Mixed Workloads

[RSEngineer]

IHAC with mixed workloads with a great disparity in their desktop composition. Have some 3D CAD graphics virtual desktops and then some lower-level knowledge and task workers who do not require the same kind of CPU speed. For CAD desktops, we are talking about 6 vCPUs per desktop. And with CAD, high speed cores are needed, like 3.2Ghz, at least. High speed cores dont usually come in high quantities, so CPU is a limiting factor when it comes to consolidation ratios.

That said, there will be the smaller task and knowledge worker desktops sharing the same V-Series VxRail platforms with the CAD desktops. Each type will use 2 vCPUs per desktop. That will allow me to increase the consolidation ratio more than it would be if all the desktops were the 3D CAD ones. BUT it seems like a colossal waste to have them using the 3.2 Ghz cores. Besides building a totally separate cluster for the task and knowledge worker desktops with slower CPU cores (and more of them), can they be deployed on the same cluster with the CAD desktops BUT in a way so as to use only a portion of the 3.2 Ghz cores and perhaps allow me to double up their number (or something like that)  and further increase consolidation ratios?

Am I making sense or talking kaka?

[techguy129]

When a VM needs CPU cycles, it needs all available cores, based on its vCPU, available all at once on the host to schedule. It's not necessarily based on CPU speed. When mixing VM's with different vCPU counts on the same host, your biggest concern is the co-scheduling. You don't want contention between the VMs waiting for available CPUs. Since this is VDI, you can consolidate at a higher rate then typical server workloads. Something like 10vCPU to 1 physical core. You have to watch your CPU ready % and Co-Stop % to see if CPU is really an overallocated.

With that all said...With your scenario, if you have a few CAD machines at 6vCPU and a lot of 2vCPU, you have potential to have the above conflict. As far as a waste, if it has higher Ghz then the smaller workloads will be completed faster.

[RSEngineer]

Hi and thanks. I cant get 10:1 overcommit because the CAD app requires high CPU speed, and because of that, the core counts are low. All I can get on VxRail V-Series are 2s/24 cores @ 3.2Ghz. With 2:1 overcommit, all I can get are 8 CAD desktops per appliance.Or a combination of a fewer (than 😎 CAD desktops and a small handful of others.

I wouldnt do more tan 2:1 precisely for the reasons you mentioned above. Imagine having, say, 4:1 overcommit while still having to schedule 6 vCPUs for every workload to execute...

So would it make more sense to build a separate physical cluster just for the CAD VDIs given the limitation on overall core count they impose? If I separate the normal workloads in their own cluster and use CPUs at slower speeds, then I can leverage like 6:1 overcommit (I wouldnt go higher than that to start off with), and get CPU with much higher core counts to boot, and get like 50 or more VMs on the same appliance.