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 the creation of GPU-enabled node pools to run these compute-intensive workloads in Kubernetes. For more information on available GPU-enabled VMs, see GPU optimized VM sizes in Azure. For AKS nodes, we recommend a minimum size of Standard_NC6.

Note

GPU-enabled VMs contain specialized hardware that is subject to higher pricing and region availability. For more information, see the pricing tool and region availability.

Currently, using GPU-enabled node pools is only available for Linux node pools.

Before you begin

This article assumes that you have an existing AKS cluster with nodes that support GPUs. Your AKS cluster must run Kubernetes 1.10 or later. If you need an AKS cluster that meets these requirements, see the first section of this article to create an AKS cluster.

You also need the Azure CLI version 2.0.64 or later installed and configured. RunĀ az --version to find the version. If you need to install or upgrade, seeĀ Install Azure CLI.

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.

First, create a resource group for the cluster using the az group create command. The following example creates a resource group name myResourceGroup in the eastus region:

az group create --name myResourceGroup --location eastus

Now create an AKS cluster using the az aks create command. The following example creates a cluster with a single node of size Standard_NC6:

az aks create \
    --resource-group myResourceGroup \
    --name myAKSCluster \
    --node-vm-size Standard_NC6 \
    --node-count 1

Get the credentials for your AKS cluster using the az aks get-credentials command:

az aks get-credentials --resource-group myResourceGroup --name myAKSCluster

Install nVidia drivers

Before the GPUs in the nodes can be used, you must deploy a DaemonSet for the NVIDIA device plugin. This DaemonSet runs a pod on each node to provide the required drivers for the GPUs.

First, create a namespace using the kubectl create namespace command, such as gpu-resources:

kubectl create namespace gpu-resources

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

apiVersion: extensions/v1beta1
kind: DaemonSet
metadata:
  name: nvidia-device-plugin-daemonset
  namespace: kube-system
spec:
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      # 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.  This annotation works in tandem with the toleration below.
      annotations:
        scheduler.alpha.kubernetes.io/critical-pod: ""
      labels:
        name: nvidia-device-plugin-ds
    spec:
      tolerations:
      # Allow this pod to be rescheduled while the node is in "critical add-ons only" mode.
      # This, along with the annotation above marks this pod as a critical add-on.
      - key: CriticalAddonsOnly
        operator: Exists
      - key: nvidia.com/gpu
        operator: Exists
        effect: NoSchedule
      containers:
      - image: nvidia/k8s-device-plugin:1.11
        name: nvidia-device-plugin-ctr
        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

Now use the kubectl apply command to create the DaemonSet and confirm the nVidia device plugin is created successfully, as shown in the following example output:

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

daemonset "nvidia-device-plugin" created

Confirm that GPUs are schedulable

With your AKS cluster created, confirm that GPUs are schedulable in Kubernetes. First, list the nodes in your cluster using the kubectl get nodes command:

$ kubectl get nodes

NAME                       STATUS   ROLES   AGE   VERSION
aks-nodepool1-28993262-0   Ready    agent   13m   v1.12.7

Now use the kubectl describe node command to confirm that the GPUs are schedulable. Under the Capacity section, the GPU should list as nvidia.com/gpu: 1.

The following condensed example shows that a GPU is available on the node named aks-nodepool1-18821093-0:

$ kubectl describe node aks-nodepool1-28993262-0

Name:               aks-nodepool1-28993262-0
Roles:              agent
Labels:             accelerator=nvidia

[...]

