Hierarchical three-way decision fusion for multigranularity GPU-CPU coscheduling in hybrid computing systems

In heterogeneous computing environments, coscheduling of the graphics processing unit (GPU) and central processing unit (CPU) poses substantial challenges because of the diverse hardware architectures and dynamic workload patterns. To address this, we propose a novel hierarchical three-way decision fusion (H3WDF) strategy that integrates multigranularity workload predictions and adaptive scheduling policies. H3WDF employs a three-tier decision-making process, achieving global coordination through selective aggregation of localized decisions while establishing a balance between efficiency and quality of service. Results of experimental evaluation in a heterogeneous environment comprising several GPUs demonstrate the superior performance of H3WDF across multiple metrics. For "large language model" workloads, H3WDF achieves remarkable prediction accuracy both for short- and long-term forecasts. H3WDF's three-way decision mechanism effectively distributes workloads, balancing between batched executions for training tasks and immediate executions for inference workloads. Resource utilization exhibits significant improvements across all GPU types, with particularly strong performance in the case of high-end GPUs. Compared with the state-of-the-art baselines, H3WDF substantially reduces job completion times, enhances energy efficiency, and consistently maintains high fairness in resource allocation.