Evaluation of virtual machine performance on NUMA multicore systems

An increasing number of new multicore server systems use the Non-uniform Memory Access (NUMA) architecture due to its scalable memory performance. However, the multicore NUMA systems introduce complex interplay among data locality, contention on shared resources, and inter-socket data sharing overhead, which makes the achievement of an optimal and predictable system performance difficult. Virtualization technology further complicates the scheduling problem. In this paper, we first investigate the impact of virtual machine (VM) scheduling on multicore NUMA systems. Then, we analyze different VM mapping combinations and find that the best VM scheduling strategy does not only depend on the data sharing and memory characteristics of the VM itself, it is also impacted by the dynamic behavior of co-running VMs.