eOn on EESSI with EESSI-extend and eb –robot

Install eOn, and the dependencies the European Environment for Scientific Software Installations (EESSI) lacks, into a user or project prefix on top of EESSI. EasyBuild’s robot resolves the chain. Then build eOn from a working tree against the same modules for development.

The build ran on 2026-09-21 and the development build on 2026-09-25, both against EESSI 2026.06 on an AMD Zen 5 (x86_64/amd/zen5) Slurm node with EasyBuild 5.4.0 from EESSI. The robot dry run was repeated on 2026-09-25 on an Intel laptop (x86_64/intel/haswell). The easyconfigs come from the open easyconfigs pull request #26480 (eOn 2.17.10 on foss/2026.1). Until that pull request is merged, --from-pr 26480 is how they reach the robot.

What EESSI provides and what gets built

EESSI 2026.06 provides foss/2026.1 (GCC 15.2.0, OpenMPI, FlexiBLAS, FFTW), Meson, Ninja, CMake, Eigen, Rust and Python. The robot builds the rest, eleven easyconfigs at GCCcore-15.2.0 plus eOn. A dry run (eb -D with the command below) reports 88 modules already present and these 12 to build:

Module

Why EESSI lacks it

Highway/1.4.0, inih/62, fmt/12.2.0, Catch2/3.14.0

not in the 2026.06 stack

cargo-c/0.10.23

build dependency of readcon-core

robin-map/1.4.1, nanobind/2.13.0

nanobind needs robin-map

CapnProto/1.4.0, quill/11.1.0, readcon-core/0.13.1, rgpot/3.0.0

new in pull request #26480

eOn/2.17.10-foss-2026.1

new in pull request #26480

The first full run built the chain in 13 minutes 38 seconds on 32 cores. With the dependencies installed, a rebuild of eOn alone takes about 2.5 minutes.

Prerequisites

  1. EESSI mounted at /cvmfs/software.eessi.io through the CernVM File System (CVMFS). See Without root when no administrator can install the CVMFS client.

  2. A prefix for the installs. It must exist before EESSI-extend loads, or the module refuses to load:

    export EESSI_USER_INSTALL=$HOME/EESSI      # or EESSI_PROJECT_INSTALL=/project/...
    mkdir -p "$EESSI_USER_INSTALL"
    
  3. Optionally, a shallow clone of the easyconfigs develop branch as an extra robot path, for recipes newer than the EasyBuild release in EESSI:

    scripts/fetch-upstream-trees.sh
    set -a
    source "${EB_UPSTREAM_CACHE:-$HOME/.cache/eb-stack/upstream}/env.sh"
    set +a
    # "$EB_EASYCONFIGS_DEVELOP" is the pinned easyconfigs checkout
    

Build eOn with the robot

eb-local eb sources EESSI’s init, loads EESSI-extend, isolates Cargo from host wrappers, and execs eb. On this checkout the build runs through the local Podman target, which invokes eb --robot on the planned eOn bundle. The bundle path has to sit under /tmp/eb-stack, because that is the directory the target mounts.

podman build -t eb-stack-rocky9 \
  -f skills/new-package/container/rocky9/Containerfile \
  skills/new-package/container/rocky9

eb-stack campaign run \
  --bundle /tmp/eb-stack/bundles/eon-tutorial \
  --config examples/targets/local-podman.toml \
  --target local-rocky9 \
  --state /tmp/eb-stack/bundles/eon-tutorial.campaign.json

Plan that bundle with the commands in the tutorial before this run. campaign run stays in the foreground and starts eb --robot inside the image.

What each option is for:

  • --from-pr 26480 puts the pull request’s easyconfigs first on the robot path. Drop it once the recipes are in an EasyBuild release.

  • --robot-paths=<develop tree>: with the trailing colon appends the default robot path, so EESSI’s own easyconfigs still resolve. -r turns the robot on.

  • --umask=022 matters for user installs: EESSI-extend sets umask 077 and the sticky bit when EESSI_USER_INSTALL is set. That is fine for a private prefix, but see Without root for the case where it breaks module loads.

  • --dump-test-report writes the report EasyBuild would post to the pull request.

The test report’s summary on success:

Build succeeded for 1 out of 1 (total: 2 mins 27 secs)
 * SUCCESS eOn-2.17.10-foss-2026.1.eb

Use the installed eOn

Read the campaign while it runs, from another shell:

eb-stack campaign status --state /tmp/eb-stack/bundles/eon-tutorial.campaign.json

The campaign installs each attempt into its own prefix, <work_root>/install-attempts/eOn-2.17.10/attempt-<N>, and passes it to eb as --installpath, which takes precedence over the path EESSI-extend derives from $EESSI_USER_INSTALL. The modules land under modules/all in that prefix, and the verification commands find them there. Unlike the $EESSI_USER_INSTALL tree, which has one directory per $EESSI_SOFTWARE_SUBDIR, that path does not name the CPU target.

