GPU Acceleration¶

GPU acceleration can be utilised in cluster worker nodes with little additional effort. This document describes the details for using GPU acceleration in Kubernetes.

For additional documentation on GPU, see GPU Support in Virtual Servers.

GPU support is available beginning from the following CCKS cluster versions:

v1.31.11
v1.32.7
v1.33.3

Note

C2-GPU flavors are not currently supported in CCKS. If a worker node is deployed using a C2-GPU flavor, the nodes will start but GPU acceleration will be unavailable.

Creating a Cluster¶

For the most part this is the same as deploying a standard Kubernetes cluster, except that a suitable GPU flavor needs to be provided for the worker nodes. The following example deploys a simple cluster with two GPU-backed worker nodes using the c3-gpu.c24r96g1 flavor:

openstack coe cluster create cluster1 \
--cluster-template kubernetes-v1.33.3 \
--master-count 3 \
--node-count 2 \
--flavor c3-gpu.c24r96g1

Once the cluster is created, monitor its status until it reports healthy:

$ openstack coe cluster show cluster1 -f value -c health_status
HEALTHY

Download the Kubernetes config file for the cluster from the Catalyst Cloud dashboard.

Run the NVIDIA CUDA Helm chart to install CUDA support. Note that the version used can be quite particular; the current version tested by Catalyst Cloud is v24.6.2.

helm install --wait --generate-name -n gpu-operator --create-namespace nvidia/gpu-operator --version=v24.6.2 --set driver.enabled=false --set toolkit.enabled=true

At this point the cluster is ready to run GPU accelerated applications.

Testing GPU Acceleration¶

A simple method for testing GPU acceleration is to use NVIDIA’s CUDA sample apps.

For example, run the following to create a pod running vectorAdd:

$ kubectl create -f - << EOF
apiVersion: v1
kind: Pod
metadata:
  name: cuda-vectoradd
spec:
  restartPolicy: OnFailure
  containers:
  - name: cuda-vectoradd
    image: "nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda12.5.0-ubi8"
    resources:
      limits:
        nvidia.com/gpu: 1
EOF

After a minute or two the cuda-vectoradd pod logs shdould show a successful result:

$ kubectl logs cuda-vectoradd

[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory
Test PASSED
Done