Use Alibaba Cloud Kubernetes GPU node labels for scheduling

When you implement GPU computing through a Kubernetes cluster, you can schedule an application to the node installed with GPU devices as needed by using GPU node labels.

1. Log on to the Container Service console.
2. In the left-side navigation pane under Container Service-Kubernetes, choose Clusters > Nodes.
   Note In this example, the cluster has three Worker nodes of which two Worker nodes are mounted with GPU devices. You need to view the node IP addresses for verification.

   
   
   

3. Select a GPU node, and choose More > Details in the action column. Then, you can view the GPU node label on the Kubernetes dashboard.

   
   
   

   You can also log on to a Master node and run the following command to view the GPU node label:

   # kubectl get nodes
   NAME                                STATUS    ROLES     AGE       VERSION
   cn-beijing.i-2ze2dy2h9w97v65uuaft   Ready     master    2d        v1.11.2
   cn-beijing.i-2ze8o1a45qdv5q8a7luz   Ready     <none>    2d        v1.11.2             #Compare this node with the node displayed in the console to determine the GPU node.
   cn-beijing.i-2ze8o1a45qdv5q8a7lv0   Ready     <none>    2d        v1.11.2
   cn-beijing.i-2ze9xylyn11vop7g5bwe   Ready     master    2d        v1.11.2
   cn-beijing.i-2zed5sw8snjniq6mf5e5   Ready     master    2d        v1.11.2
   cn-beijing.i-2zej9s0zijykp9pwf7lu   Ready     <none>    2d        v1.11.2
   					

   Select a GPU node and run the following command to view the GPU node label:

   # kubectl describe node cn-beijing.i-2ze8o1a45qdv5q8a7luz
   Name:               cn-beijing.i-2ze8o1a45qdv5q8a7luz
   Roles:              <none>
   Labels:             aliyun.accelerator/nvidia_count=1                          #This field is important.
                       aliyun.accelerator/nvidia_mem=12209MiB
                       aliyun.accelerator/nvidia_name=Tesla-M40
                       beta.kubernetes.io/arch=amd64
                       beta.kubernetes.io/instance-type=ecs.gn4-c4g1.xlarge
                       beta.kubernetes.io/os=linux
                       failure-domain.beta.kubernetes.io/region=cn-beijing
                       failure-domain.beta.kubernetes.io/zone=cn-beijing-a
                       kubernetes.io/hostname=cn-beijing.i-2ze8o1a45qdv5q8a7luz
    ......

   In this example, the GPU node contains the following three node labels:

   Key	Value
   aliyun.accelerator/nvidia_count	Number of GPU cores
   aliyun.accelerator/nvidia_mem	GPU memory in MiB
   aliyun.accelerator/nvidia_name	Name of the GPU computing card of the NVIDIA device

   The GPU cloud servers of the same type share the same GPU computing card name. Therefore, you can use this label to filter nodes.

   # kubectl get no -l aliyun.accelerator/nvidia_name=Tesla-M40
   NAME                                STATUS    ROLES     AGE       VERSION
   cn-beijing.i-2ze8o1a45qdv5q8a7luz   Ready     <none>    2d        v1.11.2
   cn-beijing.i-2ze8o1a45qdv5q8a7lv0   Ready     <none>    2d        v1.11.2
   					

4. Return to the Container Service console home page. Then, in the left-side navigation pane, choose Applications > Deployments, and click Create by Template in the upper-right corner.
   1. Create a TensorFlow application and schedule this application to the GPU node.

      
      
      

      In this example, the YAML template is orchestrated as follows:

      ---
      # Define the tensorflow deployment
      apiVersion: apps/v1
      kind: Deployment
      metadata:
        name: tf-notebook
        labels:
          app: tf-notebook
      spec:
        replicas: 1
        selector: # define how the deployment finds the pods it manages
          matchLabels:
            app: tf-notebook
        template: #Define the pod specifications.
          metadata:
            labels:
              app: tf-notebook
          spec:
            nodeSelector:                                                  #This field is important.
              aliyun.accelerator/nvidia_name: Tesla-M40
            containers:
            - name: tf-notebook
              image: tensorflow/tensorflow:1.4.1-gpu-py3
              resources:
                limits:
                  nvidia.com/gpu: 1                                        #This field is important.
              ports:
              - containerPort: 8888
                hostPort: 8888
              env:
                - name: PASSWORD
                  value: mypassw0rdv

   2. You can also avoid deploying an application to a GPU node. The following deploys an Nginx pod and schedules it by using the node affinity feature. For more information about node affinity, see Create a deployment application by using an image.

      The example YAML template is orchestrated as follows:

      apiVersion: v1
      kind: Pod
      metadata:
        name: not-in-gpu-node
      spec:
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: aliyun.accelerator/nvidia_name
                  operator: DoesNotExist
        containers:
        - name: not-in-gpu-node
          image: nginx

5. In the left-side navigation pane, choose Applications > Pods, and select the target cluster and namespace.