Develop eOn from a working tree

The installed modules are a development environment as well. eOn builds with Meson, and its develop branch needs Meson 1.8 or newer, rgpot 3.2 or newer and readcon-core 0.14.9 or newer. That is newer than the rgpot 3.0.0 and readcon-core 0.13.1 in pull request #26480, so Meson takes those two from its wraps (the compute node needs network for that) and everything else comes from EESSI and the prefix.

This ran on 2026-09-25 as a 16-core Slurm job on EESSI 2026.06, with develop at 1916f9a9: meson test --suite eon reported 43 passed and 0 failed. The build directory was outside $HOME, CC, CXX, and FC named the EESSI compilers, and CARGO_HOME was a fresh directory. Iterate with meson compile -C $WORK/build after each edit. Nothing here installs into the EESSI prefix, so the code under development lives in the working tree alone. module use on the prefix does what module load EESSI-extend does for MODULEPATH, without the EasyBuild settings that module also sets.

Three host settings broke earlier attempts, each with its own error:

Error

Cause

Fix

ModuleNotFoundError: No module named 'mesonbuild'

the compat layer’s usr/bin was prepended to PATH; its python3 ignores EBPYTHONPREFIXES, so Meson’s shebang found a Python without Meson

leave the compat layer off the front of PATH; EESSI’s GCC finds its linker without it

gcc: error: unrecognized command-line option '-fuse-ld=/usr/bin/mold'

Cargo reads .cargo/config.toml from every parent of its working directory, so a build under $HOME picked up the host’s linker settings

build outside $HOME and set CARGO_HOME to a fresh directory

compilers run through the host’s sccache

Meson wraps cc in a ccache it finds on PATH

name the compilers in CC, CXX and FC

On EESSI 2025.06, the one Meson 1.8 module sits at GCCcore-14.3.0, which means foss/2025b rather than foss/2025a.

Without root

Where no administrator installs the CVMFS client, an unprivileged user runs cvmfs2 against a private cache with -o libfuse=2, then bind-mounts that directory at the canonical path inside a user namespace. This checkout has /dev/fuse and no cvmfs2 package in its apt repositories, so the build above uses the local Podman target instead of that mount.

The libfuse 2 path matters on some hosts: CVMFS 2.13.3’s libfuse 3 path served directory listings but never answered a file read against fuse3 3.18.2, while the libfuse 2 path answered at once. Inside the namespace the process is uid 0, which has four consequences for EasyBuild:

Symptom

Cause

Fix

EasyBuild refuses to start

it will not run as root

--allow-use-as-root-and-accept-consequences; files still belong to the real user

sanity check: fake module unknown

Lmod as uid 0 skips a module directory without read and search for others; EESSI-extend’s umask 077 makes it 0700

--umask=022

tar: Cannot change ownership to uid 1000

the namespace maps uid 0 alone, so restoring archive owners fails (cargo-c, CapnProto)

export TAR_OPTIONS=--no-same-owner

every run rebuilds everything

a prefix first created under umask 077 hides all later modules from Lmod

once, as the real user: find $EESSI_USER_INSTALL -type d ! -perm -o=rx -exec chmod o+rx {} +

A cancelled eb run also leaves its lock under software/.locks/ in the prefix; remove it before the next run.

After pull request #26480 is merged

EESSI’s software-layer easystacks pin recipes that are not in an EasyBuild release with from-commit and a comment naming the pull request. The entries for easystacks/software.eessi.io/2026.06/eessi-2026.06-eb-5.4.0-2026.1.yml:

- Highway-1.4.0-GCCcore-15.2.0.eb
- inih-62-GCCcore-15.2.0.eb
- fmt-12.2.0-GCCcore-15.2.0.eb
- cargo-c-0.10.23-GCCcore-15.2.0.eb
- Catch2-3.14.0-GCCcore-15.2.0.eb
- robin-map-1.4.1-GCCcore-15.2.0.eb
- nanobind-2.13.0-GCCcore-15.2.0.eb
- eOn-2.17.10-foss-2026.1.eb:
    options:
      # see https://github.com/easybuilders/easybuild-easyconfigs/pull/26480
      from-commit: <merge commit of #26480>

The four other new recipes (CapnProto, quill, readcon-core, rgpot) carry the same from-commit block. The EESSI build bot then builds the easystack for each CPU target, and the prefix install above is no longer needed.