Co-Design of a Power State-Aware Scheduler and an Intelligent Power Manager for Energy-Efficient HPC Systems

Power consumption is a major challenge in large-scale High Performance Computing systems, where idle nodes contribute to a significant portion of the overall system’s power consumption. This highlights the need for effective power management strategies capable of deciding when and which nodes should be powered on or off. While simple timeout-based heuristics and more advanced Reinforcement Learning (RL)-based approaches have been explored to enable autonomous power control, existing designs are often not paired with power-state-aware schedulers, hindering their effectiveness. To address these issues, this research introduces a co-designed solution comprising the Power-State-Aware Scheduler (PSAS) and an Intelligent Power Manager (IPM), collectively referred to as PSAS+IPM. PSAS offers clean integration with external power policies by ensuring resource and power-state awareness, while IPM augments scheduling with predictive power control and energy-aware node selection to minimize delay and waste. Extensive simulation results show that PSAS+IPM outperforms both traditional timeout-based and RL-driven strategies. PSAS+IPM also exhibits greater stability and robustness across a wide range of hardware configurations, including varying power transition latencies, power consumption profiles, and system sizes. These findings highlight that tightly coupled scheduler-power manager co-design can yield greater efficiency than more complex but decoupled RL approaches, underscoring coordination as a key determinant of sustainable exascale computing.