Using NVIDIA GPUs with minikube

Using the docker driver

• Ensure you have an NVIDIA driver installed, you can check if one is installed by running nvidia-smi, if one is not installed follow the NVIDIA Driver Installation Guide

• Check if bpf_jit_harden is set to 0

  sudo sysctl net.core.bpf_jit_harden
  

  • If it’s not 0 run:

  echo "net.core.bpf_jit_harden=0" | sudo tee -a /etc/sysctl.conf
  sudo sysctl -p
  

• Install the NVIDIA Container Toolkit on your host machine

• Configure Docker:

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

• Start minikube:

  minikube start --driver docker --container-runtime docker --gpus all
  

Using the ’none’ driver

NOTE: This approach used to expose GPUs here is different than the approach used to expose GPUs with --driver=kvm. Please don’t mix these instructions.

• Install minikube.

• Install the nvidia driver, nvidia-docker and configure docker with nvidia as the default runtime. See instructions at https://github.com/NVIDIA/nvidia-docker

• Start minikube:

  minikube start --driver=none --apiserver-ips 127.0.0.1 --apiserver-name localhost
  

• Install NVIDIA’s device plugin:

  minikube addons enable nvidia-device-plugin
  

Using the kvm driver

When using NVIDIA GPUs with the kvm driver, we passthrough spare GPUs on the host to the minikube VM. Doing so has a few prerequisites:

• You must install the kvm driver If you already had this installed make sure that you fetch the latest docker-machine-driver-kvm binary that has GPU support.

• Your CPU must support IOMMU. Different vendors have different names for this technology. Intel calls it Intel VT-d. AMD calls it AMD-Vi. Your motherboard must also support IOMMU.

• You must enable IOMMU in the kernel: add intel_iommu=on or amd_iommu=on (depending to your CPU vendor) to the kernel command line. Also add iommu=pt to the kernel command line.

• You must have spare GPUs that are not used on the host and can be passthrough to the VM. These GPUs must not be controlled by the nvidia/nouveau driver. You can ensure this by either not loading the nvidia/nouveau driver on the host at all or assigning the spare GPU devices to stub kernel modules like vfio-pci or pci-stub at boot time. You can do that by adding the vendorId:deviceId of your spare GPU to the kernel command line. For ex. for Quadro M4000 add pci-stub.ids=10de:13f1 to the kernel command line. Note that you will have to do this for all GPUs you want to passthrough to the VM and all other devices that are in the IOMMU group of these GPUs.

• Once you reboot the system after doing the above, you should be ready to use GPUs with kvm. Run the following command to start minikube:

  minikube start --driver kvm --kvm-gpu
  

  This command will check if all the above conditions are satisfied and passthrough spare GPUs found on the host to the VM.

  If this succeeded, run the following commands:

  minikube addons enable nvidia-gpu-device-plugin
  minikube addons enable nvidia-driver-installer
  

  This will install the NVIDIA driver (that works for GeForce/Quadro cards) on the VM.

• If everything succeeded, you should be able to see nvidia.com/gpu in the capacity:

  kubectl get nodes -ojson | jq .items[].status.capacity
  

Where can I learn more about GPU passthrough?

See the excellent documentation at https://wiki.archlinux.org/index.php/PCI_passthrough_via_OVMF

Why are so many manual steps required to use GPUs with kvm on minikube?

These steps require elevated privileges which minikube doesn’t run with and they are disruptive to the host, so we decided to not do them automatically.