rrosenkoetter
Enthusiast
Enthusiast

So how different is hosting desktops from servers?

Jump to solution

I've got a good-sized ESX farm for servers (16 hosts, 300 VMs), and now we're looking at hosting about 100 desktops...

So I get 20:1 ratio for servers (hosts are 4-proc, 32 GB RAM)

What kind of VM to host ratio can I expect for desktops? I'm thinking it may actually be LESS than my server ratio, since people will be RDPing to these VMs and actually interacting directly with them all day long.

These are programmers using these desktops by the way.

I know "it depends"... but a rough estimate would be helpful. What kind of ratio are you guys seeing? What are the bottlenecks? CPU, memory? What the main differences between hosting servers and hosting desktops?

Thanks for any help

0 Kudos
1 Solution

Accepted Solutions
conyards
Expert
Expert

with what would appear (on the face of it) a very similar setup, I have been able to achieve a ratio of 30:1 (3.75 VMs per logical core, as the servers in question are 4 Dual core CPU boxes)

During periods of maintenence performance hasn't taken a big hit until ratios of 50:1 are encountered. (6.25 VMs per logical core).

Simon

https://virtual-simon.co.uk/

View solution in original post

0 Kudos
8 Replies
nick_couchman
Immortal
Immortal

Just a sort of a side note here - I'd measure the ratio in VMs per core, not VMs per host. It makes a little more sense, in my mind, anyway, since you can have hosts with different numbers of CPUs, cores, etc. If your hosts are all the same, that might not be an issue for you, but I have four ESX servers of three different types. Same basic CPU, but one of them has single core CPUs, one has dual core CPUs, and two of them are CPUs with the Intel VT technology. The speeds are different on the first two, as well, and this all plays into how many VMs I can get per host.

Just my thought...

conyards
Expert
Expert

with what would appear (on the face of it) a very similar setup, I have been able to achieve a ratio of 30:1 (3.75 VMs per logical core, as the servers in question are 4 Dual core CPU boxes)

During periods of maintenence performance hasn't taken a big hit until ratios of 50:1 are encountered. (6.25 VMs per logical core).

Simon

https://virtual-simon.co.uk/

View solution in original post

0 Kudos
davlloyd
Hot Shot
Hot Shot

We have successfuly tested 80 guests running on a single host of the same specification. This was not in a lab but a real world test with live users on about three quarters of the guests and stress and performance agents running on the others (just to test the capacity). Interestinglhy the host was happy and not one negative response from a user.

In production things are more conservative but still hosting 10 guests per core. Bis thing you need to think about with 100 guests is what i the contingency capability if you lose a host. If you split it over two hosts and one is lost (or if you just nee to run some maintenance) then you will need to be able to swing the 100 guests to the other host, better to go a bit wider and shallower with smaller hosts to ensure performance and contingency.

heybuzzz
Enthusiast
Enthusiast

davlloyd,

How much RAM do you guys intially give to your users per VM?

0 Kudos
tbrouwer
Hot Shot
Hot Shot

VMware suggests I think 6-8 VMs per core. Obviously depends on what they are doing. For me, I have 20 XP VM's running right now, and I guess the bottleneck will end up being the RAM.

0 Kudos
heybuzzz
Enthusiast
Enthusiast

I have some home transcribers who usually have 5 to 6 applications open at once multi tasking, printing, copy pasting, and found that 1 gig of ram per VM may not be enough. They usually all work at the same time too. Has anyone experienced a similar situation?

0 Kudos
tbrouwer
Hot Shot
Hot Shot

That is a good example of how it completely depends on the environment. I have 512 for all my XP VM's, and that is more than enough.

0 Kudos
davlloyd
Hot Shot
Hot Shot

davlloyd,

How much RAM do you guys intially give to your users

per VM?

We run between 512 and 1024 with no issues with 10 machines per core. We increased the memory allocation just to make the hosts work harder at it more then due to user requirements.

0 Kudos