什么是kubeflow gpu资源节点分配标准？

问题描述

我对Kubeflow GPU资源感到好奇。我正在做下面的工作。 我指定GPU资源的唯一部分是在只有一个GPU的第一个容器上。但是，事件消息告诉我0/4 nodes are available: 4 Insufficient nvidia.com/gpu。 尽管我仅指定了1个GPU资源，但为何此作业搜索4个节点？我的解释有问题吗？提前谢谢。

仅供参考）我有3个工作节点，每个1 gpu。

apiVersion: batch/v1
kind: Job
metadata:
  name: saint-train-3
  annotations:
    sidecar.istio.io/inject: "false"
spec:
  template:
    spec:
      initContainers:
      - name: dataloader
        image: <AWS CLI Image>
        command: ["/bin/sh","-c","aws s3 cp s3://<Kubeflow Bucket>/kubeflowdata.tar.gz /s3-data; cd /s3-data; tar -xvzf kubeflowdata.tar.gz; cd kubeflow_data; ls"]
        volumeMounts:
        - mountPath: /s3-data
          name: s3-data
        env:
        - name: AWS_ACCESS_KEY_ID
          valueFrom:
            secretKeyRef: {key: AWS_ACCESS_KEY_ID,name: aws-secret}
        - name: AWS_SECRET_ACCESS_KEY
          valueFrom:
            secretKeyRef: {key: AWS_SECRET_ACCESS_KEY,name: aws-secret}
      containers:
      - name: trainer
        image: <Our Model Image>
        command: ["/bin/sh","wandb login <ID>; python /opt/ml/src/main.py --base_path='/s3-data/kubeflow_data' --debug_mode='0' --project='kubeflow-test' --name='test2' --gpu=0 --num_epochs=1 --num_workers=4"]
        volumeMounts:
        - mountPath: /s3-data
          name: s3-data
        resources:
          limits:
            nvidia.com/gpu: "1"
      - name: gpu-watcher
        image: pytorch/pytorch:latest
        command: ["/bin/sh","--"]
        args: [ "while true; do sleep 30; done;" ]
        volumeMounts:
        - mountPath: /s3-data
          name: s3-data
      volumes:
      - name: s3-data
        persistentVolumeClaim:
          claimName: test-claim
      restartPolicy: OnFailure
  backoffLimit: 6

解决方法

0/4个节点可用：4个nvidia.com/gpu不足

这意味着您没有带有nvidia.com/gpu标签的节点