Optimizing virtual machines scheduling on high performance network NUMA systems
Junsheng Tan, Fuzong Wang · 2017
Non-Uniform Memory Access (NUMA) architecture is being widely used in virtualization and cloud computing due to its scalability. In NUMA system, previous works mainly focused on how affinity between processor and memory could be optimized by scheduling. However, we observed that in NUMA system with high-performance virtualized network enable, the network affinity is also an important factor affecting the performance of VM. In this paper, we analyze the affinities among processor, memory and NIC in depth. Based on the analysis, we designed and implemented a VM scheduler, named as vNAASched, in KVM. vNAASched monitors and collects performance events of VMs at runtime and schedules VMs based on the affinities among processor, memory and NIC. Experiment results showed that vNAASched is able to improve performance of real-world Apache Spark workloads by up to 36%.