Skip to main content

Using GPUs on Hikube

This guide presents the two methods of using GPUs: with virtual machines and with Kubernetes clusters.

🎯 Usage Methods​

Hikube offers two approaches for using GPUs:

  1. GPU with VM: Direct attachment of a GPU to a virtual machine
  2. GPU with Kubernetes: GPU allocation to workers for use by pods

🖥️ Method 1: GPU with Virtual Machine​

Step 1: Create the disk​

vm-disk.yaml
apiVersion: apps.cozystack.io/v1alpha1
kind: VMDisk
metadata:
  name: ubuntu-gpu-disk
spec:
  source:
    http:
      url: https://cloud-images.ubuntu.com/noble/current/noble-server-cloudimg-amd64.img
  optical: false
  storage: 50Gi
  storageClass: "replicated"

Step 2: Create the VM with GPU​

vm-gpu.yaml
apiVersion: apps.cozystack.io/v1alpha1
kind: VMInstance
metadata:
  name: vm-gpu-example
spec:
  runStrategy: Always
  instanceProfile: ubuntu
  instanceType: u1.xlarge  # 4 vCPU, 16 GB RAM
  gpus:
    - name: "nvidia.com/AD102GL_L40S"
  disks:
    - name: ubuntu-gpu-disk
  external: true
  externalMethod: PortList
  externalPorts:
    - 22
  sshKeys:
    - "ssh-rsa AAAAB3NzaC... your-public-key"
  cloudInit: |
    #cloud-config
    users:
      - name: ubuntu
        sudo: ALL=(ALL) NOPASSWD:ALL
        shell: /bin/bash
    
    package_update: true
    packages:
      - curl
      - wget
      - build-essential
    
    runcmd:
      # Install NVIDIA drivers
      - wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.0-1_all.deb
      - dpkg -i cuda-keyring_1.0-1_all.deb
      - apt-get update
      - apt-get install -y cuda-toolkit nvidia-driver-535
      - nvidia-smi -pm 1

Step 3: Deploy​

kubectl apply -f vm-disk.yaml
kubectl apply -f vm-gpu.yaml

# Check status
kubectl get vminstance vm-gpu-example

Step 4: Access and test​

# SSH access
virtctl ssh ubuntu@vm-gpu-example

# Check GPU
nvidia-smi

☸️ Method 2: GPU with Kubernetes​

Step 1: Create a cluster with GPU workers​

cluster-gpu.yaml
apiVersion: apps.cozystack.io/v1alpha1
kind: Kubernetes
metadata:
  name: cluster-gpu
spec:
  controlPlane:
    replicas: 1
  
  nodeGroups:
    # GPU workers
    gpu-nodes:
      minReplicas: 1
      maxReplicas: 3
      instanceType: "u1.xlarge"
      ephemeralStorage: 100Gi
      gpus:
        - name: "nvidia.com/AD102GL_L40S"
    
    # Standard workers (optional)
    standard-nodes:
      minReplicas: 1
      maxReplicas: 2
      instanceType: "s1.medium"
      ephemeralStorage: 50Gi

  storageClass: "replicated"

  addons:
    # Essential: installs the NVIDIA drivers and the device plugin
    # that expose nvidia.com/gpu to the pods of the tenant cluster
    gpuOperator:
      enabled: true
gpuOperator addon required

Without the gpuOperator addon enabled, the GPUs attached to the workers are not exposed to pods (nvidia.com/gpu stays at 0). It installs the NVIDIA drivers and the device plugin in the tenant cluster.

Step 2: Deploy the cluster​

kubectl apply -f cluster-gpu.yaml

# Wait for cluster to be ready
kubectl get kubernetes cluster-gpu -w

Step 3: Configure access​

# Retrieve kubeconfig
kubectl get secret cluster-gpu-admin-kubeconfig \
  -o go-template='{{ printf "%s\n" (index .data "super-admin.conf" | base64decode) }}' \
  > cluster-gpu-kubeconfig.yaml

# Use GPU cluster
export KUBECONFIG=cluster-gpu-kubeconfig.yaml
kubectl get nodes

Step 4: Deploy a GPU pod​

pod-gpu.yaml
apiVersion: v1
kind: Pod
metadata:
  name: gpu-test
spec:
  containers:
  - name: gpu-container
    image: nvidia/cuda:12.0-runtime-ubuntu20.04
    command: ["sleep", "infinity"]
    resources:
      limits:
        nvidia.com/gpu: 1
      requests:
        nvidia.com/gpu: 1
kubectl apply -f pod-gpu.yaml

# Check GPU allocation
kubectl describe pod gpu-test

# Test GPU
kubectl exec -it gpu-test -- nvidia-smi

📋 Practical Comparison​

AspectVM GPUKubernetes GPU
Setup time~5 minutes~10 minutes
ComplexitySimpleModerate
IsolationTotalPartial
FlexibilityLimitedHigh
ScalingManualAutomatic

🔧 Available GPU Types​

Configuration by usage​

# For inference/development
gpus:
  - name: "nvidia.com/AD102GL_L40S"  # 48 GB GDDR6

# For ML training (A100, two variants depending on the hardware)
gpus:
  - name: "nvidia.com/GA100_A100_PCIE_80GB"  # 80 GB HBM2e (PCIe)
  # or
  - name: "nvidia.com/GA100_A100_SXM4_80GB"  # 80 GB HBM2e (SXM4)

# For LLM/exascale computing
gpus:
  - name: "nvidia.com/GB202GL_RTX_PRO_6000_BLACKWELL_SERVER_EDITION"  # 96 GB GDDR7

✅ Post-Deployment Verification​

VM GPU​

# Check GPU
virtctl ssh ubuntu@vm-gpu-example -- nvidia-smi

# CUDA test
virtctl ssh ubuntu@vm-gpu-example -- nvcc --version

Kubernetes GPU​

# See available GPU resources
kubectl describe nodes

# Check allocation
kubectl top nodes

Cleanup​

Delete a GPU VM​

kubectl delete -f vm-gpu.yaml
kubectl delete -f vm-disk.yaml

Delete a GPU Kubernetes cluster​

kubectl delete -f cluster-gpu.yaml
warning

These actions delete the GPU resources and all associated data. These operations are irreversible.

🚀 Next Steps​

To deepen VM GPU:​

To deepen Kubernetes GPU:​

💡 Tips​

  • VM GPU: Ideal for prototyping and legacy applications
  • Kubernetes GPU: Recommended for scalable production workloads
  • Start with L40S to test before using A100 or RTX PRO 6000 (Blackwell)
  • Use replicated storage class for production

Prochaines etapes

Voir aussi