Train a model with GPUs on GKE Standard mode


This quickstart tutorial shows you how to deploy a training model with GPUs in Google Kubernetes Engine (GKE) and store the predictions in Cloud Storage. This tutorial uses a TensorFlow model and GKE Standard clusters. You can also run these workloads on Autopilot clusters with fewer setup steps. For instructions, see Train a model with GPUs on GKE Autopilot mode.

This document is intended for GKE administrators who have existing Standard clusters and want to run GPU workloads for the first time.

Before you begin

  1. Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.
  2. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

    Go to project selector

  3. Make sure that billing is enabled for your Google Cloud project.

  4. Enable the Kubernetes Engine and Cloud Storage APIs.

    Enable the APIs

  5. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

    Go to project selector

  6. Make sure that billing is enabled for your Google Cloud project.

  7. Enable the Kubernetes Engine and Cloud Storage APIs.

    Enable the APIs

  8. In the Google Cloud console, activate Cloud Shell.

    Activate Cloud Shell

    At the bottom of the Google Cloud console, a Cloud Shell session starts and displays a command-line prompt. Cloud Shell is a shell environment with the Google Cloud CLI already installed and with values already set for your current project. It can take a few seconds for the session to initialize.

Clone the sample repository

In Cloud Shell, run the following command:

git clone https://github.com/GoogleCloudPlatform/ai-on-gke/ ai-on-gke
cd ai-on-gke/tutorials-and-examples/gpu-examples/training-single-gpu

Create a Standard mode cluster and a GPU node pool

Use Cloud Shell to do the following:

  1. Create a Standard cluster that uses Workload Identity Federation for GKE and installs the Cloud Storage FUSE driver:

    gcloud container clusters create gke-gpu-cluster \
        --addons GcsFuseCsiDriver \
        --location=us-central1 \
        --num-nodes=1 \
        --workload-pool=PROJECT_ID.svc.id.goog
    

    Replace PROJECT_ID with your Google Cloud project ID.

    Cluster creation might take several minutes.

  2. Create a GPU node pool:

    gcloud container node-pools create gke-gpu-pool-1 \
        --accelerator=type=nvidia-tesla-t4,count=1,gpu-driver-version=default \
        --machine-type=n1-standard-16 --num-nodes=1 \
        --location=us-central1 \
        --cluster=gke-gpu-cluster
    

Create a Cloud Storage bucket

  1. In the Google Cloud console, go to the Create a bucket page:

    Go to Create a bucket

  2. In the Name your bucket field, enter the following name:

    PROJECT_ID-gke-gpu-bucket
    
  3. Click Continue.

  4. For Location type, select Region.

  5. In the Region list, select us-central1 (Iowa) and click Continue.

  6. In the Choose a storage class for your data section, click Continue.

  7. In the Choose how to control access to objects section, for Access control, select Uniform.

  8. Click Create.

  9. In the Public access will be prevented dialog, ensure that the Enforce public access prevention on this bucket checkbox is selected, and click Confirm.

Configure your cluster to access the bucket using Workload Identity Federation for GKE

To let your cluster access the Cloud Storage bucket, you do the following:

  1. Create a Google Cloud service account.
  2. Create a Kubernetes ServiceAccount in your cluster.
  3. Bind the Kubernetes ServiceAccount to the Google Cloud service account.

Create a Google Cloud service account

  1. In the Google Cloud console, go to the Create service account page:

    Go to Create service account

  2. In the Service account ID field, enter gke-ai-sa.

  3. Click Create and continue.

  4. In the Role list, select the Cloud Storage > Storage Insights Collector Service role.

  5. Click Add another role.

  6. In the Select a role list, select the Cloud Storage > Storage Object Admin role.

  7. Click Continue, and then click Done.

Create a Kubernetes ServiceAccount in your cluster

In Cloud Shell, do the following:

  1. Create a Kubernetes namespace:

    kubectl create namespace gke-ai-namespace
    
  2. Create a Kubernetes ServiceAccount in the namespace:

    kubectl create serviceaccount gpu-k8s-sa --namespace=gke-ai-namespace
    

Bind the Kubernetes ServiceAccount to the Google Cloud service account

In Cloud Shell, run the following commands:

  1. Add an IAM binding to the Google Cloud service account:

    gcloud iam service-accounts add-iam-policy-binding gke-ai-sa@PROJECT_ID.iam.gserviceaccount.com \
        --role roles/iam.workloadIdentityUser \
        --member "serviceAccount:PROJECT_ID.svc.id.goog[gke-ai-namespace/gpu-k8s-sa]"
    

    The --member flag provides the full identity of the Kubernetes ServiceAccount in Google Cloud.

