9.18.11. Enabling descheduler evictions on virtual machines

The KubeVirtRelieveAndMigrate profile evicts pods from high-cost nodes to reduce overall resource expenses and enable workload migration. It also periodically rebalances workloads to help maintain similar spare capacity across nodes, which supports better handling of sudden workload spikes. Nodes can experience the following costs:

The profile enables the LowNodeUtilization strategy with the EvictionsInBackground alpha feature. The profile also exposes the following customization fields:

Example configuration:

apiVersion: operator.openshift.io/v1
kind: KubeDescheduler
metadata:
  name: cluster
  namespace: openshift-kube-descheduler-operator
spec:
  managementState: Managed
  deschedulingIntervalSeconds: 30
  mode: "Automatic"
  profiles:
    - KubeVirtRelieveAndMigrate
  profileCustomizations:
    devEnableSoftTainter: true
    devDeviationThresholds: AsymmetricLow
    devActualUtilizationProfile: PrometheusCPUCombined

The KubeVirtRelieveAndMigrate profile requires PSI metrics to be enabled on all worker nodes. You can enable this by applying the following MachineConfig custom resource (CR):

Example MachineConfig CR:

apiVersion: machineconfiguration.openshift.io/v1
kind: MachineConfig
metadata:
  labels:
    machineconfiguration.openshift.io/role: worker
  name: 99-openshift-machineconfig-worker-psi-karg
spec:
  kernelArguments:
    - psi=1

You can use this profile with the SoftTopologyAndDuplicates profile to also rebalance pods based on soft topology constraints, which can be useful in hosted control plane environments.