env-installer

Spack environments for the software stacks of NumPEx applications together with everything needed to install them on HPC platforms:

• ready-to-use Spack environments for each software stack (projects/);
• a unified installer, install.sh, which builds a stack on a supported HPC platform, combining the stack definition with site-specific configuration (externals, compilers, package repositories);
• an OCI binary cache on ghcr.io/numpex/spack-stack, continuously filled by CI, so that matching packages are downloaded instead of rebuilt from source.

Install a stack on an HPC platform

install.sh combines three inputs into a Spack environment named <project>-<arch>--<version> (version from version.txt):

Input	Option	Files	Content
Platform	--platform	platforms/<site>/config.sh	Spack location, parallel jobs, Slurm settings
Architecture	--arch	platforms/<site>/<arch>/spack/*.yaml	externals, package preferences, package repositories
Project	--project	projects/<project>/spack.yaml	the software stack to install

Example on CINES Adastra:

$ ./install.sh --platform cines/adastra --arch mi300 --project <project>

By default it runs the following steps in order (run only one with --step <step>):

1. spack — clone Spack (release pinned in spack-version.txt) and bootstrap it (needs internet)
2. env — (re)create the environment, update the package repositories, optionally detect system packages (needs internet)
3. concretize — concretize and report the lockfile
4. mirror — download all sources to a local mirror (Slurm mode only)
5. install — build and install

On sites where compute nodes have no internet access (e.g. Jean-Zay), enable Slurm mode with --slurm-account <account> (or the SLURM_ACCOUNT variable): sources are mirrored from the login node, and concretize/install run on GPU compute nodes through srun.

Concurrent installations of the same environment are prevented by a lock file, and on completion the installation tree is shared with the site group.

Supported platforms

Site	Machine	Architectures
CINES	Adastra	genoa (CPU), mi250, mi300 (AMD GPU)
IDRIS	Jean-Zay	h100 (NVIDIA GPU, Slurm mode)
MDLS	Mandelbrot	rtx6000 (NVIDIA GPU)

The Adastra and Jean-Zay architectures also ship a toolchain.cmake with the recommended compiler settings, reusable downstream with cmake -C <toolchain.cmake>.