Parallel and Distributed Systems Group

Computer Science Department of Telecom SudParis

Uncore Power Scavenger: A Runtime for Uncore Power Conservation on HPC Systems

Reading group: Thomas Collignon presented "Uncore Power Scavenger: A Runtime for Uncore Power Conservation on HPC Systems" (SC'19) at 1C27 the 21/10/2022 at 10h30.

Abstract

The US Department of Energy (DOE) has set a power target of 20-30MW on the first exascale machines. To achieve one exaflop under this power constraint, it is necessary to minimize wasteful consumption of power while striving to improve performance.

Toward this end, we investigate uncore frequency scaling (UFS) as a potential knob for reducing the power footprint of HPC jobs. We propose Uncore Power Scavenger (UPSCavenger), a runtime system that dynamically detects phase changes and automatically sets the best uncore frequency for every phase to save power without significant impact on performance. Our experimental evaluations on a cluster show that UPSCavenger achieves up to 10% energy savings with under 1% slowdown. It achieves 14% energy savings with the worst case slowdown of 5.5%. We also show that UPSCavenger achieves up to 20% speedup and proportional energy savings compared to Intel’s RAPL with equivalent power usage making it a viable solution even for power-constrained computing.