Efficient Resource Utilization through GPU Resource Pooling Management and Control Engine
Qiang Zhang, Xuehua Chen, Decheng Wang, Xurun Lin · 2025
Aiming at the problem of low efficiency of GPU resource scheduling in cloud computing, this paper proposes a GPU Resource Pooling Management and Control Engine based on Kubernetes. The traditional GPU scheduling method of subsystem is difficult to coordinate resources, and faces the problem of inefficient utilization of GPU resources caused by the difference of task priorities. In business scenarios such as remote sensing image processing and AI model training, flexible dynamic allocation and release mechanisms of GPU resources are required to improve resource utilization. On the basis of resource virtualization based on MIG technology of Kubernetes and NVIDIA A100 GPU, the engine realizes unified scheduling and efficient utilization of GPU resources through key technologies such as pooling management, resource release strategy and task execution strategy. Compared with the cloud platform without any scheduling platform, the throughput of the cluster is doubled, and the utilization of GPU resources is increased by $\mathbf{2 0 \% \sim 2 8 \%}$. The GPU Resource Pooling Management and Control Engine provides a new idea and method for the efficient management and utilization of GPU resources in cloud environment.