NVIDIA官方文档：NVIDIA/k8s-device-plugin：用于 Kubernetes 的 NVIDIA 设备插件

前提条件

安装NVIDIA Container Toolkit

官方地址：https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html

curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo | \
  sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
yum install -y nvidia-container-toolkit

更新运行时

docker

nvidia-ctk runtime configure --runtime=docker
systemctl restart docker

无根模式下运行的docker

nvidia-ctk runtime configure --runtime=docker --config=$HOME/.config/docker/daemon.json
systemctl --user restart docker
nvidia-ctk config --set nvidia-container-cli.no-cgroups --in-place

Containerd

nvidia-ctk runtime configure --runtime=containerd

systemctl daemon-reload
systemctl restart containerd

仅用于nerdctl无需配置，直接运行nerdctl run --gpus=all

如果以上修改不生效，修改默认运行时

# 编辑 /etc/containerd/config.toml

搜索 plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc

runtime_type = "io.containerd.runc.v2" 改成 io.containerd.runtime.v1.linux

搜索io.containerd.runtime.v1.linux

runtime = "runc"改成runtime = "nvidia-container-runtime"

systemctl daemon-reload
systemctl restart containerd

CRI-O

nvidia-ctk runtime configure --runtime=crio
systemctl restart crio

部署nvidia-device-plugin

kubectl create -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.16.2/deployments/static/nvidia-device-plugin.yml

查看GPU所在节点的pod日志

正常情况（将GPU注册到kubelet）：

I0820 08:26:47.926058       1 main.go:317] Retrieving plugins.
I0820 08:26:51.674618       1 server.go:216] Starting GRPC server for 'nvidia.com/gpu'
I0820 08:26:51.682183       1 server.go:147] Starting to serve 'nvidia.com/gpu' on /var/lib/kubelet/device-plugins/nvidia-gpu.sock
I0820 08:26:51.690851       1 server.go:154] Registered device plugin for 'nvidia.com/gpu' with Kubelet

未找到GPU（尝试将nvidia修改为默认运行时）：

I0820 08:26:53.370854       1 main.go:317] Retrieving plugins.
E0820 08:26:53.370997       1 factory.go:87] Incompatible strategy detected auto
E0820 08:26:53.371004       1 factory.go:88] If this is a GPU node, did you configure the NVIDIA Container Toolkit?
E0820 08:26:53.371010       1 factory.go:89] You can check the prerequisites at: https://github.com/NVIDIA/k8s-device-plugin#prerequisites
E0820 08:26:53.371016       1 factory.go:90] You can learn how to set the runtime at: https://github.com/NVIDIA/k8s-device-plugin#quick-start
E0820 08:26:53.371021       1 factory.go:91] If this is not a GPU node, you should set up a toleration or nodeSelector to only deploy this plugin on GPU nodes
I0820 08:26:53.371029       1 main.go:346] No devices found. Waiting indefinitely.

更新运行时后，重启device-plugin daemonset

kubectl rollout restart ds -n kube-system nvidia-device-plugin-daemonset

测试

# gpu节点打标签
kubectl label nodes k8s-dell-r740-worker01 nvidia.com/gpu=true
apiVersion: v1
kind: Pod
metadata:
  name: gpu-pod
spec:
  restartPolicy: IfNotPresent
  containers:
    - name: cuda-container
      image: nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda10.2
      resources:
        limits:
          nvidia.com/gpu: 1
      command: ['nvidia-smi']
  tolerations:
  - key: nvidia.com/gpu
    operator: Exists
    effect: NoSchedule
  #nodeSelector:
  #  viadia.com/gpu: "true"

查看pod日志