OPTIMIZATION OF GPU RESOURCE SHARING SCHEDULING FOR CONTAINER CLUSTERS
In a container cluster environment,the entire physical GPU resource can usually only be scheduled exclusively by a single container,and there is a lot of waste of resources.Existing GPU sharing scheduling schemes still have problems of scheduling failure,high resource overhead and lack of resource isolation.Improved GPU sharing used the LD_PRELOAD mechanism to effectively isolate GPU memory resources,and it optimized the original scheduling algorithm,so that the utilization of cluster video memory resources was greatly improved.The experimental results verify the effectiveness of the improved GPU Sharing in the realization of resource isolation.At the same time,the improved GPU sharing has only 1.008％extra overhead for executing applications on the physical machine,and the optimized scheduling algorithm has increased by 53.01％GPU memory utilization.
国家科技期刊平台
NSTL
万方数据
