VMware edges out Microsoft in virtualization performance test

The limitations imposed by the hypervisor vendors on the number of available vCPUs come from two areas. First, keeping track of VM guests with very large CPU needs also involves enormous memory management and large amount of inter-CPU communications (including processor cache, instruction pipelines and I/O state controls) that are exceedingly difficult. Secondly, the demand for VM guest hosting has been perceived to be a server consolidation action - and servers that need consolidating are often single CPU machines.

These limitations in hypervisor hardware resource allocations set the stage for how we could take advantage of the 16-CPU HP DL580G5 server in our test bed.

As previously noted, Microsoft officially supports its own operating systems and Novell's SLES 10 (editions running Service Packs 1 and 2) as guest instances. That accounts for why we tested with only Windows 2008 and SLES 10.2 VMs. Other operating systems (Red Hat Linux, Debian Linux and NetBSD) may work, but organizations seeking debugging or tech support are on their own if they use them.

While we were testing, Microsoft introduced its Hyper-V Linux Interface Connector (Hyper-V LinuxIC) kit, which is a set of drivers that help optimize CPU, memory, disk and network I/O for SLES guest instances. We did see a boost in performance with the kit in place, but only in the case of one vCPU per guest. Hyper-V LinuxIC isn't supported for SMP environments.

The cost of virtualization

No one is claiming the buzz about server virtualization is unsubstantiated. It lets you pack multiple operating system instances onto the same hardware that previously only hosted one instance. And it helps in deploying a standard operating system profile across the data center, if that is your goal.

But nothing is free. Hypervisors become the basic operating system of the servers that they virtualize, which taxes performance. Our first test measures the cost of virtualization by comparing transactional performance when an operating system is running on bare metal with the performance of that same operating system when a hypervisor serves as a buffer between the operating system and the system. The difference in performance amounts to a theoretical tax imposed by the hypervisor's innate management role.

In our tests, the performance hit when we moved from a native operating system instance to a virtualized one with a single vCPU allotted, ranged from about 2.5 percent when ESX was running Windows 2008 to more than 12 percent when Hyper-V was running SLES. The foundational performance 'cost' of each hypervisor varied, but VMware wins this theoretical round. It's theoretical because there are few cases for running a virtual machine platform with only a single guest limited to a single CPU.