容器集群GPU资源共享调度优化

OPTIMIZATION OF GPU RESOURCE SHARING SCHEDULINGFOR CONTAINER CLUSTERS

在容器集群环境中,整块的物理GPU资源通常只能被单个容器独享调度,存在大量的资源浪费。现有的GPU共享调度方案中仍存在调度失败、资源开销大或没有实现资源隔离的问题,改进的GPU Sharing利用LD_PRELOAD机制有效地实现了GPU显存资源的隔离,并优化了原有的调度算法,极大提高了集群显存资源的利用率。实验结果验证了改进后GPU Sharing在资源隔离实现上的有效性,同时,改进后的GPU Sharing同在物理机上执行应用程序只多了1.008%的额外开销,而且优化后的调度算法提高了53.01%的GPU显存利用率。

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.