Reload a mismatched stack

This tutorial runs the guard from a checkout. The prefix, eessi_archdetect.sh, and LMOD_CMD are the stubs in tests/. The guard itself is src/eessi_arch_guard.py. You need Python 3.9 or newer and bash.

The loaded stack is EESSI 2025.06 plus R/4.4.2, recorded for x86_64/amd/zen4. The fake cpuinfo is a zen2 node, and both CPU trees exist in the prefix. $PREFIX in the transcript is the temporary directory the script created.

Check the version

$ python3 src/eessi_arch_guard.py --version
0.1.1

Run the guard on the mismatched stack

$ python3 docs/examples/walkthrough.py mismatch

The script prints:

case: mismatch
loaded: x86_64/amd/zen4
node cpuinfo: zen2
command: python3 -I src/eessi_arch_guard.py
exit: 0
--- stdout ---
export EBROOTR=$PREFIX/prefix/software/linux/x86_64/amd/zen2/software/R/4.4.2
export EESSI_ACCEL_SUBDIR=
export EESSI_ARCH_GUARD_RELOADED=x86_64/amd/zen4->x86_64/amd/zen2
export EESSI_SOFTWARE_SUBDIR=x86_64/amd/zen2
export MODULEPATH=$PREFIX/prefix/software/linux/x86_64/amd/zen2/modules/all
export PATH=$PREFIX/prefix/software/linux/x86_64/amd/zen2/software/R/4.4.2/bin:$PREFIX/prefix/software/linux/x86_64/amd/zen2/bin:$PREFIX/bin:/usr/bin:/bin
export _LMFILES_=
export _ModuleTable001_=table-for-x86_64-amd-zen2
print eessi-arch-guard: reloaded modules for x86_64/amd/zen2 (loaded for x86_64/amd/zen4)
--- stderr ---
(empty)

Stdout is the TaskProlog protocol. export lines are the environment delta. The print line is the message slurmstepd writes to the task’s stdout when SLURM_LOCALID is 0 or unset. EESSI_ARCH_GUARD_RELOADED is was->now. EBROOTR and MODULEPATH now point at the zen2 tree. The process exits 0, so Slurm continues with the task.

How the CPU walk chooses zen2 is covered in Detection and matching.

Run it again on the corrected environment

The example applies those export lines the way tests/slurm_model.py reads a TaskProlog, then runs the guard on the result.

$ python3 docs/examples/walkthrough.py again
case: again
command: python3 -I src/eessi_arch_guard.py
exit: 0
--- stdout ---
(empty)
--- stderr ---
(empty)

The loaded subdirectory is the node the second run detects. A match prints nothing.

Remove the temporary prefix when you are done:

$ python3 docs/examples/walkthrough.py clean
case: clean
removed: docs/examples/_work

Where to go next

A real job shows the same print line. The paths in that line use the site prefix. When the line says the environment was left as it was, follow Correct a guard message. A site that has yet to install the prolog can run the same decision from a batch script with --shell.