NVIDIA Network Operator v24.4.0
Network Operator Application Notes v24.4.0

Life Cycle Management

Once the Network Operator is deployed, and a NicClusterPolicy resource is created, the operator will reconcile the state of the cluster until it reaches the desired state, as defined in the resource.

Alignment of the cluster to the defined policy can be verified in the custom resource status.

a “Ready” state indicates that the required components were deployed, and that the policy is applied on the cluster.

Status Field Example of a NICClusterPolicy Instance

Get the NicClusterPolicy status:

Copy
Copied!
            

kubectl get -n nvidia-network-operator nicclusterpolicies.mellanox.com nic-cluster-policy -o yaml

Copy
Copied!
            

status: appliedStates: - name: state-pod-security-policy state: ignore - name: state-multus-cni state: ready - name: state-container-networking-plugins state: ignore - name: state-ipoib-cni state: ignore - name: state-whereabouts-cni state: ready - name: state-OFED state: ready - name: state-SRIOV-device-plugin state: ignore - name: state-RDMA-device-plugin state: ready - name: state-NV-Peer state: ignore - name: state-ib-kubernetes state: ignore - name: state-nv-ipam-cni state: ready state: ready

Note

An “Ignore” state indicates that the sub-state was not defined in the custom resource, and thus, it is ignored.

Before upgrading to Network Operator v1.0 or newer with SR-IOV Network Operator enabled, the following manual actions are required:

Copy
Copied!
            

$ kubectl -n nvidia-network-operator scale deployment network-operator-sriov-network-operator --replicas 0 $ kubectl -n nvidia-network-operator delete sriovnetworknodepolicies.sriovnetwork.openshift.io default

The network operator provides limited upgrade capabilities, which require additional manual actions if a containerized OFED driver is used. Future releases of the network operator will provide an automatic upgrade flow for the containerized driver.

Since Helm does not support auto-upgrade of existing CRDs, the user must follow a two-step process to upgrade the network-operator release:

  • Upgrade the CRD to the latest version

  • Apply Helm chart update

Downloading a New Helm Chart

To obtain new releases, run:

# Download Helm chart $ helm fetch https://helm.ngc.nvidia.com/nvidia/charts/network-operator-24.4.0.tgz $ ls network-operator-*.tgz | xargs -n 1 tar xf

Upgrading CRDs for a Specific Release

It is possible to retrieve updated CRDs from the Helm chart or from the release branch on GitHub. The example below shows how to upgrade CRDs from the downloaded chart.

Copy
Copied!
            

$ kubectl apply \ -f network-operator/crds \ -f network-operator/charts/sriov-network-operator/crds

Preparing the Helm Values for the New Release

Edit the values-<VERSION>.yaml file as required for your cluster. The network operator has some limitations as to which updates in the NicClusterPolicy it can handle automatically. If the configuration for the new release is different from the current configuration in the deployed release, some additional manual actions may be required.

Known limitations:

  • If component configuration was removed from the NicClusterPolicy, manual clean up of the component’s resources (DaemonSets, ConfigMaps, etc.) may be required.

  • If the configuration for devicePlugin changed without image upgrade, manual restart of the devicePlugin may be required.

These limitations will be addressed in future releases.

Warning

Changes that were made directly in the NicClusterPolicy CR (e.g. with kubectl edit) will be overwritten by the Helm upgrade due to the force flag.

Applying the Helm Chart Update

To apply the Helm chart update, run:

Copy
Copied!
            

$ helm upgrade -n nvidia-network-operator network-operator nvidia/network-operator --version=<VERSION> -f values-<VERSION>.yaml --force

Warning

The –devel option is required if you wish to use the Beta release.

OFED Driver Manual Upgrade

Restarting Pods with a Containerized OFED Driver

Warning

This operation is required only if containerized OFED is in use.

When a containerized OFED driver is reloaded on the node, all pods that use a secondary network based on NVIDIA NICs will lose network interface in their containers. To prevent outage, remove all pods that use a secondary network from the node before you reload the driver pod on it.

The Helm upgrade command will only upgrade the DaemonSet spec of the OFED driver to point to the new driver version. The OFED driver’s DaemonSet will not automatically restart pods with the driver on the nodes, as it uses “OnDelete” updateStrategy. The old OFED version will still run on the node until you explicitly remove the driver pod or reboot the node:

Copy
Copied!
            

$ kubectl delete pod -l app=mofed-<OS_NAME> -n nvidia-network-operator

It is possible to remove all pods with secondary networks from all cluster nodes, and then restart the OFED pods on all nodes at once.

The alternative option is to perform an upgrade in a rolling manner to reduce the impact of the driver upgrade on the cluster. The driver pod restart can be done on each node individually. In this case, pods with secondary networks should be removed from the single node only. There is no need to stop pods on all nodes.

For each node, follow these steps to reload the driver on the node:

  1. Remove pods with a secondary network from the node.

  2. Restart the OFED driver pod.

  3. Return the pods with a secondary network to the node.

When the OFED driver is ready, proceed with the same steps for other nodes.

