Comment 11 for bug 1519878

Hey Serguei - so the patch I linked is only a test. It was the quickest way for me to confirm that the NUMA fitting logic itself is not flawed with the exact data it would use in your case. It does not seem to be on my fresh checkout of 12.0.0

If you look at the test added in the patch - it does not make the exact same call to numa_fit_instance_to_host that the filter does - filter considers overcommit ratios and requested pci devices in the general case, but that should be fine since overcommit is not considered for CPU pining anyway.

The only thing left to confirm would be if you are requesting any PCI devices with your instance. You don't seem to be doing it from the flavor, but it is possible that you are requesting a Neutron port that will result in Nova needing to request a PCI device (--vnic-type=direct). The reason this would fail is that if you request for both CPU pinning AND a pci device, Nova will refuse to pin an instance to CPUs that are on a different NUMA node than the available PCI device is.

Could you confirm if you are in fact requesting a PCI nic? The quickest way would be to paste contents of the pci_requests column of the instance_extra table (so SELECT pci_requests FROM instance_extra WHERE deleted=0;)