Analysis of SMP VM CPU Scheduling

System virtualization allows resources of a single physical machine to be shared among multiple virtual machines (VMs). This can increase system flexibility, but it can also invalidate some assumption that the OS running in a VM has about the system it is executing on. This paper surveys the implementation of symmetric multiprocessing (SMP) in a virtualized environment. It examines two proposals for optimizing the implementation of SMP in VMs, and it analyzes the implementation of SMP VMs for two popular virtualization systems: VMware ESX server and Xen. The benchmark results show that SMP VMs do experience greater overhead than uniprocessor VMs for computationally intensive workloads. However, optimizations in ESX server and Xen reduce the overhead below what it would be for the most naive implementation.