Skip to content

Instantly share code, notes, and snippets.

@singlecheeze
Last active November 16, 2022 19:25
Show Gist options
  • Save singlecheeze/d9b1f0b02b650e4499ac5e72937d7256 to your computer and use it in GitHub Desktop.
Save singlecheeze/d9b1f0b02b650e4499ac5e72937d7256 to your computer and use it in GitHub Desktop.
Nvidia MPS Device Plugin on OCP.

Disable Default Nvidia Device Plugin in Nvidia GPU Operator ClusterPolicy:

devicePlugin:
    config:
      default: ''
      name: ''
    enabled: false
    imagePullPolicy: Always

Label nodes for alternate device plugin:

oc label node r730ocp3.localdomain k8s.kuartis.com/accelerator=vgpu

Download the alternative device plugin (https://github.com/kuartis/kuartis-virtual-gpu-device-plugin):

git clone https://github.com/kuartis/kuartis-virtual-gpu-device-plugin.git

Edit the device-plugin.yml manifest to get the latest (as of now) CUDA image, this should match CUDA being called from workload pods and Nvidia Operator. Note:

image: nvidia/cuda:11.8.0-runtime-ubuntu22.04
--vgpu=<number_of_virtual_gpu_one_pyhsical_gpu_can_have> # Default is 10, Max 48
--allowmultigpu=<true|false> # Default is false. Prevents vGPU resources that one container can have to fall on different physical gpus.
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: kuartis-virtual-gpu-device-plugin-daemonset
  namespace: kube-system
spec:
  selector:
    matchLabels:
      app.kubernetes.io/name: kuartis-virtual-gpu-device-plugin
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      labels:
        app.kubernetes.io/name: kuartis-virtual-gpu-device-plugin
    spec:
      hostIPC: true
      nodeSelector:
        k8s.kuartis.com/accelerator: vgpu
      tolerations:
        - key: k8s.kuartis.com/vgpu
          operator: Exists
          effect: NoSchedule
      priorityClassName: "system-node-critical"
      initContainers:
        - name: set-compute-mode
          image: nvidia/cuda:11.8.0-runtime-ubuntu22.04
          command: ["nvidia-smi", "-c", "EXCLUSIVE_PROCESS"]
          securityContext:
            capabilities:
              add: ["SYS_ADMIN"]
      containers:
        - image: ghcr.io/kuartis/kuartis-virtual-gpu-device-plugin:0.7.0
          name: kuartis-virtual-gpu-device-plugin-ctr
          command:
            - /usr/bin/virtual-gpu-device-plugin
          args:
            - --vgpu=1
            - --allowmultigpu=false
          securityContext:
            allowPrivilegeEscalation: false
            capabilities:
              drop: ["ALL"]
          env:
            - name: NODE_NAME
              valueFrom:
                fieldRef:
                  fieldPath: spec.nodeName
          ports:
            - containerPort: 8080
              name: metrics
          volumeMounts:
            - name: device-plugin
              mountPath: /var/lib/kubelet/device-plugins
            - name: proc
              mountPath: /host/proc
              readOnly: true
            - mountPath: /var/run/dockershim.sock
              name: dockershimsock
        - image: nvidia/mps
          name: mps
          volumeMounts:
            - name: nvidia-mps
              mountPath: /tmp/nvidia-mps
      volumes:
        - name: device-plugin
          hostPath:
            path: /var/lib/kubelet/device-plugins
        - name: nvidia-mps
          hostPath:
            path: /tmp/nvidia-mps
        - name: proc
          hostPath:
            path: /proc
        - name: dockershimsock
          hostPath:
            path: /var/run/dockershim.sock
---
apiVersion: v1
kind: Service
metadata:
  labels:
    app.kubernetes.io/name: kuartis-virtual-gpu-device-plugin
  name: kuartis-virtual-gpu-device-plugin
  namespace: kube-system
spec:
  ports:
    - name: metrics
      port: 8080
      targetPort: 8080
      protocol: TCP
  selector:
    app.kubernetes.io/name: kuartis-virtual-gpu-device-plugin
  sessionAffinity: None
  type: ClusterIP
---
apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  labels:
    app.kubernetes.io/name: kuartis-virtual-gpu-device-plugin
  name: kuartis-virtual-gpu-device-plugin
  namespace: kube-system
spec:
  endpoints:
    - interval: 15s
      path: /metrics
      port: metrics
  namespaceSelector:
    matchNames:
      - kube-system
  selector:
    matchLabels:
      app.kubernetes.io/name: kuartis-virtual-gpu-device-plugin

Apply manifest to cluster:


Troubleshooting: image

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment