While testing and evaluating the new beta of Hyper-V that is included with Windows Server 2008 I have run across some interesting performance monitor stuff.

First a little background - I have been testing and evaluating the performance of VMs on VMware's ESX server since early 2004 with the first whitepaper I'm aware of that took an detailed look at the performance of VMotion. So I have been using esxtop ever since to monitor the performance of ESX servers.

Windows Server 2008 RC1 has a new Perfmon that is much improved over previous versions, so I was really looking forward to using it for my Hyper-V testing. The first testing that I did with W2K8 RC0 showed that Hyper-V (called Windows Server Virtualization at the time) was able to evenly distribute limited resources across 8 VMs all under load. When using perfmon to monitor the CPU during these tests - it at first looked like the server wasn't loaded at all. I was looking at the standard %Processor Time for the entire server. I knew that the load I was running should be pushing the server well over %50 utilization, so I started digging into this new perfmon to see what was going on.

I found a set of counter objects that were labeled as hypervisor. Here there were Hypervisor Virtual Processor, Hypervisor Root Virtual Processor, Hypervisor Root Partition, Hypervisor Partition, and Hypervisor Logical Processor. It seemed that surely one of these would give me the overall CPU utilization that I was looking for, but it wasn't really clear which one was the counter to look at. As this is still beta code, there was no documentation for these counters to clearly state what each one meant. I decided in the end that the CPU utilization wasn't really that important to report as part of my results, and I just decided to look at it again when the next beta was released.

It seemed like a big issue to me because it would be possible for admins to look at %Processor Time and see a low utilization number while the system was actually very heavily loaded with several very active virtual machines.

With the Hyper-V beta that is included with W2K8 RC1 (released in Dec 07) I am currently doing a new round of testing to see if there is any change in performance. I'm doing the same test with the DVD Store database, but this time the perfmon results are very different. The %Processor Time for the server now seems to reflect the activity of the virtual machines (or partitions as they are referred to in perfmon).

I'm still working on finalizing the results of the RC1 testing - I'll have an entry here when they are ready.

Todd