Accommodating High vCPU VM

[sgunelius]

I've got a request to support a VM with 42 vCPU on a 2-node dedicated ESXi 7U3 cluster.  Each node is a ProLiant BL460c Gen9 with (2) Intel Xeon E5-2697v4 (2.3GHz/18-core), so the host has 36 physical CPUs available.  I confirmed Hyperthreading is enabled, so the vSphere client indicates there are 72 logical CPUs available on each node.

The VM will be supporting a large number of LiDAR sensors and the ISV requirements indicates 1 vCPU per sensor, so that's driving the configuration.  The cluster is supporting another 6 VMs that range in size from 4-12 vCPU and I don't see any evidence of high %Ready values on the cluster.

We did run a "What-If Analysis" from Aria Operations (formerly vRealize Operations) and the analysis indicates we could support this VM operating at 80% utilization. Is there a direct correlation between vCPU and Logical CPU when doing capacity planning of this nature?  Is there a best practice to avoid oversubscribing processor resources.

Thank you.

[Sachchidanand · Accepted solution]

Please go through the following documents:

https://docs.vmware.com/en/VMware-vSphere/8.0/vsphere-vm-administration/GUID-3CDA4DEF-3DE0-4A64-89C7...

https://blogs.vmware.com/performance/2017/03/virtual-machine-vcpu-and-vnuma-rightsizing-rules-of-thu...

As defined in the document:

A virtual machine cannot have more virtual CPUs than the actual number of logical CPUs on the host. The number of logical CPUs means the number of physical processor cores or two times that number if hyperthreading is enabled.

Regards,

Sachchidanand

[EvertAM]

Keep in mind that you'll be giving up your redundancy capacity as well. If either one of those hosts fails, there's no way you're keeping all VM's running on the other one.

[sgunelius]

Yes, I'd been concerned about that as well.  It's no good having native HA or vMotion if there isn't enough capacity to support the entire VM workloads should we lose a node.

Thank you.

[sgunelius]

Thanks for reminding me about the direct correlation between vCPU and lCPU; I'd obviously forgotten about that.  So, with hyperthreading enabled, I've got 72 lCPUs and can support that number of vCPUs between the existing VMs (42 vCPU total) and the new 42 vCPU VM they want to provision in the cluster.

You can't argue the math, it just won't fit, especially if we want enough capacity that a single host in the cluster can support the entire workload if we need to perform maintenance or a host experiences a PSOD and HA restarts all VMs on the remaining host.

We could replace the processors in these BL460c Gen9s with Intel Xeon E5-2699v4 (2.2GHz/22-core), to give us 88 lCPUs, but these hosts are already 7 years old and nearing End of Support Life.  We've been replacing these hosts with Synergy 480 Gen10 equipped with Intel Xeon Gold 6248R (3.0GHz/24-core) processors, giving us 96 lCPUs, so that'll probably be a better long-term investment.