README.md

NVIDIA GPU Smoke tests

The image created by the Dockerfile in this folder compiles a few CUDA samples for both x86_64 and aarch64. This container image can be used as a TestSys workload test to validate the Bottlerocket *-nvidia variants running on NVIDIA hosts.

Build requirements

Due to missing packages needed for cross-compiling, and unexpected handling in the sample Makefile logic when compiling for one platform for another, the build for this image requires two Buildx hosts - one for each architecture.

In order to build the image for both architectures, you need to set up the buildx docker plugin. This guide describes the installation process.

By default, the install of docker buildx uses the docker driver type. This driver is not able to build for all platform targets we need. You will need to run the following to create a new builder using the docker-container driver:

docker buildx create --use --bootstrap
docker buildx ls

You should see a * next to your newly created builder denoting it as the currently active builder to use. If it is not, you can switch builder contexts by running:

docker buildx use $BUILDER_NAME

You then need to add another "context" to the builder for a second host that can build the other platform architecture. If you are running these steps on an amd64 host, you will need to add an arm64 host to the builder, or vice versa.

First, verify you are able to access the remote host via SSH:

docker -H ssh://user@hostname info

The output from that command should show information about the Docker instance running on the remote host. You can then add that host to your Buildx builder by running:

docker buildx create --name $BUILDER_NAME --append ssh://user@hostname

You must bootstrap the new builder.

docker buildx inspect --bootstrap --builder $BUILDER_NAME

Supported platforms can be verified by running:

docker buildx inspect | grep Platforms

The output from this command should show both linux/amd64 and linux/arm64 platforms.

By default, the image will be tagged <PUBLISH_IMAGES_REGISTRY>/gpu-tests:<TAG>, so make sure you already have a gpu-tests repository in your registry. You can change the name of the repo by overriding the PUBLISH_IMAGES_REGISTRY env variable while building the image.

NOTE: This assumes that you have already configured your credentials to be able to perform a docker push to your registry.

Building the image

To build the image for both x86_64 and aarch64, run the following:

PUBLISH_IMAGES_REGISTRY=<YOUR_REGISTRY> make

The command will build, tag, and push the images to your PUBLISH_IMAGES_REGISTRY using the name gpu-tests:<TAG>.

Since this is a multi-arch image, you can use it in both x86_64 and aarch64 clusters.