Since the addition of CPU hotplug, vcpu placement is now defaulting to "static" placement, ex:
  <vcpu placement='static' current='4'>16</vcpu>

this allows vcpus to be scheduled across the entire host, which in cases of smaller VMs can mean higher cross numa penalties for workloads.

Libvirt offers an "auto" vcpu placement option, ex:
  <vcpu placement='auto' current='4'>16</vcpu>

this option makes a "one shot numad" call when the guest is started in order to apply a cpu affinity and memory binding which can significantly help performance when the VM can fit inside 1 NUMA node (or a subset of total host NUMA nodes). 
However note, the host node needs numad installed to properly take advantage of this (not currently in default RHCOS).

All supported