VMware edges out Microsoft in virtualization performance test

VMware started out ahead in this race with Windows 2008 and SLES 10.2 virtual performance nearly as fast as native performance, and held close to that pace with three guest operating systems. Hyper-V with three VMs in place was about 1,400 bops off VMWare's pace with Windows 2008 guests and 1,800 bops down from ESX mark with SLES VMs.

At six VM guests, both hypervisors are starting to struggle to deliver performance comparable to what a native operating system running directly on the server can pull off. But Microsoft kept its performance drop a bit more in check as it appears to have mastered a more linear distribution of hypervisor resources when VMs get piled on.

In reality, consolidated instances aren't necessarily as burdened at the pace we placed on the instance by running concurrent SPECjbb2005 tests. Many operating system and application instances typically have far less constant CPU utilization than SPECjbb2005 places on them, and the utilization is often more random in nature. We've stressed the VMs and the hypervisors supporting them to amplify how each hypervisor reacts under enormous loads.

In the second round of iterative VM tests we allowed each VM to have access to four vCPUs, the maximum allowed by either hypervisor under test. Each VM was still limited by 2GB of memory as it's a common ceiling when consolidating and testing an operating system. This test scenario more readily demonstrates how VMs would be used in virtualized database applications, rendering farms, high-volume transaction systems and other applications needing strong CPU availability.

As before, we started with a single VM guest to establish a baseline, then added two more VMs for a total of three instances, then three more for a total of six VMs. In the first test, as we noted in our cost of virtualization test, VMware pulls slightly ahead when hosting Windows 2008 clients and has almost an 1100 bops advantage when hosting SLES 10.2 VMs. Because Microsoft's LinuxIC kit isn't supported for SMP environments, Hyper-V's performance with SLES is dampened without the boost it provided in the tests where we could allocate a single vCPU to each VM.

In the test where three VMs were each using four vCPUs, 12 vCPUs were in play. Because there were 16 physical CPU cores on the server in our test bed that could be virtualized by the hypervisors under test, there were four CPUs sitting idle. Hyper-V pulls ahead of VMware ESX in this instance with on average 6,500 more bops. Our test results suggest that Hyper-V could see those extra available hardware resources and tapped into them, whereas ESX could not.