Hello,
I’m not sure, if I understand the NUMA thing correctly. I have the following situation:
- HP DL380 ESXi 6.0U2 hosts with 2 Intel pCPUs with 12 pCores each = 24 pCores per host, HT enabled
- 256 GB RAM per host
- VMs with 6vCPU, 32 GB vRAM per VM (virtual Xenapp-Servers, Win 2012 R2)
As far as I found out, my ESXi hosts have 2 NUMA nodes. Now I’m trying to figure out, how many VMs per host will perform best.
My understanding was, if I place 4 of these VMs on one host, both NUMA nodes would be used. But this doesn’t seem to be the case:
Most of the time, all VMs are placed on one NUMA node, in this case the N%L seems to be quite good, %RDY also. What I randomly see are short switches of a VM to NHN 0, with decreasing N%L, but a minute later, it’s moved back to NHN 1.
So I tried with 5 of these VMs. Initially after vMotion, both NUMA nodes were used, but minutes later, all VMs had been migrated to NHN 0(!!), with lower N%L values and %RDY values getting worse:
From my understanding, I expected both NUMA nodes to be used, not necessarily 50%-50%, but definitely better than 100%-0%.
Can anyone guess why there is this imbalance?
