GNU Guix is a transactional package manager, with support for per-user package installations. Users can install their own packages without interfering with each other, yet without unnecessarily increasing disk usage or rebuilding every package. Users can in fact create as many software environments as they like—think of it as VirtualEnv but not limited to Python, or modules but not limited to your sysadmin-provided modules.

The software environments created with Guix are fully reproducible: a package built from a specific Guix commit on your laptop will be exactly the same as the one built on the HPC cluster you deploy it to, usually bit-for-bit.

We believe this makes Guix a great foundation for reproducible software deployment in high-performance computing (HPC). Here’s how to get started.

Installing Guix

You can install Guix on your laptop in 5 minutes: just follow the binary install instructions.

You’re a cluster sysadmin and would like to have a cluster-wide install? Check out this guide.

Installing Packages

Say you’re searching for a sparse solver among the 50,000+ packages that come with Guix and with third-party channels:

$ guix search sparse solver
name: mumps
version: 5.5.1
outputs:
+ out: everything
systems: x86_64-linux i686-linux
dependencies: gfortran@10.3.0 metis@5.1.0 openblas@0.3.20
+ scotch@7.0.1
location: gnu/packages/maths.scm:3848:2
homepage: http://mumps.enseeiht.fr
license: CeCILL-C
synopsis: Multifrontal sparse direct solver
description: MUMPS (MUltifrontal Massively Parallel sparse
+ direct Solver) solves a sparse system of linear equations
+ A x = b using Guassian elimination.
relevance: 12

…

To install it along with the latest GNU compiler tool chain:

$ guix install mumps gcc-toolchain
The following packages will be installed:
   mumps         5.5.1
   gcc-toolchain 12.2.0

The following derivation will be built:
   /gnu/store/kipa9k61zkhw4s3frs92w683ps23hpjj-profile.drv
3.1 MB will be downloaded

…

hint: Consider setting the necessary environment variables by running:

     GUIX_PROFILE="$HOME/.guix-profile"
     . "$GUIX_PROFILE/etc/profile"

Alternately, see `guix package --search-paths -p "$HOME/.guix-profile"'.

Spawning One-Off Environments

Sometimes all you want is to try out a program without installing it in your profile. That’s where guix shell comes in. To create an environment containing Python, NumPy, and Scikit-learn, run:

$ python3
bash: python3: Command not found
$ guix shell python python-numpy python-scikit-learn
The following derivation will be built:
   /gnu/store/2g3mj1xdlq2rj8j0crl4sa68bqhmfsmd-profile.drv
building directory of Info manuals...
building database for manual pages...
[env]$ python3
Python 3.9.9 (main, Jan  1 1970, 00:00:01)
[GCC 10.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy
>>> import sklearn
>>>

Customizing Packages

Occasionally you’ll want to customize the way packages are built. From the command line, you can apply transformations, such as replacing one dependency with another one in the dependency graph. The example below replaces openmpi with openmpi-thread-multiple in the dependency graph of mumps-openmpi:

$ guix install mumps-openmpi \
     --with-input=openmpi=openmpi-thread-multiple

The expressivity of the command line is limited, but you can go further by writing your own package definitions.

Defining Packages

To add a package, provide a package definition, which looks like this:

(define-public scalapack
  (package
    (name "scalapack")
    (version "2.0.2")
    (source
     (origin
       (method url-fetch)
       (uri (string-append "http://www.netlib.org/scalapack/scalapack-"
                           version ".tgz"))
       (sha256
        (base32
         "0p1r61ss1fq0bs8ynnx7xq4wwsdvs32ljvwjnx6yxr8gd6pawx0c"))))
    (build-system cmake-build-system)
    (inputs
     (list openmpi gfortran lapack))
    (arguments
     `(#:configure-flags `("-DBUILD_SHARED_LIBS:BOOL=YES")))
    (home-page "http://www.netlib.org/scalapack/")
    (synopsis "Library for scalable numerical linear algebra")
    (description
     "ScaLAPACK is a Fortran 90 library of high-performance linear algebra
routines on parallel distributed memory machines.  ScaLAPACK solves dense and
banded linear systems, least squares problems, eigenvalue problems, and
singular value problems.")
    (license (license:bsd-style "file://LICENSE"
                                "See LICENSE in the distribution."))))

Generate it from a third-party repository using guix import, or write it with the help of the Guix Packager web interface.

Your own package collection can be published as a channel.

Sending Packages to Guix-less Machines

What if the target supercomputer lacks Guix? You can still enjoy Guix’s reproducibility and customizability by sending your package binaries there, leveraging relocatable binaries:

laptop$ scp `guix pack -RR hwloc -S /bin=bin` supercomputer:hwloc.tar.gz
…
supercomputer$ mkdir -p ~/.local
supercomputer$ (cd ~/.local; tar xf ~/hwloc.tar.gz)
supercomputer$ ~/.local/bin/lstopo

Other options include building Singularity or Docker images.

Learning More

Find the main commands in the quick reference card. Learn more in the reference manual: Deutsch | English | español | français.

Joining

Read about on-going Guix-HPC developments on our blog.

Guix-HPC and GNU Guix are collaborative efforts. You are welcome to join!

  • MDC
  • Inria
  • UBC
  • UTHSC