Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices
Authors: Solomon Bekele (Argonne National Laboratory); Nick Eliopoulos (Purdue University); and Swann Perarnau, Servesh Muralidharan, and Brice Videau (Argonne National Laboratory)
Abstract: Node-level power capping is a scalable mechanism for enforcing electrical, thermal, and scheduling constraints on exascale systems. On a heterogeneous node, however, a single cap hides a second decision: how platform firmware distributes the constrained budget among GPUs, CPU cores, memory, and the CPU uncore. This paper characterizes the hidden allocation on the Aurora supercomputer. More importantly, under constrained node caps, firmware can keep the CPU uncore near its minimum even when application progress depends on data movement, increasing runtime enough to raise total board energy. We introduce U-PullUp, a lightweight GEOPM agent that uses recent GPU copy-engine activity as a data-movement signal and requests a higher uncore frequency while leaving the node cap active. Across the nine SPEChpc workloads, U-PullUp reduces board energy by up to 34.2% relative to default power control at the same cap, while the best configuration for each workload reduces energy by 19.6% on average.
Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing
Back to Full Workshop Archive Listing