Deep RL for HPC job scheduling — statistical evaluation finds MaskablePPO best on heterogeneous trace

KEY POINTS

1. This arXiv preprint compares six deep reinforcement learning (DRL) treatments (PPO, A2C, DQN, each with and without action masking) against three heuristics and a seed-paired masked uniform random control on two real Slurm workload traces (homogeneous and heterogeneous) using a non‑parametric framework.
2. Masked variants completed evaluation while unmasked ones were truncated; on the homogeneous trace no masked DRL outperformed the masked random control, whereas on the heterogeneous trace MaskablePPO improved average waiting time, average bounded slowdown, and maximum bounded slowdown (max slowdown 22% lower than LCFS), had the best average waiting-time rank among complete treatments, the lowest mean training time among masked variants, and MaskableDQN showed lowest decision latency.
3. Clarifies that action masking and a masked random control are essential to isolate true learned scheduling gains and shows MaskablePPO can yield measurable benefit on heterogeneous HPC workloads but not universally.