Research on power saving and energy efficiency for data-centric computing on production HPC systems

Abstract

Energy consumption has become one of the four key challenges when Department of Energy (DOE) plans to deploy its first exascale supercomputer by 2021. With advanced power management technologies, building a power-saving and energy-efficient exascale computer with faster computing capability becomes possible. To achieve this goal, it is crucial to characterize the power and energy consumption on modern computer hardware and runtime systems. In my dissertation research, I configure the execution environment in various settings and profile and characterize the power and energy consumption on Intel Haswell Enterprise processor, which is currently used in the new Trinity supercomputer hosted at Los Alamos National Laboratory. I also study the power and energy characteristics of a data-centric HPC runtime system, i.e., the Legion runtime and applications, jointed developed by Stanford University and Los Alamos National Laboratory. The experimental results show that the settings of P-States and hyperthreading have a significant impact on the power and energy consumption. Additionally, the data-centric Legion runtime and applications achieve better energy efficiency compared with the MPI computing paradigm.