Removing Pods with a Secondary Network from the Node

To remove pods with a secondary network from the node with node drain, run the following command:

Copy
Copied!
            

$ kubectl drain <NODE_NAME> --pod-selector=<SELECTOR_FOR_PODS>

Warning

Replace <NODE_NAME> with -l “network.nvidia.com/operator.mofed.wait=false” if you wish to drain all nodes at once.

Restarting the OFED Driver Pod

Find the OFED driver pod name for the node:

Copy
Copied!
            

$ kubectl get pod -l app=mofed-<OS_NAME> -o wide -A

Example for Ubuntu 20.04:

Copy
Copied!
            

kubectl get pod -l app=mofed-ubuntu20.04 -o wide -A

Deleting the OFED Driver Pod from the Node

To delete the OFED driver pod from the node, run:

Copy
Copied!
            

$ kubectl delete pod -n <DRIVER_NAMESPACE> <OFED_POD_NAME>

Warning

Replace <OFED_POD_NAME> with -l app=mofed-ubuntu20.04 if you wish to remove OFED pods on all nodes at once.

A new version of the OFED pod will automatically start.

Returning Pods with a Secondary Network to the Node

After the OFED pod is ready on the node, you can make the node schedulable again.

The command below will uncordon (remove node.kubernetes.io/unschedulable:NoSchedule taint) the node, and return the pods to it:

Copy
Copied!
            

$ kubectl uncordon -l "network.nvidia.com/operator.mofed.wait=false"

Automatic OFED Driver Upgrade

To enable automatic OFED upgrade, define the UpgradePolicy section for the ofedDriver in the NicClusterPolicy spec, and change the OFED version.

nicclusterpolicy.yaml:

apiVersion: mellanox.com/v1alpha1 kind: NicClusterPolicy metadata: name: nic-cluster-policy namespace: nvidia-network-operator spec: ofedDriver: image: doca-driver repository: nvcr.io/nvidia/mellanox version: 24.04-0.6.6.0-0 upgradePolicy: # autoUpgrade is a global switch for automatic upgrade feature # if set to false all other options are ignored autoUpgrade: true # maxParallelUpgrades indicates how many nodes can be upgraded in parallel # 0 means no limit, all nodes will be upgraded in parallel maxParallelUpgrades: 0 # cordon and drain (if enabled) a node before loading the driver on it safeLoad: false # describes the configuration for waiting on job completions waitForCompletion: # specifies a label selector for the pods to wait for completion podSelector: "app=myapp" # specify the length of time in seconds to wait before giving up for workload to finish, zero means infinite # if not specified, the default is 300 seconds timeoutSeconds: 300 # describes configuration for node drain during automatic upgrade drain: # allow node draining during upgrade enable: true # allow force draining force: false # specify a label selector to filter pods on the node that need to be drained podSelector: "" # specify the length of time in seconds to wait before giving up drain, zero means infinite # if not specified, the default is 300 seconds timeoutSeconds: 300 # specify if should continue even if there are pods using emptyDir deleteEmptyDir: false

Apply NicClusterPolicy CRD:

Copy
Copied!
            

$ kubectl apply -f nicclusterpolicy.yaml

Warning

To be able to drain nodes, make sure to fill the PodDisruptionBudget field for all the pods that use it. On some clusters (e.g. Openshift), many pods use PodDisruptionBudget, which makes draining multiple nodes at once impossible. Since evicting several pods that are controlled by the same deployment or replica set, violates their PodDisruptionBudget, those pods are not evicted and in drain failure.

To perform a driver upgrade, the network-operator must evict pods that are using network resources. Therefore, in order to ensure that the network-operator is evicting only the required pods, the upgradePolicy.drain.podSelector field must be configured.

Node Upgrade States

The status upgrade of each node is reflected in its nvidia.com/ofed-driver-upgrade-state label . This label can have the following values:

Name

Description

Unknown (empty) The node has this state when the upgrade flow is disabled or the node has not been processed yet.
upgrade-done Set when OFED POD is up-to-date and running on the node, the node is schedulable.
upgrade-required Set when OFED POD on the node is not up-to-date and requires upgrade. No actions are performed at this stage.
cordon-required Set when the node needs to be made unschedulable in preparation for driver upgrade.
wait-for-jobs-required Set on the node when waiting is required for jobs to complete until the given timeout.
drain-required Set when the node is scheduled for drain. After the drain, the state is changed either to pod-restart-required or upgrade-failed.
pod-restart-required Set when the OFED POD on the node is scheduled for restart. After the restart, the state is changed to uncordon-required.
uncordon-required Set when OFED POD on the node is up-to-date and has “Ready” status. After uncordone, the state is changed to upgrade-done
upgrade-failed Set when the upgrade on the node has failed. Manual interaction is required at this stage. See Troubleshooting section for more details.
Warning

Depending on your cluster workloads and pod Disruption Budget, set the following values for auto upgrade:

apiVersion: mellanox.com/v1alpha1 kind: NicClusterPolicy metadata: name: nic-cluster-policy namespace: nvidia-network-operator spec: ofedDriver: image: doca-driver repository: nvcr.io/nvidia/mellanox version: 24.04-0.6.6.0-0 upgradePolicy: autoUpgrade: true maxParallelUpgrades: 1 drain: enable: true force: false deleteEmptyDir: true podSelector: ""

Safe Driver Loading

Warning

The state of this feature can be controlled with the ofedDriver.upgradePolicy.safeLoad option.

