High-performance computing (HPC) systems face escalating challenges in resource allocation as workloads grow increasingly heterogeneous and computational demands reach exascale levels. Traditional scheduling methods struggle to effectively manage diverse and fluctuating workloads, while recent machine learning approaches fail to fully capture the complex interactions between jobs, resources, and system states. We present GNN-RL, a framework that integrates Graph Attention Networks (GAT) with Proximal Policy Optimization (PPO) to address fundamental limitations in HPC scheduling. Our approach complements recent uncertainty-aware and digital twin-based schedulers by providing efficient graph-structured state representations that capture complex job dependencies and resource topologies. Through multi-head attention mechanisms and sophisticated state representations, GNN-RL simultaneously optimizes resource utilization, job throughput, makespan, and system fairness. A comprehensive evaluation of 377,234 jobs from three Argonne Leadership Computing Facility (ALCF) systems, Polaris, Mira, and Cooley, demonstrates significant improvements over EASY-backfilling: 13.8%–20.2% resource utilization gains, 15.0%–50.0% throughput increases, and 12.1%–12.8% makespan reductions, with statistical validation confirming significance and large effect sizes. Job-level analysis reveals that medium-sized batch jobs (16–128 nodes) benefit most from graph-based scheduling (＋15.3 percentage points utilization), while large jobs ( > 128 nodes) show smaller gains with a slight fairness trade-off. The framework maintains 76% effectiveness at 50 K concurrent jobs, demonstrating strong potential for practical application in HPC environments.

Exascale high-performance computing (HPC) systems face critical resource management challenges such as massive energy consumption in megawatts per facility, performance variability for identical jobs, and resource utilization inefficiencies. Traditional single-objective schedulers cannot address these multifaceted challenges effectively. This paper introduces HARMONIC (Holistic Adaptive Resource Management Optimizing Next-generation Interconnected Computing), a novel framework that simultaneously optimizes performance, energy efficiency, and resilience through uncertainty-aware multi-objective optimization. Our approach distinguishes aleatoric uncertainty (inherent system variability) from epistemic uncertainty (modeling limitations) using Bayesian neural networks and employs graphbased representations to capture complex system dependencies. Experimental validation in both simulated environments and controlled testbeds demonstrates significant improvements over state-of-the-art schedulers: 10-19% energy reduction, 16-25% throughput improvement and 18-32% performance variability reduction. These results translate to potential annual savings of multimillion dollars per exascale facility while enhancing scientific productivity through improved experimental reproducibility.