The LRZ AI Systems use the Enroot container framework provided by NVIDIA, to run jobs in isolated environments. Enroot allows users to define custom software stacks and operates entirely in user space.

No root privileges are needed to build or run containers. This approach ensures flexibility while maintaining system security and performance. Enroot supports various container image formats, including Docker images.

This means you can use any standard Docker image, such as those from Docker Hub or the NVIDIA NGC Cloud. Using NVIDIA NGC containers requires registration. For details, see section 4.2 Enroot - Images from Nvidia NGC.

Please note:

Enroot is not available on the SSH login nodes of the LRZ AI Systems. To use Enroot on the compute nodes, please start an interactive session, as described in 5.1 Slurm Interactive Jobs.

Tools like module, conda, and pip are not officially supported on the LRZ AI Systems. The only supported method for installing custom software is through the use of containers with Enroot.

The Enroot Workflow

Running containerized applications with the Enroot framework follows a slightly different workflow compared to other container technologies. We recommend reviewing the Enroot documentation for a deeper understanding ( see here). However, in most common use cases, you’ll only need three commands:

enroot import
enroot create
enroot start

Other frequently used commands include enroot export, enroot list, and enroot remove. Since the Enroot container runtime is not available on the SSH login nodes of the LRZ AI Systems, please start an interactive session, as described in 5.1 Slurm Interactive Jobs, before proceeding with any of the steps below.

Importing an Enroot Container Image

To use a container image from a registry (e.g., Docker or NVIDIA NGC), start by importing it with the enroot import command. This converts the image into an Enroot-compatible .sqsh file, which can be reused to launch multiple containers.

enroot import docker://ubuntu

Creating an Enroot Container

Once you have an Enroot image, use the enroot create to create the container. The command will expand the container image into a local container filesystem. Each container image requires a separate run of this command. To assign custom names to containers, see the official documentation. The following command will create the container, which will inherit the image name (without the .sqsh extension). You can assign a custom name to a container by using the --name CUSTOM_NAME option after the enroot create command.

enroot create ubuntu.sqsh

Start an Enroot Container

Once you have an Enroot container, you can run applications within its defined software environment using the enroot start command.
You can either execute a predefined hook script inside the container (refer to the container image documentation for its location)
or directly specify the executable as a command-line argument.

enroot start ubuntu

Install Software inside an Enroot Container

If you need to run a command as root inside the container, for example to install software, you can use the --root option. Note that this grants root privileges only within the container environment, not on the host system.

enroot start --root ubuntu

With root privileges inside the container you can now install your software stack. The software is installed on the writable layer of the container. For example, the following command installs essential tools to build Python packages like xformers or faiss from source.

apt update && apt install -y build-essential cmake

Store an Enroot Container

After exiting, you can list your containers with enroot list. To make the changes to container persistent, it needs to be exported into a container image (.sqsh file).

enroot export -o ubuntu-build.sqsh ubuntu

Finally, remove not needed containers with enroot remove -f ubuntu.

This is an educational example. To run your AI/ML workload efficiently use the highly optimized docker containers provided directly from Nvidia NGC.

4. Enroot - Prev | Next - 4.2 Enroot - Images from Nvidia NGC