SIParCS 2021 - Spencer Diamond
Lowering the Cost of Climate Research: Energy Consumption vs Clock Speed for Various Application Profiles
High Performance Computing has been a boon to the world of climate research, but its use comes with large costs, not only financial but also environmental costs. One way to reduce these costs is by lowering the energy consumption of these systems. However, most HPC optimization is focused on time efficiency, with energy efficiency largely being of secondary importance. The goal of this project was to develop a methodology for collecting reliable power data during execution in order to understand how to improve the energy efficiency of HPC systems without significantly impacting performance. In order to streamline the development of this methodology, this project’s scope was limited to 64GB nodes on NCAR’s Cheyenne computer. Experiments were focused on measuring the effect that CPU clock frequency has on energy consumption for single nodes. The experiments were performed on nodes running three common performance benchmarks as well as on idle nodes. Data was collected from the node itself during execution, and from the power supplies feeding the node. Although further exploration is necessary, the results of this project suggest that managing CPU frequency both during execution and while idle could be a viable way to boost energy efficiency and reduce the cost of HPC systems without sacrificing performance.
Mentors: Dave Hart, Rory Kelly, & Ben Matthews
Slides and poster