With Kubernetes topologyManagerPolicy set to restricted we can make sure that pods using SR-IOV CNI plugin will be able to start only in the same NUMA as SR-IOV card. If all CPU in this NUMA are reserved by CPU manager pod will not start.
Does anyone know if pods will have PriorityClass with higher priority and PreemptLowerPriority
it will try to evict only pods in the same NUMA with lower priority or will evict random pods from any NUMA that will not help to schedule new pod?
I have access only to cluster with topologyManagerPolicy best-effort and the pods are eventually scheduld in different NUMA so evistion does not start.
Cluster information:
Kubernetes version: v1.19.5
Cloud being used: bare-metal
Installation method:
Host OS: Centos 7