Guix@GRICAD

Software packaging and deployment at GriCAD

Exa-DI Annuel Meeting - 24/02/2026

P.-A. Bouttier

TOC

  1. A bit of history
  2. Guix for users
  3. Guix for operators

TOC

  1. A bit of history
  2. Guix for users
  3. Guix for operators

GRICAD supercomputing context

  • 4 compute clusters in-house (over 25 000 CPU cores, approximately 100 GPUs)
  • ...connected to each other and to lab clusters via a local computing grid
  • Uses:
    • HPC, HTC
    • Data processing
    • Visualization, training & development
  • High heterogeneity in use cases, communities, user skill levels, and consequently, software stacks
  • The challenge: finding a software environment management tool that meets very heterogeneous constraints

Open Science context

  • Open Science: How to ensure reproducibility of digital processing?

    • Good software development practices
    • And the software environment?

  • Practice changes

    • Increasing use of non-HPC platforms: JupyterHub, BinderHub, cloud computing...
    • ...within the same processing chain
    • Development of distributed compute/data platforms

The Good Ol' days

  • Until 2015, use of the module system

    • Classic, well-known
    • Used at tier-1 and tier-2
    • Usages and communities (many) were more homogeneous

  • But issues started to surface:

    • Very average portability (essential for our small grid!)
    • Duplication of effort proportional to the number of clusters
    • Marginal community aspect (e.g. no direct sharing of packages, binaries)
    • ...

The utopia of software environments?

  • Isolation from the host system, at execution AND during construction
  • Maintenance (well-managed dependency tree), reproducibility (unique package version - src, compile, desc, defs,... - same output everywhere), portability
  • Completely functional in user space
  • Automated workflow: custom packages, CI/CD and automatic rebuilds, from personal PC to clusters
  • OS-independent

Welcome to the Jungle

center

Our final choices

  • Installation and deployment of Nix in 2015, Guix in 2018
  • Today, Nix and Guix are available, with some modules for certain users

TOC

  1. A bit of history
  2. Guix for users
  3. Guix for operators

Guix on the user side, Pros and Cons (1/2)

Cons

  • A new tool: requires support (documentation, assistance)
  • Support for changes in practices (e.g. detaching from version numbers)
  • Some important software stacks still need to be packaged: Intel OneApi, AI frameworks => Nix to the rescue

Guix on the user side, Pros and Cons (2/2)

Pros

  • Reproducible and therefore truly portable (unlike Spack or Conda)
  • Very practical for development environments (like venv, but for everything)
  • Easy-to-use Guix commands
  • Large package base (between 20,000 and 50,000 if you add specific channels)
  • User autonomous (if package exists)

TOC

  1. A bit of history
  2. Guix for users
  3. Guix for operators

Guix on the support team side, Pros and Cons (1/2)

Cons

  • Steep learning curve for packaging
  • Some software, especially proprietary software, is instrinsicly difficult to package

Guix on the support team side, Pros and Cons (2/2)

Pros

  • Greater user autonomy
  • A much more portable deployment solution
  • Community aspect:
    • sharing packages, recipes, problem-solving
    • Shared documentation
    • Support for support (mailing lists, dedicated Mattermost, CafĂ© Guix, Guix-HPC days, etc.)
    • Much more exciting!

Thank you all for your attention!