Hi All
Over the past 6month I've looked alot into the performance of single thread processes.
My finding are that single thread apps can suffer greatly in a virtual machine.
But before you say "Well then multi thread your apps", I would like to point out that despite my tests focusing on single thread apps, I think that my findings could assist in overall performance in many areas. Using a single thread app is just an easy way to expose the issue.
Here's the problem:
Test1: Windows guest, with 2vcpus (2500mhz). Run a cpu intensive single core process, i.e 7zip
Run a 200meg zip job.
Findings:
As evident by task manager, the process is thrown between both vcpus. The vSPhere performance chart shows only ~1600mhz being used at most.
zip job takes significantly longer than a 2500mhz physical machine.
Test 2: Same setup as test 1. BUT, set the affinity of the 7zip process to just one vcpu, using windows task manager.
FIndings:
Kaboom! vSphere performance chart shows FULL cpu usage: 2500 mhz.
And the job finishes in expected time, similar to that of a physical machine.
Can any one else reporduce these result?
Am I the only one who sees this as a concern??
My point is this: If someone of sufficient expertise investigates this, I beleive that we may find that the CPU scheduling of guest OS's could be greatly improved for virtual environments.
For example, if i am using my laptop, then yes, I want the OS to be nice to each core. But on a VM, where resources and power are already shared, then it's just a waste.
Cheers
Michael
Fyi, if you want more detail, I posted some time ago about my findings: http://communities.vmware.com/message/1683965#1683965
