Use GPUs for compute-intensive workloads on Azure Kubernetes Service (AKS)

Graphical processing units (GPUs) are often used for compute-intensive workloads, such as graphics and visualization workloads. AKS supports GPU-enabled Linux node pools to run compute-intensive Kubernetes workloads.

Supported GPU-enabled VMs

To view supported GPU-enabled VMs, see GPU-optimized VM sizes in Azure. For AKS node pools, we recommend a minimum size of Standard_NC6s_v3. The NVv4 series (based on AMD GPUs) aren't supported on AKS.

Limitations

• If you're using an Azure Linux GPU-enabled node pool, automatic security patches aren't applied. Refer to your current AKS API version for the default behavior of node OS upgrade channel.

• NVadsA10 v5-series are not a recommended SKU for GPU VHD.
• Updating an existing node pool to add GPU isn't supported.

Before you begin

This article helps you provision nodes with schedulable GPUs on new and existing AKS clusters. This article assumes that you have an existing AKS cluster. If you need an AKS cluster, see the AKS quickstart using the Azure CLI, using Azure PowerShell, or using the Azure portal.

Make sure your Azure subscription could create NCv3 serial VMs, e.g. Standard_NC6s_v3, otherwise you may file a support ticket to enable that kind of VM size for your Azure subscription.

Create an AKS cluster

If you need an AKS cluster that meets the minimum requirements (GPU-enabled node and Kubernetes version 1.10 or later), complete the following steps. If you already have an AKS cluster that meets these requirements, skip to the next section.

Options for using NVIDIA GPUs

Using NVIDIA GPUs involves the installation of various NVIDIA software components such as the NVIDIA device plugin for Kubernetes, GPU driver installation, and more.

NVIDIA device plugin installation

NVIDIA device plugin installation is required when using GPUs on AKS. In some cases, the installation is handled automatically, such as when using the NVIDIA GPU Operator or the AKS GPU image (preview). Alternatively, you can manually install the NVIDIA device plugin.

Manually install the NVIDIA device plugin

You can deploy a DaemonSet for the NVIDIA device plugin, which runs a pod on each node to provide the required drivers for the GPUs. This is the recommended approach when using GPU-enabled node pools for Azure Linux.

1. Create a namespace using the kubectl create namespace command.

   kubectl create namespace gpu-resources
   

2. Create a file named nvidia-device-plugin-ds.yaml and paste the following YAML manifest provided as part of the NVIDIA device plugin for Kubernetes project:

   apiVersion: apps/v1
   kind: DaemonSet
   metadata:
     name: nvidia-device-plugin-daemonset
     namespace: kube-system
   spec:
     selector:
       matchLabels:
         name: nvidia-device-plugin-ds
     updateStrategy:
       type: RollingUpdate
     template:
       metadata:
         labels:
           name: nvidia-device-plugin-ds
       spec:
         tolerations:
         - key: "sku"
           operator: "Equal"
           value: "gpu"
           effect: "NoSchedule"
         # Mark this pod as a critical add-on; when enabled, the critical add-on
         # scheduler reserves resources for critical add-on pods so that they can
         # be rescheduled after a failure.
         # See https://kubernetes.io/docs/tasks/administer-cluster/guaranteed-scheduling-critical-addon-pods/
         priorityClassName: "system-node-critical"
         containers:
         - image: mcr.azk8s.cn/oss/nvidia/k8s-device-plugin:v0.14.1
           name: nvidia-device-plugin-ctr
           env:
             - name: FAIL_ON_INIT_ERROR
               value: "false"
           securityContext:
             allowPrivilegeEscalation: false
             capabilities:
               drop: ["ALL"]
           volumeMounts:
           - name: device-plugin
             mountPath: /var/lib/kubelet/device-plugins
         volumes:
         - name: device-plugin
           hostPath:
             path: /var/lib/kubelet/device-plugins
   

3. Create the DaemonSet and confirm the NVIDIA device plugin is created successfully using the kubectl apply command.

   kubectl apply -f nvidia-device-plugin-ds.yaml
   

4. Now that you successfully installed the NVIDIA device plugin, you can check that your GPUs are schedulable and run a GPU workload.

