Workshops Archive

Decay-Driven Multi-Objective Reinforcement Learning for Preference-Aware Power Control in HPC


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

Authors: Akhilesh Raj (Vanderbilt University, ORNL); Swann Perarnau (Argonne National Laboratory); and Aniruddha Gokhale (Vanderbilt University)

Abstract: Energy efficiency is a fundamental constraint in HPC systems, where electricity costs rival hardware costs and power budgets limit deployable performance. RL suits runtime power control, but online training is impractical in production, and offline RL needs a separate agent per operating point since preference weights are fixed beforehand. We present DDMORL, a Decay-Driven Multi-Objective RL framework for single-node energy management that exposes one application- and hardware-agnostic control: a user-specified scalar acceptable performance decay δ ∈ [0,1]. A lightweight, application-specific progress-to-power-cap model generates preference-aligned signals from this target, removing the need for learned interpolators or per-point agents, while letting the offline-trained policy control power live via Intel's RAPL. We evaluate DDMORL on compute- and memory-bound HPC benchmarks, showing decay levels map reliably to power caps delivering substantial energy savings within the requested degradation limit; the approach is application- and hardware-agnostic and generalizes to other nodes for production deployment.


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