We help IT Professionals succeed at work.

VMware ESXI massive performance loss??

We recently virtualised a few old physical servers and all seemed well, until the customer complained that the VM seems slower than the old servers. Which to us was hard to believe as on paper as the new server should outperform it with ease.

We went and grabbed the old server and booted them up and compared some application load times and the VM was definately slower. We then benchmarked both machines using 'Novabench' and the old physical server posted much higher numbers.

The old server (Dell T3500) is running a single X5670 and the new server (Poweredge R720) is running 2x e5-2640. Even with no other VM's running on the machine and 8 cores assigned to the VM. The old server scored 958 for CPU and the VM scored 627.

But the biggest hit seems to be RAM speed, the old server scored 12116 mb/s and the VM only got 6509 mb/s. Both have DDR3 ram but the new server has faster RAM, yet benchmarking at half the speed.

We have tried a tonne of recommended performance tweaks, like turning off power management, disable hyperthreading, playing around with different amounts of cores and sockets, bios settings etc. Have seen some improvment but the old physical machine is still well ahead.

Is this normal for VM's? Am I wasting my time benchmarking a VM with conventional benchmark software?
Comment
Watch Question

Senior IT Advisor
Top Expert 2008
Commented:
This is common, especially if you leave the amount of vCPUs equal to the amount you have on physical servers.

Keep in mind that VMware is a scheduler for your VMs using the same hardware, if your CPU count is large it will may experience issues with queing, think of a restaurant with 2 seat tables, if you have an 8 person party come in it will have to wait until several tables free up to allow for the 8 person party to have a seat.

good article..check your CPU % ready, wait costop

http://searchservervirtualization.techtarget.com/answer/Reducing-CPU-ready-time-will-get-your-hungry-VMs-a-seat-at-the-table

https://vmware2112.wordpress.com/2012/07/05/what-does-the-cpu-co-stop-counter-measure-cstp/

Author

Commented:
Well like I said, I have ran numerous tests with just that single VM running so I dont believe vCPU queing is an issue, I have tried using only 4 cores and even just 1 core and as expected it just performs worse.
Andrew Hancock (VMware vExpert / EE Fellow)VMware and Virtualization Consultant
Fellow 2018
Expert of the Year 2017
Commented:
Is this normal for VM's? Am I wasting my time benchmarking a VM with conventional benchmark software?

All I can say is too be expected!

and yes you are comparing apples and pears!

The question you should be asking is, does it affect service, does it meet user expectations ?

Then we have to look at the old server configuration versus new VM, and what is the application and service being delivered?

Also what is the Host specification, Virtualisation is a Compromise between performance, and hosting many servers on a single box.
What version of ESXi is the host running? What OS and application is the VM running? How many Vms are on the host? What type of storage is the VM using? is NUMA enabled or disabled?