Performance tab reporting inaccurate info

[k_shady]

Hi,

I have created 3 ESXi 5 hosts in HA and DRS cluster in vcenter 5. I have guest VM's in these hosts which are used as an Load test servers. During Load test there is discrepencies in the CPU usage which is reported in the performance tab of the host to the actual used by the server.The discrepencies is something around 20 - 25 % than the actual value reported in the server.

Any help is much appreciated.

Thanks

[vLarus]

The info you see on the hosts are the accurate numbers.

The OS performance statistics are incorrect.

The CPU scheduler will give the VM the CPU cycles it has available at any time, and also with several memory saving techniques (compression, shared memory pages, swapping, ballooning) you will not have matching values in both metrics.

If you are creating load on a ESXi host, just overload it with machines and watch host performance metrics.

[k_shady]

Hi Larus,

Thanks for your response.

Will the below explanation fit for discrepencies seen on the performance tab of individual load sever VM guest server and performance seen on the OS level.

For example: if i m using server1 running linux as my load test server and i m running a load test on it. The CPU usage caught by vmstat and the one reported in the performance tab of that particular server has something like 20 -25 % discrepencies during the test.

Thanks in advance

[vLarus]

That explanation is valid for the discrepencies.

The discrepencies will vary when the vKernel (that runs the CPU scheduler) is under load, or there is contentation over CPU cycles (to many vCPUs on one core=CPU scheduler overload, to much load etc.)

[k_shady]

Hi,

Can you please explain this with an example plz as I am still unclear with your explanation.

Thanks

[vLarus]

This excellent post from http://vpivot.com/2010/02/10/inaccuracy-of-in-guest-performance-counters/ should explain this better.

The main point from that article is:

"The reason the host shows lower utilization than the guest is because the guest is unaware that it is only getting a fraction of the host’s CPU, time-sliced by ESX’s scheduler."

Hope this clarifies the issue somewhat.