AI-Driven Multi-Dimensional Computing Power Scheduling: Adaptive Resource Coordination via Task Demand Prediction in Heterogeneous Clusters

Abstract

This paper introduces an AI-driven computing power scheduling framework that innovatively integrates multidimensional resource optimization with machine-learning-based task-demand prediction to significantly enhance computational efficiency and resource utilization. Unlike prior works that primarily focus on Graphics Processing Unit (GPU) allocation, our method pioneers a holistic resource-coordination mechanism that dynamically balances GPU, memory, Central Processing Unit (CPU), and other critical resources according to their joint impact on task performance and cluster efficiency. A core innovation is our data-driven resource predictor, which autonomously analyzes historical task patterns and forecasts future demand, enabling the scheduler to adaptively scale resources so that user over-provisioning is reduced and under-allocation bottlenecks are avoided. Experimental validation demonstrates that this closed-loop prediction–scheduling paradigm achieves breakthroughs in both scale and efficiency: a 25.7% increase in concurrent task deployment, a 15.6% increase in task completion rate, and substantial relative utilization improvements of 7.5% for GPU and 8.0% for memory, outperforming conventional single-resource optimization approaches. These advancements establish a new direction for intelligent resource management in large-scale heterogeneous computing environments.OPEN ACCESS Received: 15/05/2025 Accepted: 11/10/2025 Published: 23/01/2026

Document