Analysis of the Impact of CPU Virtualization on Parallel Applications in Xen

Recently, Infrastructure as a Service (IaaS) is being used for parallel computing. In IaaS clouds, parallel applications are run in virtual machines (VMs), which own virtual CPUs (vCPUs). Application threads are scheduled to vCPUs and then vCPUs are scheduled to physical CPUs (pCPUs). This CPU virtualization can affect the performance of parallel applications. According to our experiments in Xen, the scalability of parallel applications in a VM was largely different from that in a physical machine (PM). In this paper, we analyze the root cause of such a difference in scalability. As a result of our investigation, we found that the root cause was both resource conflicts between pCPUs and conflict in vCPU scheduling. In addition, we provide three methods for avoiding these conflicts and improving scalability. Furthermore, we confirmed that the optimal vCPU scheduling could be effective for not only applications parallelized by Tascell but also most of NAS Parallel Benchmarks.