GPU-enabled clusters


With CAPZ you can create GPU-enabled Kubernetes clusters on Microsoft Azure.

Before you begin, be aware that:

  • Scheduling GPUs is a Kubernetes beta feature
  • NVIDIA GPUs are supported on Azure NC-series, NV-series, and NVv3-series VMs
  • NVIDIA GPU Operator allows administrators of Kubernetes clusters to manage GPU nodes just like CPU nodes in the cluster.

To deploy a cluster with support for GPU nodes, use the nvidia-gpu flavor.

An example GPU cluster

Let’s create a CAPZ cluster with an N-series node and run a GPU-powered vector calculation.

Generate an nvidia-gpu cluster template

Use the clusterctl generate cluster command to generate a manifest that defines your GPU-enabled workload cluster.

Remember to use the nvidia-gpu flavor with N-series nodes.

AZURE_LOCATION=southcentralus \
clusterctl generate cluster azure-gpu \
  --kubernetes-version=v1.22.1 \
  --worker-machine-count=1 \
  --flavor=nvidia-gpu > azure-gpu-cluster.yaml

Create the cluster

Apply the manifest from the previous step to your management cluster to have CAPZ create a workload cluster:

$ kubectl apply -f azure-gpu-cluster.yaml --server-side serverside-applied serverside-applied serverside-applied serverside-applied serverside-applied serverside-applied serverside-applied serverside-applied
configmap/nvidia-clusterpolicy-crd serverside-applied
configmap/nvidia-gpu-operator-components serverside-applied serverside-applied

Wait until the cluster and nodes are finished provisioning. The GPU nodes may take several minutes to provision, since each one must install drivers and supporting software.

$ kubectl get cluster azure-gpu
azure-gpu   Provisioned
$ kubectl get machines
NAME                             PROVIDERID                                                                                                                                     PHASE     VERSION
azure-gpu-control-plane-t94nm    azure:////subscriptions/<subscription_id>/resourceGroups/azure-gpu/providers/Microsoft.Compute/virtualMachines/azure-gpu-control-plane-nnb57   Running   v1.22.1
azure-gpu-md-0-f6b88dd78-vmkph   azure:////subscriptions/<subscription_id>/resourceGroups/azure-gpu/providers/Microsoft.Compute/virtualMachines/azure-gpu-md-0-gcc8v            Running   v1.22.1

Install a CNI of your choice. Once the nodes are Ready, run the following commands against the workload cluster to check if all the gpu-operator resources are installed:

$ clusterctl get kubeconfig azure-gpu > azure-gpu-cluster.conf
$ export KUBECONFIG=azure-gpu-cluster.conf
$ kubectl get pods | grep gpu-operator
default                  gpu-operator-1612821988-node-feature-discovery-master-664dnsmww   1/1     Running                 0          107m
default                  gpu-operator-1612821988-node-feature-discovery-worker-64mcz       1/1     Running                 0          107m
default                  gpu-operator-1612821988-node-feature-discovery-worker-h5rws       1/1     Running                 0          107m
$ kubectl get pods -n gpu-operator-resources
NAME                                       READY   STATUS      RESTARTS   AGE
gpu-feature-discovery-66d4f                1/1     Running     0          2s
nvidia-container-toolkit-daemonset-lxpkx   1/1     Running     0          3m11s
nvidia-dcgm-exporter-wwnsw                 1/1     Running     0          5s
nvidia-device-plugin-daemonset-lpdwz       1/1     Running     0          13s
nvidia-device-plugin-validation            0/1     Completed   0          10s
nvidia-driver-daemonset-w6lpb              1/1     Running     0          3m16s

Then run the following commands against the workload cluster to verify that the NVIDIA device plugin has initialized and the resource is available:

$ kubectl -n kube-system get po | grep nvidia
kube-system   nvidia-device-plugin-daemonset-d5dn6                    1/1     Running   0          16m
$ kubectl get nodes
NAME                            STATUS   ROLES    AGE   VERSION
azure-gpu-control-plane-nnb57   Ready    master   42m   v1.22.1
azure-gpu-md-0-gcc8v            Ready    <none>   38m   v1.22.1
$ kubectl get node azure-gpu-md-0-gcc8v -o jsonpath={.status.allocatable} | jq
  "attachable-volumes-azure-disk": "12",
  "cpu": "6",
  "ephemeral-storage": "119716326407",
  "hugepages-1Gi": "0",
  "hugepages-2Mi": "0",
  "memory": "115312060Ki",
  "": "1",
  "pods": "110"

Run a test app

Let’s create a pod manifest for the cuda-vector-add example from the Kubernetes documentation and deploy it:

$ cat > cuda-vector-add.yaml << EOF
apiVersion: v1
kind: Pod
  name: cuda-vector-add
  restartPolicy: OnFailure
    - name: cuda-vector-add
      image: ""
 1 # requesting 1 GPU
$ kubectl apply -f cuda-vector-add.yaml

The container will download, run, and perform a CUDA calculation with the GPU.

$ kubectl get po cuda-vector-add
cuda-vector-add   0/1     Completed   0          91s
$ kubectl logs cuda-vector-add
[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory

If you see output like the above, your GPU cluster is working!