This is a placeholder epic for numa distance based VM placement in nova.

This is a complex topic which we have discussed several times in the past.

the simplest explanation is that numa distance aware shduling will compute a metric for the CPU to numa node to device affinity and minimise that metric when selecting devices to use.

one incarnation of this would be a shduler weighed that operates on allocation candidates and host numa topligy info to try and minimise the distance between PCI devices and cores.

this would complement the numa topology filter and if we had numa in placement could be done entirely based on the allocation candates.

on a compute node level that would also involve ensuring the cpus selected are the ones that most optimally minimise the distance between cpus and passhtough devices.

libvirt provides detailed CPU distance info but a simple algorithm can be done by using a proxy metric.

cores on the die/amd ccx can have a distance or 1 on the same numa node but different ccx have a distance fo 10
on the same CPU socket but a different numa node a distance to 100. and different sokcet a distance 1000

when selection cpus the distance between each CPU and the distance form the cpus and the devices should be minimised.

the exact alghrotim is not important to cautre in this epic but that is the high level intent.