VMware edges out Microsoft in virtualization performance test

This advantage is lost, however, when we oversubscribe as we did in the final round of testing. Oversubscription is a method that allocates more physical CPU than is available, allowing VMs to "share" their allocated vCPUs with other VM guests. It's a process that is useful when VMs are running applications that use CPU power randomly, as it lets you stuff more VMs while hopefully (dependent on guest activities) offering performance at or above what the guests did before they were virtualized.

Six VM guests each using four vCPUs oversubscribes the 16 physical CPU cores in our test rig. Both hypervisors are starting to buckle under an extreme load as CPU power is at a premium in this stressful test. But VMware seems to deal with oversubscription better than Hyper-V as it could still pull down an average of 16,136 bops with Windows 2008 guests (compared with Hyper-V's 14,588 bops) and 17,089 bops with SLES guests (compared with Hyper-V's 11,122 bops). Microsoft also is slightly disadvantaged in oversubscription because a native instance of Windows 2008 Server (we used Enterprise Edition) needs to be active to run the Hyper-V hypervisor system - using up its own space and CPU.

The disk I/O seen in a VM light

We also tracked disk throughput of hosted VMs with Intel's IOMeter (pre-compiled Windows and Linux versions). IOMeter exercises disk subsystems by spawning worker threads that read and write to the subsystem in a tester-defined routine. Measurements are summarized in terms of IOs per second as recorded by IOmeter at the end of a test run. The results are expressed in terms of IO's per second. A higher number of IOs is better.

In a virtualized world, VM guest instances must contend with either internal disk or storage-area network resources. When the hardware is re-represented to guest operating systems through virtualization, the hypervisor layer between the hardware and guest VMs uses its own disk driver to manage disk activity. Adding virtualized guests divides the hardware resources among the guest VM operating system/applications instance. Even though native operating system drivers might be good, the ability for a hypervisor to manage the communication needs among a number of guests becomes a very sophisticated business, and latency and efficiency issues will be seen as application performance slow-downs.

We ran IOmeter in each VM instance to gauge how the hypervisor could "breathe" data to disk. We used a tougher-than-real-world ratio of 70 percent writes vs. 30 percent reads. We favored writes in our configuration because they aren't heavily cached by the operating system (so their contents don't evaporate during power outages or hardware resets), and read-based cache can distort measurements.