Customization Options and CRDs
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:
kubectl get -n network-operator nicclusterpolicies.mellanox.com nic-cluster-policy -o yaml
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
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:
$ 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-24.1.1.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.
$ 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.
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:
$ helm upgrade -n nvidia-network-operator network-operator nvidia/network-operator --version=<VERSION> -f values-<VERSION>.yaml --force
The --devel option is required if you wish to use the Beta release.
OFED Driver Manual Upgrade
Restarting Pods with a Containerized OFED Driver
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:
$ 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:
Remove pods with a secondary network from the node.
Restart the OFED driver pod.
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:
$ kubectl drain <NODE_NAME> --pod-selector=<SELECTOR_FOR_PODS>
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:
$ kubectl get pod -l app=mofed-<OS_NAME> -o wide -A
Example for Ubuntu 20.04:
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:
$ kubectl delete pod -n <DRIVER_NAMESPACE> <OFED_POD_NAME>
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:
$ 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: 24.01
-0.3
.3.1
-10
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:
$ kubectl apply -f nicclusterpolicy.yaml
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. |
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: 24.01
-0.3
.3.1
-10
upgradePolicy:
autoUpgrade: true
maxParallelUpgrades: 1
drain:
enable: true
force: false
deleteEmptyDir: true
podSelector: ""
Safe Driver Loading
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. |
|
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:
helm uninstall network-operator -n network-operator
You should now see all the pods being deleted:
kubectl get pods -n network-operator
Make sure that the CRDs created during the operator installation have been removed:
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:
oc get subscription -n nvidia-network-operator nvidia-network-operator -o yaml |
grep
currentCSVExample output:
currentCSV: nvidia-network-operator.v24.
1.1
Delete the subscription:
oc delete subscription -n nvidia-network-operator nvidia-network-operator
Example output:
subscription.operators.coreos.com
"nvidia-network-operator"
deletedDelete the CSV using the currentCSV value from the previous step:
subscription.operators.coreos.com
"nvidia-network-operator"
deletedExample output:
clusterserviceversion.operators.coreos.com "nvidia-network-operator.v
10.0
" 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
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:
$ 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.