Skip to content

Evaluate a GNU 15 / Open MPI 5 variant #94

Open

Description

@mahf708

UBI9 ships gcc-toolset 13, 14 and 15. A gnu15-openmpi5 variant would let E3SM test newer Fortran/C++ compilers and Open MPI 5 (PRRTE launcher). Add it as a new variant, not a replacement, and not as part of the rename. GCC 13 is what the E4S cache is built with; gcc 15 means building most of the stack from source, so this is much cheaper after #69.

Detailed plan (for an agent or contributor)
  1. Add a variant (catalog Define the variant list once and derive the workflow matrices from it #70): COMPILER=gcc GCC_TOOLSET=15 MPI_SPEC=openmpi@5.0.x LAPACK_SPEC=openblas WITH_CUDA=no WITH_MOAB=yes, both arches. Pick the Open MPI 5 version spack/builtin v2026.06.0 marks preferred.
  2. Watch for: core_compilers in ghci/env/Dockerfile step 9 (already adds the toolset's version); boost reuse (cache's boost is gcc 13-built; header-only use is fine); Open MPI 5 needs PRTE_MCA_* (the exit-sync var is already set) and --oversubscribe/--allow-run-as-root flag differences in smoke tests; PMIx/hwloc externals.
  3. Smoke must pass; then run the E3SM cases from Nightly / on-demand job that builds and runs a short E3SM case in each image #93 against it (manually via workflow_dispatch if that exists). Record build time (cold and with Publish our own Spack binary cache on GHCR #69).
  4. Decide on the PR whether to publish it (tag per Rename images to e3sm-image:<compiler><ver>-<mpi><ver>[-cuda<ver>][-dev] #80: gnu15-openmpi5) or keep it as a CI-only candidate.

Acceptance: variant builds and passes smoke on both arches; E3SM small cases compile/run (or failures are documented as E3SM issues).

Ground rules for whoever picks this up: one focused PR per issue (reviewers want one thing at a time); push and let CI validate across all five images and both arches rather than blocking on slow local image builds; keep old consumer tags working unless the issue says otherwise; keep comments in the Dockerfiles/workflows in the existing explanatory style.

Part of the follow-ups from the repo review; see #68 for the first batch of fixes.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestghcigithub ci

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions