techniques can be used to model the power and performance of a scientific application, and then the application parameters can be tuned to achieve the best energy efficiency possible, based on metrics such as the performance-to-power ratio. In this paper, we utilize multi-variable regression to model the power and performance of the high-performance LINPACK (HPL) benchmark. We then tune the HPL parameters for energy efficiency and compare them to the energy efficiency achieved at maximum possible performance(R
_max
). Our results show that statistical regression modeling can be used for predicting the HPL configuration for achieving the maximum energy efficiency with very high accuracy.