Comment thread started by heinrich321 on Virtual Machine Compute Optimizer

Full comments
May 07, 2021

We used the VMCO fling to optimize our VMs and made many changes over the past month. A lot of our VMs are 4+ cores, many 8 and 16 cores, and it suggested we move all our VMs to 1 socket and (4, 8, and 16) core setups.

However, our CPU ready time was astronomical... Like 80,000+ ms. Our CPU utilization on each host was only ~25%. We are running single-socket EPYCs host, each with 64 cores/128 threads.

We have since moved almost all of our VMs back to (1) core-per-socket and 4, 8, and 16 sockets. Now our CPU ready time is ~100ms or less for most of our VMs, 300-400ms for some of the heavy hitters.

Since moving back to (1) core-per-socket the VMs feel snappier but we made the changes yesterday and haven't had time to do proper testing. Any suggestions?

May 19, 2021

Heinrich, I apologize for the slow reply! Can you confirm what version of vSphere you are running in your environment? There are some versions that had performance issues w/ the EPYC processors, so a good place to start troubleshooting. See this KB: https://kb.vmware.com/s/article/67996.

Thanks!

May 19, 2021

Mark thanks for the reply. We are using vSphere 7.0 U2 (7.0.2.00100) and have AMD EPYC 7742 (Rome) processors.

We looked through this guide (Page 27) https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/techpaper/performance/vsphere70u2-cpu-sched-amd-epyc.pdf

We followed these recommendations:

- NUMA nodes per socket = 1
- CCX-as-NUMA = Disabled

These past few weeks the systems seem to be running much better, CPU Ready time is down considerably. Unfortunately I don't have a good benchmark test (before/after) to show...