WeeklyTelcon_20240611

Open MPI Weekly Telecon ---

• Dialup Info: (Do not post to public mailing list or public wiki)

Attendees (on Web-ex)

• Tommy Janjusic (NVIDIA)
• Jeff Squyres (CISCO)
• Luke Robison (Amazon)
• Edgar Gabriel (AMD)
• Todd Kordenbrock
• Manu Shantharam (AMD)
• Howard Pritchard (LANL)
• Wenduo Wang (Amazon)
• George Bosilca (NVIDIA)
• Josheph Schuhart (SBU)
• Todd Kordenbrock
• Aurelien Bouteiller (UTK)

---

v4.1.x

• v4.1.x Issues; v4.1.x Questions
• v4.1.x Open PRs
• Next release when cycles allow it, or if there is a emergency

v5.0.x

• New issues since v5.0.3

• Open PRs since v5.0.3

• Closed/Merged PRs since v5.0.3

• v5.0.x Open PRs

• v5.0.x Issues; v5.0.x Questions

• PRTE issue if PMIx upgrades from v4.x to v5.x without also rebuilding PRTE

  • PMIx changed some internal data structures between v4.x and v5.x (specifically, they are not part of the PMIx public API).
  • The PMIx public API did not change between v4.x and v5.x, so the libpmix .so number did not change to reflect an incompatibility.
  • PRTE uses the internal PMIx data structures, however, and is affected by the changes between PMIx v4.x and v5.x.
  • Meaning: if you build+install PRTE against PMIx v4.x and then upgrade to PMIx v5.x without also re-building PRTE, PRTE will almost certainly segv.
  • Since PRTE is the back-end of mpirun, this affects Open MPI as well.
  • Upstream PRTE is likely to put in a run-time check to print an error and abort (before segv'ing) when detecting this case.
  • See https://github.com/openpmix/prrte/pull/1982 , pull this PR into our fork
  • Runtime test suite, from IBM check with Josh H.
    • Check requirements, there was an IBM CI running on PRRTE (sept. 2023),

Main branch

• Main Issues
• Main PRs

Additional Agenda Items

• Now we have our own PRRTE. People should watch out for pull requests https://github.com/open-mpi/prrte/pulls

• ompi/pull/12356

• cuda reduction, cuda mem topic (Joseph)

  • https://github.com/open-mpi/ompi/pull/12356 , merge
  • Add CUDA/HIP implementations of reduction operators #12569 (large PR, needs reviewers)
    • only have the CUDA backend, generate HIP *

• acoll PR, needs to be merged for v6.0

  • https://github.com/open-mpi/ompi/pull/12484
    • needs to be reviewed, disabled by default

• Large count stuff before fork

• **Fork sometime in v6.0 July **

• Introduce new JSON parser utility #12531

TODO

• Need to discuss MPI 4.1
• Need to discuss MPI 4.0
• Discuss set v5.0.4 timeline
• Discuss v6.0 feature list and timeline
• Discuss 6.0: BigCount
• Explore cherry-pick bots
• Explore an MTT alternative

Notes