NUMA misses performance implications

[Menno]

Hello,

I'm experiencing many NUMA misses on a few of our Proxmox servers and reading the documentation[0] this might be resolved by enabling NUMA emulation for the VMs so the resources are properly distributed.

Before changing all the VM configs I'd like to know if this is safe to do on already existing VMs, i.e., won't this cause stability issues because of the change of virtual hardware.

I assume the Linux VMs won't cause much trouble but I'm more worried about the Windows VMs because I have less experience with them.

Is there anyone who has experience with NUMA misses and enabled the option on already installed VMs to resolve this?

If so did performance actually improve? Reading RedHat's virtualization guide[1] they mention NUMA misses performance impact are quite large, "generally starting at a 10% performance hit or higher."

[0] https://pve.proxmox.com/wiki/Qemu/KVM_Virtual_Machines
[1] https://access.redhat.com/documenta...ning_optimization_guide-numa-numa_and_libvirt

 

[Stefan_R]

Yes, Windows VMs do support NUMA. Enabling NUMA should not break anything, but as with any config change, testing is essential, also to see if it actually improves performance for you.

 

[Menno]

Thanks Stefan,

just to update in case anyone else finds this useful, we've tested this on a Proxmox machine that had quite a few NUMA misses according to numastat but after enabling numa emulation in Proxmox for most of the VMs on it we did not see these misses go down.

For now we've decided not to use this feature on our production clusters as it doesn't seem to benefit us, though I'd be interested if anyone else has the opposite experience.