Skip GPU driver installation (preview)

If you want to control the installation of the NVidia drivers or use the NVIDIA GPU Operator, you can skip the default GPU driver installation. Microsoft doesn't support or manage the maintenance and compatibility of the NVidia drivers as part of the node image deployment.

1. Register or update the aks-preview extension using the az extension add or az extension update command.

   # Register the aks-preview extension
   az extension add --name aks-preview
   
   # Update the aks-preview extension
   az extension update --name aks-preview
   

2. Create a node pool using the az aks nodepool add command with the --skip-gpu-driver-install flag to skip automatic GPU driver installation.

   az aks nodepool add \
       --resource-group myResourceGroup \
       --cluster-name myAKSCluster \
       --name gpunp \
       --node-count 1 \
       --skip-gpu-driver-install \
       --node-vm-size Standard_NC6s_v3 \
       --enable-cluster-autoscaler \
       --min-count 1 \
       --max-count 3
   

   Adding the --skip-gpu-driver-install flag during node pool creation skips the automatic GPU driver installation. Any existing nodes aren't changed. You can scale the node pool to zero and then back up to make the change take effect.

Use NVIDIA GPU Operator with AKS

The NVIDIA GPU Operator automates the management of all NVIDIA software components needed to provision GPU including driver installation, the NVIDIA device plugin for Kubernetes, the NVIDIA container runtime, and more. Since the GPU Operator handles these components, it's not necessary to manually install the NVIDIA device plugin. This also means that the automatic GPU driver installation on AKS is no longer required.

1. Skip automatic GPU driver installation by creating a node pool using the az aks nodepool add command with --skip-gpu-driver-install. Adding the --skip-gpu-driver-install flag during node pool creation skips the automatic GPU driver installation. Any existing nodes aren't changed. You can scale the node pool to zero and then back up to make the change take effect.

2. Follow the NVIDIA documentation to Install the GPU Operator.

3. Now that you successfully installed the GPU Operator, you can check that your GPUs are schedulable and run a GPU workload.

Use the AKS GPU image (preview)

AKS provides a fully configured AKS image containing the NVIDIA device plugin for Kubernetes. The AKS GPU image is currently only supported for Ubuntu 18.04.

1. Install the aks-preview Azure CLI extension using the az extension add command.

   az extension add --name aks-preview
   

2. Update to the latest version of the extension using the az extension update command.

   az extension update --name aks-preview
   

3. Register the GPUDedicatedVHDPreview feature flag using the az feature register command.

   az feature register --namespace "Microsoft.ContainerService" --name "GPUDedicatedVHDPreview"
   

   It takes a few minutes for the status to show Registered.

4. Verify the registration status using the az feature show command.

   az feature show --namespace "Microsoft.ContainerService" --name "GPUDedicatedVHDPreview"
   

5. When the status reflects Registered, refresh the registration of the Microsoft.ContainerService resource provider using the az provider register command.

   az provider register --namespace Microsoft.ContainerService
   

   Now that you updated your cluster to use the AKS GPU image, you can add a node pool for GPU nodes to your cluster.

6. Add a node pool using the az aks nodepool add command.

   az aks nodepool add \
       --resource-group myResourceGroup \
       --cluster-name myAKSCluster \
       --name gpunp \
       --node-count 1 \
       --node-vm-size Standard_NC6s_v3 \
       --node-taints sku=gpu:NoSchedule \
       --aks-custom-headers UseGPUDedicatedVHD=true \
       --enable-cluster-autoscaler \
       --min-count 1 \
       --max-count 3
   

   The previous example command adds a node pool named gpunp to myAKSCluster in myResourceGroup and uses parameters to configure the following node pool settings:

   • --node-vm-size: Sets the VM size for the node in the node pool to Standard_NC6s_v3.
   • --node-taints: Specifies a sku=gpu:NoSchedule taint on the node pool.
   • --aks-custom-headers: Specifies a specialized AKS GPU image, UseGPUDedicatedVHD=true. If your GPU sku requires generation 2 VMs, use --aks-custom-headers UseGPUDedicatedVHD=true,usegen2vm=true instead.
   • --enable-cluster-autoscaler: Enables the cluster autoscaler.
   • --min-count: Configures the cluster autoscaler to maintain a minimum of one node in the node pool.
   • --max-count: Configures the cluster autoscaler to maintain a maximum of three nodes in the node pool.