  2. Annotate the Kubernetes ServiceAccount:

    kubectl annotate serviceaccount gpu-k8s-sa \
        --namespace gke-ai-namespace \
        iam.gke.io/gcp-service-account=gke-ai-sa@PROJECT_ID.iam.gserviceaccount.com
    

Verify that Pods can access the Cloud Storage bucket

  1. In Cloud Shell, create the following environment variables:

    export K8S_SA_NAME=gpu-k8s-sa
    export BUCKET_NAME=PROJECT_ID-gke-gpu-bucket
    

    Replace PROJECT_ID with your Google Cloud project ID.

  2. Create a Pod that has a TensorFlow container:

    envsubst < src/gke-config/standard-tensorflow-bash.yaml | kubectl --namespace=gke-ai-namespace apply -f -
    

    This command substitutes the environment variables that you created into the corresponding references in the manifest. You can also open the manifest in a text editor and replace $K8S_SA_NAME and $BUCKET_NAME with the corresponding values.

  3. Create a sample file in the bucket:

    touch sample-file
    gcloud storage cp sample-file gs://PROJECT_ID-gke-gpu-bucket
    
  4. Wait for your Pod to become ready:

    kubectl wait --for=condition=Ready pod/test-tensorflow-pod -n=gke-ai-namespace --timeout=180s
    

    When the Pod is ready, the output is the following:

    pod/test-tensorflow-pod condition met
    
  5. Open a shell in the Tensorflow container:

    kubectl -n gke-ai-namespace exec --stdin --tty test-tensorflow-pod --container tensorflow -- /bin/bash
    
  6. Try to read the sample file that you created:

    ls /data
    

    The output shows the sample file.

  7. Check the logs to identify the GPU attached to the Pod:

    python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
    

    The output shows the GPU attached to the Pod, similar to the following:

    ...
    PhysicalDevice(name='/physical_device:GPU:0',device_type='GPU')
    
  8. Exit the container:

    exit
    
  9. Delete the sample Pod:

    kubectl delete -f src/gke-config/standard-tensorflow-bash.yaml \
        --namespace=gke-ai-namespace
    

Train and predict using the MNIST dataset

In this section, you run a training workload on the MNIST example dataset.

  1. Copy the example data to the Cloud Storage bucket:

    gcloud storage cp src/tensorflow-mnist-example gs://PROJECT_ID-gke-gpu-bucket/ --recursive
    
  2. Create the following environment variables:

    export K8S_SA_NAME=gpu-k8s-sa
    export BUCKET_NAME=PROJECT_ID-gke-gpu-bucket
    
  3. Review the training Job:

    # Copyright 2023 Google LLC
    #
    # Licensed under the Apache License, Version 2.0 (the "License");
    # you may not use this file except in compliance with the License.
    # You may obtain a copy of the License at
    #
    #      http://www.apache.org/licenses/LICENSE-2.0
    #
    # Unless required by applicable law or agreed to in writing, software
    # distributed under the License is distributed on an "AS IS" BASIS,
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    # See the License for the specific language governing permissions and
    # limitations under the License.
    
    apiVersion: batch/v1
    kind: Job
    metadata:
      name: mnist-training-job
    spec:
      template:
        metadata:
          name: mnist
          annotations:
            gke-gcsfuse/volumes: "true"
        spec:
          nodeSelector:
            cloud.google.com/gke-accelerator: nvidia-tesla-t4
          tolerations:
          - key: "nvidia.com/gpu"
            operator: "Exists"
            effect: "NoSchedule"
          containers:
          - name: tensorflow
            image: tensorflow/tensorflow:latest-gpu 
            command: ["/bin/bash", "-c", "--"]
            args: ["cd /data/tensorflow-mnist-example; pip install -r requirements.txt; python tensorflow_mnist_train_distributed.py"]
            resources:
              limits:
                nvidia.com/gpu: 1
                cpu: 1
                memory: 3Gi
            volumeMounts:
            - name: gcs-fuse-csi-vol
              mountPath: /data
              readOnly: false
          serviceAccountName: $K8S_SA_NAME
          volumes:
          - name: gcs-fuse-csi-vol
            csi:
              driver: gcsfuse.csi.storage.gke.io
              readOnly: false
              volumeAttributes:
                bucketName: $BUCKET_NAME
                mountOptions: "implicit-dirs"
          restartPolicy: "Never"
  4. Deploy the training Job:

