Your submission was sent successfully! Close

You have successfully unsubscribed! Close

Thank you for signing up for our newsletter!
In these regular emails you will find the latest updates about Ubuntu and upcoming events where you can meet our team.Close

Using GPU workers

Charmed Kubernetes can automatically detect GPU-enabled hardware and install required drivers from NVIDIA repositories. With the release of Charmed Kubernetes 1.29, the NVIDIA GPU Operator charm is available to simplify GPU software management and configuration.

This page describes recommended deployment and verification steps for using GPU workers with Charmed Kubernetes.

Deploying Charmed Kubernetes with GPU workers

GPU support varies depending on the underlying cloud, so you will need to determine a particular instance type for the kubernetes-worker application.

For example, when deploying to AWS, you may decide to use a p3.2xlarge instance from the available AWS GPU-enabled instance types. Similarly, you could choose an Azure Standard_NC6s_v3 instance from the available Azure GPU-enabled instance types.

NVIDIA updates its list of supported GPUs frequently, so be sure to cross-reference the GPU included in a specific cloud instance against the Supported NVIDIA GPUs and Systems documentation.

When deploying the Charmed Kubernetes bundle, you can use a YAML overlay file to ensure application instance types and configuration are optimised for use in a GPU-enabled environemnt. Broadly, the following should be noted when constructing an overlay file:

  • The GPU operator charm manages software packages and drivers on a host. Therefore, the kubernetes-control-plane application needs to be deployed in privileged mode.
  • The automatic configuration of NVIDIA software repositories should be disabled for the containerd application as this is now managed by the GPU operator charm.
  • A suitable GPU-enabled instance constraint will need to be specified for the kubernetes-worker application.

The following is an example overlay file that meets the above considerations for AWS:

# AWS gpu-overlay.yaml
      gpu_driver: none
      allow-privileged: "true"
    constraints: "instance-type=p3.2xlarge root-disk=128G"
    num_units: 1

Deploy Charmed Kubernetes with your overlay(s) like this:

juju deploy charmed-kubernetes --overlay ~/path/my-overlay.yaml --overlay ~/path/gpu-overlay.yaml

As demonstrated here, you can use multiple overlay files when deploying, so you can combine GPU support with an integrator charm or other custom configuration.

Adding GPU workers post deployment

It isn't necessary for all worker units to have GPU support. You can add GPU-enabled workers to an existing cluster. The recommended way to do this is to first set a new constraint for the kubernetes-worker application:

juju set-constraints kubernetes-worker instance-type=p3.2xlarge

Then add as many new GPU worker units as required. For example, to add two new units:

juju add-unit kubernetes-worker -n2

Adding GPU workers with GCP

Google supports GPUs slightly differently to most clouds. There are no GPU variations included in the general instance templates, and therefore they have to be added manually.

To begin, add a new machine with Juju. Include any desired constraints for cpu cores, memory, etc:

juju add-machine --constraints 'cores=4 mem=16G'

The command will return with the unit number of the machine that was created - take note of this number.

Next you will need to use the gcloud tool or the GCP console to stop the newly created instance, edit its configuration and then restart the machine.

Once it is up and running, add the kubernetes-worker application to it:

juju add-unit kubernetes-worker --to 10

...replacing 10 in the above with the previously noted number. As the charm installs, the GPU will be detected and the relevant support will be installed.

Deploying the GPU Operator

Operator charms use Kubernetes as a secondary cloud on your Juju controller. The primary cloud (e.g. AWS) hosts machine-style workloads like Charmed Kubernetes, which in turn provides the platform needed for Kubernetes operators.

Create an operator-cloud based on the Charmed Kubernetes configuration:

kubeconfig="$(juju ssh kubernetes-control-plane/leader -- cat config)"
controller="$(juju controller-config controller-name)"
echo "$kubeconfig" | \
    juju add-k8s operator-cloud --controller "$controller" --skip-storage

Switch to this cloud and create a model for our GPU operator:

juju switch operator-cloud
juju add-model operator-model operator-cloud

Now deploy the NVIDIA GPU Operator charm. As mentioned above, this charm will install requisite drivers and software packages on the worker machine, therefore we use the --trust flag to ensure the charm has host administrative privileges:

juju deploy nvidia-gpu-operator --channel 1.29/stable --trust

Deployment of this charm will typically take 15-30 minutes. Once complete, Kubernetes will be ready to run GPU optimised workloads.


As GPU instances can be costly, it is useful to test that they can actually be used. A simple test job can be created to run NVIDIA's hardware reporting tool. Please note that you may need to replace the image tag in the following YAML with the latest supported one.

This can also be downloaded here.

apiVersion: batch/v1
kind: Job
  name: nvidia-smi
      name: nvidia-smi
      restartPolicy: Never
      - image: nvidia/cuda:12.1.0-base-ubuntu22.04
        name: nvidia-smi
          - nvidia-smi
        - mountPath: /usr/bin/
          name: binaries
        - mountPath: /usr/lib/x86_64-linux-gnu
          name: libraries
      - name: binaries
          path: /usr/bin/
      - name: libraries
          path: /usr/lib/x86_64-linux-gnu

Download the file and run it with:

kubectl create -f nvidia-test.yaml

You can inspect the logs to find the hardware report.

kubectl logs job.batch/nvidia-smi

Tue Apr 11 22:46:04 2023
| NVIDIA-SMI 530.30.02              Driver Version: 530.30.02    CUDA Version: 12.1     |
| GPU  Name                  Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf            Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|   0  Tesla V100-SXM2-16GB            On | 00000000:00:1E.0 Off |                    0 |
| N/A   36C    P0               23W / 300W|      0MiB / 16384MiB |      0%      Default |
|                                         |                      |                  N/A |

| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|  No running processes found                                                           |

We appreciate your feedback on the documentation. You can edit this page or file a bug here.

See the guide to contributing or discuss these docs in our public Mattermost channel.