If an instance with 'hw:cpu_thread_policy=isolate' is scheduled on a non-HT host, the pinned pCPUs are not properly accounted for. This can lead to multiple instances running on the same pCPUs
The problem is that in LibvirtDriver._get_host_numa_topology() when calculating the NUMACell.siblings field we filter out single cpus. On a non-HT host this means that NUMACell.siblings is an empty list.
Later when _update_usage() runs it ends up eventually running NUMACell.pin_cpus_with_siblings(). This contains the following code:
def pin_cpus_with_siblings(self, cpus):
pin_siblings = set()
for sib in self.siblings:
if cpus & sib:
pin_siblings.update(sib)
self.pin_cpus(pin_siblings)
Since "self.siblings" is empty, we end up calling self.pin_cpus() with an empty list, which means that we don't update self.pinned_cpus.
Stephen Finucane has suggested the correct fix might be to leave single pCPUs in the NUMACell.siblings field. This needs to be verified to make sure that it doesn't cause other problems.
Fix proposed to branch: master
Review: https:/