    envsubst < src/gke-config/standard-tf-mnist-train.yaml | kubectl -n gke-ai-namespace apply -f -
    

    This command substitutes the environment variables that you created into the corresponding references in the manifest. You can also open the manifest in a text editor and replace $K8S_SA_NAME and $BUCKET_NAME with the corresponding values.

  5. Wait until the Job has the Completed status:

    kubectl wait -n gke-ai-namespace --for=condition=Complete job/mnist-training-job --timeout=180s
    

    The output is similar to the following:

    job.batch/mnist-training-job condition met
    
  6. Check the logs from the Tensorflow container:

    kubectl logs -f jobs/mnist-training-job -c tensorflow -n gke-ai-namespace
    

    The output shows the following events occur:

    • Install required Python packages
    • Download the MNIST dataset
    • Train the model using a GPU
    • Save the model
    • Evaluate the model
    ...
    Epoch 12/12
    927/938 [============================>.] - ETA: 0s - loss: 0.0188 - accuracy: 0.9954
    Learning rate for epoch 12 is 9.999999747378752e-06
    938/938 [==============================] - 5s 6ms/step - loss: 0.0187 - accuracy: 0.9954 - lr: 1.0000e-05
    157/157 [==============================] - 1s 4ms/step - loss: 0.0424 - accuracy: 0.9861
    Eval loss: 0.04236088693141937, Eval accuracy: 0.9861000180244446
    Training finished. Model saved
    
  7. Delete the training workload:

    kubectl -n gke-ai-namespace delete -f src/gke-config/standard-tf-mnist-train.yaml
    

Deploy an inference workload

In this section, you deploy an inference workload that takes a sample dataset as input and returns predictions.

  1. Copy the images for prediction to the bucket:

    gcloud storage cp data/mnist_predict gs://PROJECT_ID-gke-gpu-bucket/ --recursive
    
  2. Review the inference workload:

    # Copyright 2023 Google LLC
    #
    # Licensed under the Apache License, Version 2.0 (the "License");
    # you may not use this file except in compliance with the License.
    # You may obtain a copy of the License at
    #
    #      http://www.apache.org/licenses/LICENSE-2.0
    #
    # Unless required by applicable law or agreed to in writing, software
    # distributed under the License is distributed on an "AS IS" BASIS,
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    # See the License for the specific language governing permissions and
    # limitations under the License.
    
    apiVersion: batch/v1
    kind: Job
    metadata:
      name: mnist-batch-prediction-job
    spec:
      template:
        metadata:
          name: mnist
          annotations:
            gke-gcsfuse/volumes: "true"
        spec:
          nodeSelector:
            cloud.google.com/gke-accelerator: nvidia-tesla-t4
          tolerations:
          - key: "nvidia.com/gpu"
            operator: "Exists"
            effect: "NoSchedule"
          containers:
          - name: tensorflow
            image: tensorflow/tensorflow:latest-gpu 
            command: ["/bin/bash", "-c", "--"]
            args: ["cd /data/tensorflow-mnist-example; pip install -r requirements.txt; python tensorflow_mnist_batch_predict.py"]
            resources:
              limits:
                nvidia.com/gpu: 1
                cpu: 1
                memory: 3Gi
            volumeMounts:
            - name: gcs-fuse-csi-vol
              mountPath: /data
              readOnly: false
          serviceAccountName: $K8S_SA_NAME
          volumes:
          - name: gcs-fuse-csi-vol
            csi:
              driver: gcsfuse.csi.storage.gke.io
              readOnly: false
              volumeAttributes:
                bucketName: $BUCKET_NAME
                mountOptions: "implicit-dirs"
          restartPolicy: "Never"
  3. Deploy the inference workload:

    envsubst < src/gke-config/standard-tf-mnist-batch-predict.yaml | kubectl -n gke-ai-namespace apply -f -
    

    This command substitutes the environment variables that you created into the corresponding references in the manifest. You can also open the manifest in a text editor and replace $K8S_SA_NAME and $BUCKET_NAME with the corresponding values.

