Workshops Archive

Adaptive Control with Offline Reinforcement Learning for HPC Power Management


Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices

Authors: Elias Klett and Christian Simmendinger (HPE)

Abstract: Power-capping policies for high-performance computing must reduce energy use without imposing unacceptable delays across heterogeneous applications. PowerSched, deployed on CPU nodes of the Hunter supercomputer at HLRS, identifies per-job operating points through iterative Golden-Section Search (GSS), which requires multiple measurement intervals and can respond slowly to workload changes. We replace this search with an Implicit Q-Learning controller trained offline on approximately 13 million telemetry transitions from 15,000 historical jobs. Using only telemetry already available to PowerSched, the controller requires no application instrumentation and predicts a bounded per-socket power cap every 15 seconds. We integrate it as a prototype behind PowerSched's predictor interface and conduct preliminary single-node experiments on HPL, HPCG, and eight NAS Parallel Benchmarks. Relative to a 400 W/socket maximum-cap baseline, the controller reduces measured node energy by 7.1% across all ten benchmarks with a 1.0% average performance reduction. The results use three independent runs per benchmark.


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