This topic describes how to use a GPU container instance by using an example in which Tensorflow is used to identify a picture. ACK serverless Kubernetes provides GPU container instances by integrating with Elastic Container Instance (ECI). As a part of this service, GPU container instances can be used in a serverless Kubernetes cluster or on a virtual node of a dedicated or managed Kubernetes cluster.

Billing

GPU and CPU resources provided by ACK serverless Kubernetes are charged with the Pay-As-You-Go billing method.

GPU instances provided by ACK serverless Kubernetes are charged the same as the ECS GPU instances. You are not charged additional fees.

Overview

In an ACK serverless Kubernetes cluster, you can easily create a pod to which a GPU is attached. Specifically, for the pod configurations, you must set the annotations parameter to specify the GPU type that you want, and set the limits parameter to specify the number of GPU.

Note A GPU is exclusive to one pod.

Prerequisites

Either a serverless Kubernetes cluster is created, or a virtual node is created for a Kubernetes cluster. For more information, see Create a serverless Kubernetes cluster or Deploy a virtual node.

Procedure

In the following steps, Tensorflow is used in a serverless Kubernetes cluster to identify the following picture.

  1. Log on to the Container Service console.
  2. In the left-side navigation pane under Container Service-Kubernetes, choose Applications > Deployments. Then, in the upper-right corner, click Create by Template.
  3. Select the target cluster and namespace, select an example template or customize a template, and then click DEPLOY.
    You can use the following YAML template to create a pod:
    Note In this pod template, the GPU type is set to P4, and the number of GPU is set to 1. 
    apiVersion: v1
kind: Pod
metadata:
  name: tensorflow
  annotations:
    k8s.aliyun.com/eci-gpu-type : "P4"
spec:
  containers:
  - image: registry-vpc.cn-hangzhou.aliyuncs.com/ack-serverless/tensorflow
    name: tensorflow
    command:
    - "sh"
    - "-c"
    - "python models/tutorials/image/imagenet/classify_image.py"
    resources:
      limits:
        nvidia.com/gpu: "1"
  restartPolicy: OnFailure
  4. In the left-side navigation pane under Container Service-Kubernetes, choose Applications > Pods to verify that the pod is created.
    Note It may take a few minutes to create a pod. If you the pod is not displayed, you can refresh the page.


  5. Click the target pod, and click the Logs tab to verify that the picture is identified according to the logs marked in the red frame.

If you want to use a GPU container instance on a virtual node of a dedicated or managed Kubernetes cluster, follow these steps:
Note A GPU container instance used on a virtual node supports a larger variety of deep learning platforms such as Kubeflow, Arena, or other customized CRD.
  1. Create a pod on the virtual node or in a namespace to which the virtual-node-affinity-injection=enabled label tag is added. For more information, see Deploy a virtual node.
  2. Replace the template used in step 3 with the following template: 
    apiVersion: v1
kind: Pod
metadata:
  name: tensorflow
  annotations:
    k8s.aliyun.com/eci-gpu-type : "P4"
spec:
  containers:
  - image: registry-vpc.cn-hangzhou.aliyuncs.com/ack-serverless/tensorflow
    name: tensorflow
    command:
    - "sh"
    - "-c"
    - "python models/tutorials/image/imagenet/classify_image.py"
    resources:
      limits:
        nvidia.com/gpu: "1"
  restartPolicy: OnFailure
  nodeName: virtual-kubelet