Run a GPU-based TensorFlow training job - Elastic Container Instance

This topic describes how to use Elastic Container Instance to run a GPU-based TensorFlow training job based on a Container Service for Kubernetes (ACK) Serverless cluster.

Background information

In recent years, AI and machine learning have been widely applied in a large number of fields, and various training models have been developed. An increasing number of training jobs are run on the cloud. However, it is not easy to continuously run training jobs in a cloud environment. You may encounter the following difficulties:

Difficulties in deploying environments: You must purchase a GPU-accelerated instance and install a GPU driver on the instance. After you prepare images for training jobs, you must install the GPU runtime hook.
Lack of scalability: After you deploy an environment and run the training jobs, you may need to release idle resources to save costs. The next time you want to run training jobs, you must deploy an environment and create instances again. If compute nodes are insufficient, you must scale out the compute nodes. In this case, you must create instances and deploy the environment again.

To resolve the preceding difficulties, we recommend that you use ACK Serverless clusters and elastic container instances to run training jobs. This solution has the following benefits:

You are charged on a pay-as-you-go basis and do not need to manage resources.
You need only to prepare the configurations once. Then, you can reuse the configurations without limits.
The image cache feature allows you to create pods and start training jobs in a more efficient manner.
Training data is decoupled from training models. Training data can be persisted.

Preparations

Prepare a container image and training data for the training model.
In this example, a TensorFlow training job on GitHub is used. You can obtain the sample image eci/tensorflow from Alibaba Cloud Container Registry. For more information, see TensorFlow Model Garden.
Create an ACK Serverless cluster.
Create an ACK Serverless cluster in the ACK console. For more information, see Create an ACK Serverless cluster.
Note
If you want to pull an image from the Internet or if your training jobs need to access the Internet, you must configure a NAT gateway.
You can use kubectl and one of the following methods to manage and access the ACK Serverless cluster:
- If you want to manage the cluster from an on-premises machine, install and configure the kubectl client. For more information, see Connect to an ACK cluster by using kubectl.
- Use kubectl to manage the ACK Serverless cluster on Cloud Shell. For more information, see Use kubectl to manage ACK clusters on Cloud Shell.
Create an Apsara File Storage NAS (NAS) file system and add a mount target.
Create a NAS file system and add a mount target in the Apsara File Storage NAS console. For more information, see Manage file systems and Manage mount targets.
Note
The NAS file system and the ACK Serverless cluster must be in the same virtual private cloud (VPC).

Create an image cache

The image cache feature has been integrated into Kubernetes CustomResourceDefinitions (CRDs) to accelerate the pulling of images. For more information, see Use ImageCache to accelerate the creation of pods.

Perform the following operations:

Prepare a YAML file.

The following code shows an example of the imagecache.yaml file:

apiVersion: eci.alibabacloud.com/v1
kind: ImageCache
metadata:
  name: tensorflow
spec:
  images:
  - registry-vpc.cn-beijing.aliyuncs.com/eci/tensorflow:1.0 # The image of the training job. We recommend that you upload the image to Alibaba Cloud Container Registry. The endpoint of the VPC is used. Make sure that the VPC is the same as the one in which the cluster resides.

Create an image cache.
```
kubectl create -f imagecache.yaml
```
You must pull an image when you create an image cache. The amount of time required to pull an image is related to the image size. You can run the following command to view the creation progress of the image cache:
```
kubectl get imagecache tensorflow
```
A command output similar to the following one indicates that the image cache is created.
```
NAME         AGE   CACHEID                    PHASE   PROGRESS
tensorflow   11m   imc-2ze1xczztv7tgesg****   Ready   100%
```

Create a training job

You can use the image cache to create a training job.

Prepare the YAML file.

The following code shows an example of the gpu_pod.yaml file:

apiVersion: v1
kind: Pod
metadata:
  name: tensorflow
  annotations:
    k8s.aliyun.com/eci-use-specs: "ecs.gn6i-c4g1.xlarge"  # Specify a GPU-accelerated ECS instance type to create an elastic container instance.
    k8s.aliyun.com/eci-auto-imc: "true"             # Enable Automatically Match Image Cache.
spec:
  containers:
  - name: tensorflow
    image: registry-vpc.cn-beijing.aliyuncs.com/eci/tensorflow:1.0 # The image of the training job.
    command:
      - "sh"
      - "-c"
      - "python models/tutorials/image/imagenet/classify_image.py" # The script used to start the training job.
    resources:
      limits:
        nvidia.com/gpu: "1"   # The number of GPUs requested by the container.
    volumeMounts:
    - name: nfs-pv
      mountPath: /tmp/imagenet
  volumes:   
  - name: nfs-pv   #Persist the training results to NAS.
    flexVolume:
        driver:  alicloud/nas
        fsType: nfs
        options:
          server: 16cde4****-ijv**.cn-beijing.nas.aliyuncs.com     # The mount target of the NAS file system.
          path: /         # The mount directory.
  restartPolicy: OnFailure

Run the following command to create a pod:
```
kubectl create -f gpu_pod.yaml
```
View the execution results.
You can view events or logs.
- View events
```
kubectl describe pod tensorflow
```
- View logs
```
kubectl logs tensorflow
```

View the results

You can view the results of the training job in the console.

You can view the storage capacity occupied by the training data in the Apsara File Storage NAS console.
Note
Training results are stored in the configured NAS file system. After the NAS file system is mounted, you can obtain the results in the corresponding path.
You can view the elastic container instance that is successfully run in the Elastic Container Instance console.
Note
If an elastic container instance is successfully run, containers in the instance are stopped. Then, the system reclaims the underlying computing resources, and the billing of pods is stopped.