A toleration is Pod configuration that allows scheduling onto a node with a matching taint. Taints repel Pods that do not tolerate them; unlike node labels, they do not attract Pods to a node. A toleration permits placement but does not, by itself, direct a Pod there. section 3.4.1
The taint’s effect determines what happens without a matching toleration:
- NoSchedule prevents the Pod from being scheduled on the node. section 3.2.7
- PreferNoSchedule tells the scheduler to try to avoid the node. section 3.2.7
- NoExecute can evict a running Pod when the taint is added. section 3.2.7
Reserve a node for GPU workloads
Apply a taint with `kubectl taint node NODE workload=gpu:NoSchedule`, then add the `workload=gpu` label with `kubectl label node NODE workload=gpu`. Configure the GPU Pod with both a selector for that label and an Equal toleration matching the taint’s key, value, and effect. The selector directs the Pod toward the labeled node; the toleration lets it pass the taint. Verify placement by checking the Pod’s reported node. section 3.4.1
apiVersion: v1 kind: Pod metadata: labels: run: nginx-gpu name: nginx-gpu spec: containers: - image: nginx name: nginx-gpu restartPolicy: Always nodeSelector: workload: gpu tolerations: - key: "workload" operator: "Equal" value: "gpu" effect: "NoSchedule"
The Pod manifest places `nodeSelector` and `tolerations` together to target and tolerate the GPU node. section 3.4.1
