Have two identical VMs (one cloned from the other). One VM has numa affinity for numa0 and the other VM for numa1.
Built a VDS within vCenter and pulled it into a Host Node in NSX-T.
Send traffic to the numa0 VM and we get a benchline of performance.
Send the same traffic to the numa1 VM and we get 10-20% less throughput.
Have checked all of the esxtop statistics and there's no indication of any vCPUs nor RAM being used "cross-numa".
Suspect that the vmnics are not aligned properly somehow.
[root] vsish -e get /net/pNics/vmnic4/properties | grep 'Device NUMA Node'
Device NUMA Node:1
[root] vsish -e get /net/pNics/vmnic8/properties | grep 'Device NUMA Node'
Device NUMA Node:0
screenshot from vCenter:
from esxcfg-vswitch_-l.txt:
This is from NSX-T:
Looking at this (this is just a guide, not architecture from our system) :
1) How can I check that the QPI bus (or whatever it's called for the Xeon 8358) was traversed?
2) How can I check that everything is vertically assigned?
TIA