DRL-MOSHRS: A Deep Reinforcement Learning Approach for Multi-Objective Scheduling in Heterogeneous HPC Systems

Modern HPC systems with heterogeneous compute nodes (CPU-GPU combinations) offer enhanced processing capabilities but introduce complex resource orchestration challenges. This paper presents DRL-MOSHRS, a deep reinforcement learning approach for multi-objective scheduling in heterogeneous high-performance computing systems. DRL-MOSHRS combines a priority-based elastic node orchestration algorithm with a D3QN-based resource scheduling strategy to optimize multiple objectives, such as execution time, waiting time, and resource costs. Through experimental evaluation on public datasets, it has been demonstrated that DRL-MOSHRS attains superior efficiency in comparison to existing baseline algorithms across a range of cluster load scenarios, particularly in the context of machine learning workloads. Our approach demonstrates particular efficacy in resource-constrained environments by maintaining low average waiting times and stable rental costs through the implementation of elastic scaling capabilities.