Way to make instance more NUMA aware?

Latest response


Is there a way to make an instance more NUMA aware? That is, when I create an instance that uses 8 cores, the vcpus are assigned to physical CPUS (pcpus) across different sockets, or nodes, in the system. It would be better if the vcpus were assigned to pcpus on the same socket and share the memory on that socket or node.

If I don't delete the instance, I assume I could edit the instance's xml file to assign vcpus to pcpus in a NUMA-aware manner but this doesn't seem very flexible.

So, ideally, it'd be nice to tell the nova boot command to use certain pcpus for its vcpus and have the hypervisor pin the vcpus accordingly.

Is there a way to do this?



Hi Dave,

It looks like this feature is missing from the upstream. I found a mail thread that discusses the very topic of pCPU-vCPU pinning.