An Online Heterogeneous GPU Cluster Scheduler for Multi-Objective Optimization

The widespread adoption of large-scale deep learning models has driven an increasing demand for GPU clusters. A key challenge is how to efficiently schedule jobs in a heterogeneous GPU cluster to improve cluster utilization while meeting job deadline requirements as much as possible. To address these needs, this paper presents a Heterogeneous GPU Cluster Scheduler (HGCS), which formulates the above objectives as an optimization problem and then reformulates it as an integer linear programming problem. The solution is approached using a primal-dual framework combined with a greedy-based online scheduling algorithm. Simulation experiments on the open-source Sia simulator demonstrate that HGCS achieves significant improvements in both deadline satisfaction rate and average job cost compared to previous approaches.