Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices
Authors: Marco De Carlo (CMCC Foundation), Gabriele Padovani (University of Trento), Gabriele Accarino (Columbia University), Valentine Anantharaj (Oak Ridge National Laboratory), Paola Nassisi (CMCC Foundation), and Sandro Fiore (University of Trento)
Abstract: Hyperparameter optimization (HPO) has become a fundamental component of deep learning (DL) workflows, yet executing long-running optimization campaigns on High Performance Computing (HPC) systems remains a challenge, due to sustainability, computational costs, interruptions, heterogeneous hardware, and limited runtime visibility. Existing HPO frameworks primarily optimize the search effectiveness, while operational aspects required by HPC environments are neglected. This work introduces OBSERVE, an HPC-native framework for Bayesian Optimization (BO) offering energy-aware multi-objective optimization, fault tolerant checkpointing, and runtime observability. The framework is validated on the ClimateBench and MLPerf-DeepCAM benchmarks across heterogeneous HPC platforms with different schedulers and GPU architectures, and execution is tested by parallelizing the DL training in multi-GPU environments and the HPO process itself. Results confirm that the framework preserves the optimization effectiveness of BO while enabling resilient execution and exposing energy-performance trade-offs. More broadly, this work lays the foundation for energy-aware HPO as a runtime service in HPC infrastructures.
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