  4. Wait until the Job has the Completed status:

    kubectl wait -n gke-ai-namespace --for=condition=Complete job/mnist-batch-prediction-job --timeout=180s
    

    The output is similar to the following:

    job.batch/mnist-batch-prediction-job condition met
    
  5. Check the logs from the Tensorflow container:

    kubectl logs -f jobs/mnist-batch-prediction-job -c tensorflow -n gke-ai-namespace
    

    The output is the prediction for each image and the model's confidence in the prediction, similar to the following:

    Found 10 files belonging to 1 classes.
    1/1 [==============================] - 2s 2s/step
    The image /data/mnist_predict/0.png is the number 0 with a 100.00 percent confidence.
    The image /data/mnist_predict/1.png is the number 1 with a 99.99 percent confidence.
    The image /data/mnist_predict/2.png is the number 2 with a 100.00 percent confidence.
    The image /data/mnist_predict/3.png is the number 3 with a 99.95 percent confidence.
    The image /data/mnist_predict/4.png is the number 4 with a 100.00 percent confidence.
    The image /data/mnist_predict/5.png is the number 5 with a 100.00 percent confidence.
    The image /data/mnist_predict/6.png is the number 6 with a 99.97 percent confidence.
    The image /data/mnist_predict/7.png is the number 7 with a 100.00 percent confidence.
    The image /data/mnist_predict/8.png is the number 8 with a 100.00 percent confidence.
    The image /data/mnist_predict/9.png is the number 9 with a 99.65 percent confidence.
    

Clean up

To avoid incurring charges to your Google Cloud account for the resources that you created in this guide, do one of the following:

  • Keep the GKE cluster: Delete the Kubernetes resources in the cluster and the Google Cloud resources
  • Keep the Google Cloud project: Delete the GKE cluster and the Google Cloud resources
  • Delete the project

Delete the Kubernetes resources in the cluster and the Google Cloud resources

  1. Delete the Kubernetes namespace and the workloads that you deployed:

    kubectl -n gke-ai-namespace delete -f src/gke-config/standard-tf-mnist-batch-predict.yaml
    kubectl delete namespace gke-ai-namespace
    
  2. Delete the Cloud Storage bucket:

    1. Go to the Buckets page:

      Go to Buckets

    2. Select the checkbox for PROJECT_ID-gke-gpu-bucket.

    3. Click Delete.

    4. To confirm deletion, type DELETE and click Delete.

  3. Delete the Google Cloud service account:

    1. Go to the Service accounts page:

      Go to Service accounts

    2. Select your project.

    3. Select the checkbox for gke-ai-sa@PROJECT_ID.iam.gserviceaccount.com.

    4. Click Delete.

    5. To confirm deletion, click Delete.

Delete the GKE cluster and the Google Cloud resources

  1. Delete the GKE cluster:

    1. Go to the Clusters page:

      Go to Clusters

    2. Select the checkbox for gke-gpu-cluster.

    3. Click Delete.

    4. To confirm deletion, type gke-gpu-cluster and click Delete.

  2. Delete the Cloud Storage bucket:

    1. Go to the Buckets page:

      Go to Buckets

    2. Select the checkbox for PROJECT_ID-gke-gpu-bucket.

    3. Click Delete.

    4. To confirm deletion, type DELETE and click Delete.

  3. Delete the Google Cloud service account:

    1. Go to the Service accounts page:

      Go to Service accounts

    2. Select your project.

    3. Select the checkbox for gke-ai-sa@PROJECT_ID.iam.gserviceaccount.com.

    4. Click Delete.

    5. To confirm deletion, click Delete.

Delete the project

  1. In the Google Cloud console, go to the Manage resources page.

    Go to Manage resources

  2. In the project list, select the project that you want to delete, and then click Delete.
  3. In the dialog, type the project ID, and then click Shut down to delete the project.

What's next