search

redhat-performance / openshift-psap

Example roles and yaml files for performance-sensitive applications running on OpenShift
18 stars 14 forks source link

OCP 3.11 (through kubernetes) - priorityClassName: system-node-critical does not work on statefulsets #44

Open Tanemahuta opened 5 years ago

Tanemahuta commented 5 years ago

Summary: K8s does not schedule the pods for the daemonset in case priorityClassName is set (https://github.com/kubernetes/kubernetes/issues/78383), thus the nvidia-device-plugin will not work.

Steps to reproduce: Apply the prerequisites and the nvidia-device-plugin playbook to an OCP 3.11 cluster.

Expected state: Device plugins get started on the GPU nodes (matched by the label).

Actual state: Error message Error creating: pods "nvidia-deviceplugin-daemonset-" is forbidden: pods with system-node-critical priorityClass is not permitted in nvidia-device-plugin namespace is thrown for the daemon set.