Stefan Hajnoczi suggests (from a "best practises for VirtIO storage
configuration" PoV):

Higher and more consistent performance is achievable with NUMA pinning:

1. Define guest NUMA nodes for subset of host NUMA nodes that guest uses

2. Pin vCPUs and memory to host NUMA nodes in 1:1 mapping 

3. Create IOThreads for each guest NUMA node and pin 1:1 to host NUMA nodes 

4. Add PCI Expander Bridge (PXB) device for each guest NUMA node 

5. Assign virtio-blk and virtio-scsi devices to IOThreads so that the
   vCPUs and IOThread are pinned to their host NUMA node