Upon node startup, the OFED container takes some time to compile and load the driver. During that time, workloads might get scheduled on that node. When OFED is loaded, all existing PODs that use NVIDIA NICs will lose their network interfaces. Some such PODs might silently fail or hang. To avoid this situation, before the OFED container is loaded, the node should get cordoned and drained to ensure all workloads are rescheduled. The node should be un-cordoned when the driver is ready on it.

The safe driver loading feature is implemented as a part of the upgrade flow, meaning safe driver loading is a special scenario of the upgrade procedure, where we upgrade from the inbox driver to the containerized OFED.

When this feature is enabled, the initial OFED driver rollout on the large cluster can take a while. To speed up the rollout, the initial deployment can be done with the safe driver loading feature disabled, and this feature can be enabled later by updating the NicClusterPolicy CRD.

Troubleshooting

Issue

Required Action

The node is in upgrade-failed state.
  • Drain the node manually by running kubectl drain –ignore-daemonsets.
  • Delete the MLNX_OFED pod on the node manually, by running the following command: kubectl delete pod -n `kubectl get pods --A --field-selector spec.nodeName=<node name> -l nvidia.com/ofed-driver --no-headers | awk '{print $1 " "$2}'`.

NOTE: If the “Safe driver loading” feature is enabled, you may also need to remove the nvidia.com/ofed-driver-upgrade.driver-wait-for-safe-load annotation from the node object to unblock the loading of the driver kubectl annotate node <node_name> nvidia.com/ofed-driver-upgrade.driver-wait-for-safe-load-

  • Wait for the node to complete the upgrade.
The updated MLNX_OFED pod failed to start/ a new version of MLNX_OFED cannot be installed on the node. Manually delete the pod by using kubectl delete -n <Network Operator Namespace> <pod name>. If following the restart the pod still fails, change the MLNX_OFED version in the NicClusterPolicy to the previous version or to another working version.

Uninstalling Network Operator on a Vanilla Kubernetes Cluster

Uninstall the Network Operator:

Copy
Copied!
            

helm uninstall network-operator -n nvidia-network-operator

You should now see all the pods being deleted:

Copy
Copied!
            

kubectl get pods -n nvidia-network-operator

Make sure that the CRDs created during the operator installation have been removed:

Copy
Copied!
            

kubectl get nicclusterpolicies.mellanox.com No resources found

Uninstalling the Network Operator on an OpenShift Cluster

From the console:

In the OpenShift Container Platform web console side menu, select Operators >Installed Operators, search for the NVIDIA Network Operator, and click on it.

On the right side of the Operator Details page, select Uninstall Operator from the Actions drop-down menu.

For additional information, see the Red Hat OpenShift Container Platform Documentation.

From the CLI:

  • Check the current version of the Network Operator in the currentCSV field:

    Copy
    Copied!
                

    oc get subscription -n nvidia-network-operator nvidia-network-operator -o yaml | grep currentCSV

    Example output:

    Copy
    Copied!
                

    currentCSV: nvidia-network-operator.v24.1.0

  • Delete the subscription:

    Copy
    Copied!
                

    oc delete subscription -n nvidia-network-operator nvidia-network-operator

    Example output:

    Copy
    Copied!
                

    subscription.operators.coreos.com "nvidia-network-operator" deleted

  • Delete the CSV using the currentCSV value from the previous step:

    Copy
    Copied!
                

    subscription.operators.coreos.com "nvidia-network-operator" deleted

    Example output:

    Copy
    Copied!
                

    clusterserviceversion.operators.coreos.com "nvidia-network-operator.v10.0" deleted

The SR-IOV Network Operator uninstallation procedure is described in this document. For additional information, see the Red Hat OpenShift Container Platform Documentation.

Additional Steps

Warning

In OCP, uninstalling an operator does not remove its managed resources, including CRDs and CRs. To remove them, you must manually delete the Operator CRDs following the operator uninstallation.

Delete the Network Operator CRDs:

Copy
Copied!
            

oc delete crds hostdevicenetworks.mellanox.com macvlannetworks.mellanox.com nicclusterpolicies.mellanox.com

If the NicClusterPolicy manual update affects the device plugin configuration (e.g. NICs selectors), manual device plugin pods restart is required.

Previous CRDs
Next Advanced Configurations
© Copyright 2024, NVIDIA. Last updated on Jun 3, 2024.