Deep Reinforcement Learning for SLA-Aware GPU Scheduling Under Workload Uncertainty

GPU cluster scheduling involves making choices to meet service-level agreement (SLA) goals while minimizing queue delay and maximizing resource utilization when jobs have variable runtimes. In this paper, we introduce an SLA-aware scheduling framework, using proximal policy optimization (PPO). The proposed observation is a nine dimensional vector in the global space and a tensor of candidate job features, with thirteen features for each of ten candidate jobs. Historical runtime summaries/histograms are provided to indicate duration variability, and action masking is provided to exclude selections of resources that are not feasible. A weighted reward is a combination of penalties for queue-threshold violations and waiting time, and a utilization incentive. The following notebook contains pre-defined values for the scheduler comparisons and ablation, an example of a runtime distribution and scripted episode sequences. These materials show a way of presenting and visualizing numbers, not an actually conducted scheduling experiment. The predefined PPO entry has a lower SLA violation value than the listed comparison entries, but there is no trace-loading pipeline, scheduling simulator, PPO implementation or independent evaluation logs provided in the supplied notebook. Therefore, the numerical comparisons do not provide a basis for learned-policy superiority or convergence, effectiveness of uncertainty-models, or production-trace performance. This is a proposal for a decision making framework and an explicit statement of the evidence needed for evaluation. Empirical validation, sensitivity testing and repeated run assessment are future works.