GAS-MARL: Green-Aware job Scheduling algorithm for HPC clusters based on Multi-Action Deep Reinforcement Learning

Abstract

In recent years, the computational power of High-Performance Computing (HPC) clusters has surged. However, amidst global calls for energy conservation and emission reduction, their rapid power consumption poses a developmental bottleneck. Adopting renewable energy sources for power supply is a crucial measure to reduce carbon emissions from HPC clusters. However, due to the variability and intermittency of renewable energy, formulating effective job scheduling plans to fully utilize these sources has become urgent. To tackle this, we propose a Green-Aware job Scheduling algorithm for HPC clusters based on Multi-Action Deep Reinforcement Learning (GAS-MARL), which optimizes both renewable energy utilization and average bounded slowdown. In this algorithm, the agent outputs two actions during one decision-making period: job selection action and delay decision action. The introduction of delay decision actions enhances the flexibility of the scheduling algorithm, enabling each job to be executed during appropriate time slots. Furthermore, we have designed a new backfilling policy called Green-Backfilling to better cooperate with GAS-MARL for job scheduling. Experimental evaluations demonstrate that, compared to other algorithms, the combination of GAS-MARL and Green-Backfilling exhibits significant advantages in enhancing renewable energy utilization and decreasing average bounded slowdown.