RCS: Hybrid Co-scheduling Optimization in Virtualized System

摘要

As a support technology for cloud computing, virtualization enables multiple guest operating systems run on a system, which will improve utilization of resource. However, due to the semantic gap in the virtualization system, the mainstream of the current scheduling policy doesn't take the tasks' spin lock and cache misses into account, which leads to the performance degradation in virtual machine. In this paper, we propose an optimization of the hybrid co-scheduling in KVM called Regional Co-Scheduling (RCS). Our solution includes two aspects: co-scheduling the vCPUs of the concurrent VM within CPU cores in the same CPU socket, and optimizing the load balancing by adding two strategies, which include that vCPUs belonging to the same VM should not be in the same queue and cross-region migration should be avoided. The experiment results show that RCS significantly reduces the execution time and context switches, compared with the default scheduler CFS and Co-Scheduling, and suffers less overhead as well, while when the number of VMs is large, it is with fair-share feature.