Capacity:
 attachable-volumes-azure-disk:  24
 cpu:                            6
 ephemeral-storage:              101584140Ki
 hugepages-1Gi:                  0
 hugepages-2Mi:                  0
 memory:                         57713784Ki
 nvidia.com/gpu:                 1
 pods:                           110
Allocatable:
 attachable-volumes-azure-disk:  24
 cpu:                            5916m
 ephemeral-storage:              93619943269
 hugepages-1Gi:                  0
 hugepages-2Mi:                  0
 memory:                         51702904Ki
 nvidia.com/gpu:                 1
 pods:                           110
System Info:
 Machine ID:                 b0cd6fb49ffe4900b56ac8df2eaa0376
 System UUID:                486A1C08-C459-6F43-AD6B-E9CD0F8AEC17
 Boot ID:                    f134525f-385d-4b4e-89b8-989f3abb490b
 Kernel Version:             4.15.0-1040-azure
 OS Image:                   Ubuntu 16.04.6 LTS
 Operating System:           linux
 Architecture:               amd64
 Container Runtime Version:  docker://1.13.1
 Kubelet Version:            v1.12.7
 Kube-Proxy Version:         v1.12.7
PodCIDR:                     10.244.0.0/24
ProviderID:                  azure:///subscriptions/<guid>/resourceGroups/MC_myResourceGroup_myAKSCluster_eastus/providers/Microsoft.Compute/virtualMachines/aks-nodepool1-28993262-0
Non-terminated Pods:         (9 in total)
  Namespace                  Name                                     CPU Requests  CPU Limits  Memory Requests  Memory Limits  AGE
  ---------                  ----                                     ------------  ----------  ---------------  -------------  ---
  kube-system                nvidia-device-plugin-daemonset-bbjlq     0 (0%)        0 (0%)      0 (0%)           0 (0%)         2m39s

[...]

Run a GPU-enabled workload

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

Create a file named samples-tf-mnist-demo.yaml and paste the following YAML manifest. The following job manifest 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 - https://docs.nvidia.com/deploy/cuda-compatibility/index.html

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: microsoft/samples-tf-mnist-demo:gpu
        args: ["--max_steps", "500"]
        imagePullPolicy: IfNotPresent
        resources:
          limits:
           nvidia.com/gpu: 1
      restartPolicy: OnFailure

Use the kubectl apply command to run the job. This command parses the manifest file and creates the defined Kubernetes objects:

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

View the status and output of the GPU-enabled workload

Monitor the progress of the job using the kubectl get jobs command with the --watch argument. It may take a few minutes to first pull the image and process the dataset. When the COMPLETIONS column shows 1/1, the job has successfully finished. Exit the kubetctl --watch command with Ctrl-C:

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

NAME                    COMPLETIONS   DURATION   AGE

samples-tf-mnist-demo   0/1           3m29s      3m29s
samples-tf-mnist-demo   1/1   3m10s   3m36s

To look at the output of the GPU-enabled workload, first get the name of the pod with the kubectl get pods command:

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

NAME                          READY   STATUS      RESTARTS   AGE
samples-tf-mnist-demo-mtd44   0/1     Completed   0          4m39s

Now use the kubectl logs command to view the pod logs. The following example pod logs confirm that the appropriate GPU device has been discovered, Tesla K80. Provide the name for your own pod:

$ kubectl logs samples-tf-mnist-demo-smnr6

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
Accuracy at step 200: 0.9315
Accuracy at step 210: 0.9361
Accuracy at step 220: 0.9357
Accuracy at step 230: 0.9392
Accuracy at step 240: 0.9387
Accuracy at step 250: 0.9401
Accuracy at step 260: 0.9398
Accuracy at step 270: 0.9407
Accuracy at step 280: 0.9434
Accuracy at step 290: 0.9447
Adding run metadata for 299
Accuracy at step 300: 0.9463
Accuracy at step 310: 0.943
Accuracy at step 320: 0.9439
Accuracy at step 330: 0.943
Accuracy at step 340: 0.9457
Accuracy at step 350: 0.9497
Accuracy at step 360: 0.9481
Accuracy at step 370: 0.9466
Accuracy at step 380: 0.9514
Accuracy at step 390: 0.948
Adding run metadata for 399
Accuracy at step 400: 0.9469
Accuracy at step 410: 0.9489
Accuracy at step 420: 0.9529
Accuracy at step 430: 0.9507
Accuracy at step 440: 0.9504
Accuracy at step 450: 0.951
Accuracy at step 460: 0.9512
Accuracy at step 470: 0.9539
Accuracy at step 480: 0.9533
Accuracy at step 490: 0.9494
Adding run metadata for 499

Clean up resources

To remove the associated Kubernetes objects created in this article, use the kubectl delete job command as follows:

kubectl delete jobs samples-tf-mnist-demo

Next steps

To run Apache Spark jobs, see Run Apache Spark jobs on AKS.

For more information about running machine learning (ML) workloads on Kubernetes, see Kubeflow Labs.