Skip to content

GPU in Rahti


This version of Rahti is deprecated. The projects need to be migrated from Rahti 1 platform to the new Rahti 2 production platform by June 2024. Please consult the migration guide for more information.

Kubernetes and by extension Rahti 1 do not support GPUs and other external devices natively. Instead, they allow access to such extended resources through custom device plugins which are responsible for handling the advertisement and allocation setup of these extended resources.

The device plugin used in Rahti 1 handles Nvidia GPU with the resource name More on how it is used in practice later.

Currently, Rahti 1 has a very limited amount of A100/P100 GPUs. For this reason, your Rahti 1 projects will not have access to GPU by default. This means you will need to request a GPU quota by sending a request to stating the Rahti 1 project/namespace you need the quota for.

The Rahti 1 GPU comes with an Nvidia driver which includes the low-level shared libraries and headers needed to access/program the GPU device as well as the Nvidia device monitoring tools such as Nvidia -smi which are useful for gathering information on the usage of GPUs. These libraries and tools will be mounted and be made available to the pod automatically by the system during its creation.

Running GPU-enabled pods

As a first example let's create a pod which runs nvidia-smi command line utility that is designed to be used for managing and monitoring NVIDIA GPU devices. This is the simplest example that can be used to demonstrate the availability of GPU in your Rahti 1 project.

To run the example you need to login to Rahti 1 and switch to the project that you have an approved GPU quota as:

oc login --token=<your-token>
oc project <your-project>
oc describe limitranges core-resource-limits #to check your assigned GPU qouta

Then create a file with the following content:

# gpu-example-1.yaml
apiVersion: v1
kind: Pod
  name: gpu-example-1
  restartPolicy: Never
    - name: gpu-example-1
      image: "nvidia/cuda:11.6.0-base-ubi8"
      command: ["nvidia-smi"]

And run the following command to start the pod:

oc apply -f gpu-example-1.yaml

Finally check the output of the execution from the pod (you might need to wait for a few seconds for the pod creation to complete) and clean up after using:

oc logs gpu-example-1
oc delete pod gpu-example-1

As second example you can create a pod running Cuda based vector addition to demonstrate access to GPU through a lower-level programming model/interface.

To run the second example create a pod definition file as:

# gpu-example-2.yaml
apiVersion: v1
kind: Pod
  name: gpu-example-2
  restartPolicy: OnFailure
    - name: gpu-example-2
      image: ""
 1 # requesting 1 GPU

And run, check the log (you might need to wait for a few seconds for the task to complete), and clean up after as:

oc apply -f gpu-example-2.yaml
oc logs gpu-example-2
oc delete pod gpu-example-2

As a final example, we show how a more complete system that uses GPU resources can be deployed on Rahti 1. The example provides an API for training and deploying a text sentiment classification model. To facilitate easier deployment, we use various Kubernetes resource definitions and their explanation can be found in the general Kubernetes basics page. One of the main things to notice in this example is the Dockerfile that packages the application. The Dockerfile uses an appropriate base image (FROM tensorflow/tensorflow:2.7.0-gpu) which contains the right runtime CUDA libraries for utilizing the GPU resource. This image can be built locally and be pushed to an image repository that can be used during the deployment of the system. ABuildConfig Openshift resource can also be used as is the case in this example.

To run the example use the following command:

oc process -f
| oc create -f -

You can then follow the deployment of the application and its components from the Rahti 1 console or using oc status, and interact with the application using the UI at /docs. The route your application has been deployed to can be found from the Rahti 1 console or using oc get routes.

After exploring and testing the example deployment it should be cleaned up as:

oc delete all -l app=rahti-ml-examples


At the moment, it is not possible to request a fraction of GPU. At least, a single or more GPU has to be requested. Also, keep in mind that the GPU resources are scarce and at times your GPU request might stay in a pending state until resources become available system-wide.

Last update: December 11, 2023