NUMA-aware virtualization: Hiding physical asymmetry for virtual machines

Abstract

Virtualization has become popular to improve system utilization by consolidating multiple servers into one physical system. In addition to the improved utilization, other benefits of virtualization such as flexible resource management, fault isolation, and support for different operating systems, led to the increased interest in the virtualization. Virtualization changes the conventional computing structure. In the traditional computing structure, operating systems recognize the physical hardware asymmetry but the operating system in a virtual machine only knows the abstract hardware configuration provided by the hypervisor. In this environment, it is not meaningful that the operating system in a virtual machine recognizes the physical hardware asymmetry provided by the hypervisor. Because the virtual machine can migrate from a source machine to another machine with different hardware status. For example, virtual machines can migrate from a uniform memory access architecture machine to a non-uniform memory access architecture machine. In this situation, the hypervisor needs to be aware of the virtualized environment and provide the generalized abstract hardware to each virtual machine by converting the asymmetric hardware architecture to symmetric virtual hardware architecture. However, current virtualization technology cannot handle non-uniform memory architecture correctly. Thus we propose a scheduling policy which understands the non-uniform memory access architecture in the Xen virtual machine monitor. It is called NUMA-aware scheduling policy. In the NUMA-aware scheduling policy, the number of memory requests in each node is monitored by the hot page number sampling method. Based on the number of memory requests in each node, each virtual machine selects the favor node with frequent memory requests. The vCPUs in each virtual machine is scheduled in the favor node with a high priority. The NUMA-aware scheduling policy achieves 27% performance improvement i...