   Note

   Taints and VM sizes can only be set for node pools during node pool creation, but you can update autoscaler settings at any time.

7. Now that you successfully created a node pool using the GPU image, you can check that your GPUs are schedulable and run a GPU workload.

Confirm that GPUs are schedulable

After creating your cluster, confirm that GPUs are schedulable in Kubernetes.

1. List the nodes in your cluster using the kubectl get nodes command.

   kubectl get nodes
   

   Your output should look similar to the following example output:

   NAME                   STATUS   ROLES   AGE   VERSION
   aks-gpunp-28993262-0   Ready    agent   13m   v1.20.7
   

2. Confirm the GPUs are schedulable using the kubectl describe node command.

   kubectl describe node aks-gpunp-28993262-0
   

   Under the Capacity section, the GPU should list as nvidia.com/gpu: 1. Your output should look similar to the following condensed example output:

   Name:               aks-gpunp-28993262-0
   Roles:              agent
   Labels:             accelerator=nvidia
   
   [...]
   
   Capacity:
   [...]
    nvidia.com/gpu:                 1
   [...]
   

Install GPU plugin

kubectl create -f https://raw.githubusercontent.com/andyzhangx/demo/master/linux/gpu/nvidia-device-plugin-ds-mooncake.yaml

Run a GPU-enabled workload

To see the GPU in action, you can schedule a GPU-enabled workload with the appropriate resource request. In this example, we'll run a Tensorflow job against the MNIST dataset.

1. Create a file named samples-tf-mnist-demo.yaml and paste the following YAML manifest, which includes a resource limit of nvidia.com/gpu: 1:

   Note

   If you receive a version mismatch error when calling into drivers, such as "CUDA driver version is insufficient for CUDA runtime version", review the NVIDIA driver matrix compatibility chart.

   apiVersion: batch/v1
   kind: Job
   metadata:
     labels:
       app: samples-tf-mnist-demo
     name: samples-tf-mnist-demo
   spec:
     template:
       metadata:
         labels:
           app: samples-tf-mnist-demo
       spec:
         containers:
         - name: samples-tf-mnist-demo
           image: mcr.azk8s.cn/azuredocs/samples-tf-mnist-demo:gpu
           args: ["--max_steps", "500"]
           imagePullPolicy: IfNotPresent
           resources:
             limits:
              nvidia.com/gpu: 1
         restartPolicy: OnFailure
         tolerations:
         - key: "sku"
           operator: "Equal"
           value: "gpu"
           effect: "NoSchedule"
   

2. Run the job using the kubectl apply command, which parses the manifest file and creates the defined Kubernetes objects.

   kubectl apply -f samples-tf-mnist-demo.yaml
   

View the status of the GPU-enabled workload

1. Monitor the progress of the job using the kubectl get jobs command with the --watch flag. It may take a few minutes to first pull the image and process the dataset.

   kubectl get jobs samples-tf-mnist-demo --watch
   

   When the COMPLETIONS column shows 1/1, the job has successfully finished, as shown in the following example output:

   NAME                    COMPLETIONS   DURATION   AGE
   
   samples-tf-mnist-demo   0/1           3m29s      3m29s
   samples-tf-mnist-demo   1/1   3m10s   3m36s
   

2. Exit the kubectl --watch process with Ctrl-C.

3. Get the name of the pod using the kubectl get pods command.

   kubectl get pods --selector app=samples-tf-mnist-demo
   

4. View the output of the GPU-enabled workload using the kubectl logs command.

   kubectl logs samples-tf-mnist-demo-smnr6
   

   The following condensed example output of the pod logs confirms that the appropriate GPU device, Tesla K80, has been discovered:

