VSphere number of VMs per CPU core

[vmproteau]

I've seen various estimates on VMs per CPU core and I understand it depends on a combinations of things. I think 4-5 VMs per CPU core is thrown around. We run HP Proliant DL380 G5s (2-quad core) and it looks like that 4-5 per CPU core will be our number. Memory being the main limiting factor.

I am now looking at Proliant DL380 G6 (2-quad core) servers and considering 72GB of memory. Before I size like that, I'd like to know if 8-10 VMs per core is conceivable. Although I think my current hardware might choke if I tried that, I thought perhaps newer processor families like the Proliant G6 Nehalem along with VSphere improvments with respect to CPU scheduling might make this a possibility.

Also, I have assumed processor speed variess directly with the VM per core theroetical. As CPU speed increses # or VMs per core increases. Is that accurate?

[azn2kew · Accepted solution]

You're right it depends on many things but for general rules of thumb 4-6 VMs per core for ESX 3.x and you can safely run 6-8 VMs for vSphere, if you know your VMs are light use, then pushing to 6-10VMs might do it but you may have to monitor the activities and decrease if the resources are hammering. vSphere definitely has better CPU scheduling services and gain better performance from predecessor. I would test it on a test environment and stress it to the maximum and see how it runs out. What types of servers and workloads are you thinking about?

If you found this information useful, please consider awarding points for "Correct" or "Helpful". Thanks!!!

Regards,

Stefan Nguyen

VMware vExpert 2009

iGeek Systems Inc.

VMware, Citrix, Microsoft Consultant

[vmproteau]

We support some of our clients in a shared ESX environment so, it runs the spectrum in terms or server role and resource activities. 8-10 would be a top threshold. I don't see us getting that high because we forecast using N+1 strategy. But, because of licensing costs I see us adding more and more memory to servers while keeping CPU count the same. I wanted to make sure we could push it a little more if needed. Also, no point in building Hosts with more memory if we can't take advantage of it.

I will test but, our lab is different hardware and resource levels and toughto mimic load.

I guess this leads me to a question about CPU monitoring. Since we would push up on memory limits I have not concerened myself too much with CPU, besides routine checks. When I do test pushing limits in the lab, what metrics am I looking for. CPU usage doesn't seem to be a key number. It seems I'd want to be looking at some kind or CPU wait number or something. Then I'd need to know what is the threshold to be concerned about.

Any whitepapers to help with this?

[Chuck8773]

In ESX 3.5 there is a limit of 8 vCPU's per CPU core. vSphere4 increased this to 20 vCPU per core. So you maximum upper limit is 20 VM's per core, as each VM must have at least 1 vCPU. That is a very large number for today's processors. I see that being more attainable as CPU's get faster.

Charles Killmer, VCP

If you found this or other information useful, please consider awarding points for "Correct" or "Helpful".

[azn2kew]

You can use free built in tool such as vCenter CPU performance graphs, esxtop commands as well but you can 3rd party tool to granularly monitor specific CPU statistics such as Hyper9, Veeam, Platespin etc...I would suggest reading VMware Resource Management guide to best optimize your environment. 20 VMs per core is too large in my opinion especially shared services, you want to guarantee performance to meet SLA, so using N1 is definitely a good start and if you overcommitted number of VMs per core, then N1 can take over the load and just add another host to your N+1 cluster when fits.

If you found this information useful, please consider awarding points for "Correct" or "Helpful". Thanks!!!

Regards,

Stefan Nguyen

VMware vExpert 2009

iGeek Systems Inc.

VMware, Citrix, Microsoft Consultant