A user-level NUMA-aware scheduler for optimizing virtual machine performance

Abstract

Commodity servers deployed in the data centers are now typically using the Non-Uniform Memory Access (NUMA) architecture. The NUMA multicore servers provide scalable system performance and cost-effective property. However, virtual machines (VMs) running on NUMA systems will access remote memory and contend for shared on-chip resources, which will decrease the overall performance of VMs and reduce the efficiency, fairness, and QoS that a virtualized system is capable to provide. In this paper, we propose a "Best NUMA Node" based virtual machine scheduling algorithm and implement it in a user-level scheduler that can periodically adjust the placement of VMs running on NUMA systems. Experimental results show that our NUMA-aware virtual machine scheduling algorithm is able to improve VM performance by up to 23.4% compared with the default CFS (Completely Fair Scheduler) scheduler used in KVM. Moreover, the algorithm achieves more stable virtual machine performance. © 2013 Springer-Verlag.