   2019-05-16 16:08:31.258328: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
   2019-05-16 16:08:31.396846: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Found device 0 with properties: 
   name: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235
   pciBusID: 2fd7:00:00.0
   totalMemory: 11.17GiB freeMemory: 11.10GiB
   2019-05-16 16:08:31.396886: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1120] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: Tesla K80, pci bus id: 2fd7:00:00.0, compute capability: 3.7)
   2019-05-16 16:08:36.076962: I tensorflow/stream_executor/dso_loader.cc:139] successfully opened CUDA library libcupti.so.8.0 locally
   Successfully downloaded train-images-idx3-ubyte.gz 9912422 bytes.
   Extracting /tmp/tensorflow/input_data/train-images-idx3-ubyte.gz
   Successfully downloaded train-labels-idx1-ubyte.gz 28881 bytes.
   Extracting /tmp/tensorflow/input_data/train-labels-idx1-ubyte.gz
   Successfully downloaded t10k-images-idx3-ubyte.gz 1648877 bytes.
   Extracting /tmp/tensorflow/input_data/t10k-images-idx3-ubyte.gz
   Successfully downloaded t10k-labels-idx1-ubyte.gz 4542 bytes.
   Extracting /tmp/tensorflow/input_data/t10k-labels-idx1-ubyte.gz
   Accuracy at step 0: 0.1081
   Accuracy at step 10: 0.7457
   Accuracy at step 20: 0.8233
   Accuracy at step 30: 0.8644
   Accuracy at step 40: 0.8848
   Accuracy at step 50: 0.8889
   Accuracy at step 60: 0.8898
   Accuracy at step 70: 0.8979
   Accuracy at step 80: 0.9087
   Accuracy at step 90: 0.9099
   Adding run metadata for 99
   Accuracy at step 100: 0.9125
   Accuracy at step 110: 0.9184
   Accuracy at step 120: 0.922
   Accuracy at step 130: 0.9161
   Accuracy at step 140: 0.9219
   Accuracy at step 150: 0.9151
   Accuracy at step 160: 0.9199
   Accuracy at step 170: 0.9305
   Accuracy at step 180: 0.9251
   Accuracy at step 190: 0.9258
   Adding run metadata for 199
   [...]
   Adding run metadata for 499
   

Use Container Insights to monitor GPU usage

Container Insights with AKS monitors the following GPU usage metrics:

Metric name	Metric dimension (tags)	Description
containerGpuDutyCycle	container.azm.ms/clusterId, container.azm.ms/clusterName, containerName, gpuId, gpuModel, gpuVendor	Percentage of time over the past sample period (60 seconds) during which GPU was busy/actively processing for a container. Duty cycle is a number between 1 and 100.
containerGpuLimits	container.azm.ms/clusterId, container.azm.ms/clusterName, containerName	Each container can specify limits as one or more GPUs. It is not possible to request or limit a fraction of a GPU.
containerGpuRequests	container.azm.ms/clusterId, container.azm.ms/clusterName, containerName	Each container can request one or more GPUs. It is not possible to request or limit a fraction of a GPU.
containerGpumemoryTotalBytes	container.azm.ms/clusterId, container.azm.ms/clusterName, containerName, gpuId, gpuModel, gpuVendor	Amount of GPU Memory in bytes available to use for a specific container.
containerGpumemoryUsedBytes	container.azm.ms/clusterId, container.azm.ms/clusterName, containerName, gpuId, gpuModel, gpuVendor	Amount of GPU Memory in bytes used by a specific container.
nodeGpuAllocatable	container.azm.ms/clusterId, container.azm.ms/clusterName, gpuVendor	Number of GPUs in a node that can be used by Kubernetes.
nodeGpuCapacity	container.azm.ms/clusterId, container.azm.ms/clusterName, gpuVendor	Total Number of GPUs in a node.

Clean up resources

• Remove the associated Kubernetes objects you created in this article using the kubectl delete job command.

  kubectl delete jobs samples-tf-mnist-demo
  

Next steps

• To run Apache Spark jobs, see Run Apache Spark jobs on AKS.
• For more information on features of the Kubernetes scheduler, see Best practices for advanced scheduler features in AKS.
• For more information on Azure Kubernetes Service and Azure Machine Learning, see:
  • Configure a Kubernetes cluster for ML model training or deployment.
  • Deploy a model with an online endpoint.
  • High-performance serving with Triton Inference Server.
  • Labs for Kubernetes and Kubeflow.