Abstract:
This study employs control theory to optimize power regulation in large HPC systems. It dynamically adjusts processor power caps based on real-time application progress to enhance energy efficiency while maintaining computational performance. The approach incorporates cascaded control strategies, such as PI control and MPC, integrated into the Argo Node Resource Manager framework. Effectiveness is assessed across Grid’5000 clusters using standard HPC benchmark and Intel’s RAPL mechanism. The research aims to enhance energy efficiency in high-performance computing while meeting computational demands.
