Skip to content

2026 04 09 meeting minutes

Jump to bottom
Gabriel Dos Santos edited this page Apr 9, 2026 · 2 revisions

Attendees:

General Topics

  • 0.1 release

    • Milestone overdue by 1 month
    • A few remaining items, but we should do a release now
    • Some key additions:
      • MPI-backend exposed in the core API
      • Refactored Requests: shared API between backends, easier to write CommSpace-generic code (doc up-to-date)
      • Refactored Communicators: shared API between backends, new capabilities: split, duplicate (doc up-to-date)
      • DeepCopy packing strategy can now handle Views of arbitrary rank
      • Collectives don't have limitations on View ranks anymore (doc specifies semantics)
      • Cleaned up Kokkos Comm traits (doc up-to-date)

  • Items postponed to 0.2:

    • Spack package recipe (Cedric)
      • better to have 0.1 released anyway before writing the recipe
    • NCCL and MPI results from the same code (Gabriel)
      • Can TN Tech open a PR for their Heat3D benchmark? + any other bench they have
    • Non-GPU-aware MPI host staging (Gabriel)
      • critical to have before PASC
    • Blog post (Carl)
    • Fixing documentation (Carl)
      • Cedric opened #201
      • Things have been partially addressed in multiple PRs that have been merged recently

  • Aiming for PASC to do a 0.2 release?

Carried forward from last time

  • Kokkos Comm meeting during TN Tech visit to SNL in March

  • Kokkos Comm at HPSF 26

  • Synchronizing our release with the rest of the Kokkos ecosystem (Carl to report back)

  • Spack Package

    • Prototype, no real dependencies, so pretty straightforward for now

  • NCCL / MPI common examples/performance tests

    • Some runs can happen at CEA, need to look into runs at SNL

  • Host staging capability for non-GPU-aware MPI PR

    • still in progress, prioritized after NCCL+MPI

  • Documentation review

    • Stephen added a checklist in #201 that needed to be fixed or added
    • Automatic checks for documentation updates after 0.1?
      • Seems possible for API documentation
        • Start with the checklist for PR that the reviewer will try to hold the author to
        • A targeted use of GH Copilot to perform this check?
      • For other explainers, etc probably not possible

  • Updates on CI dev container? (Cedric)

    • Waiting for build system rework from Gabriel

  • MPICH (via Spack) + CudaSpace views, exclusive scan fails

    • SNL CI only tests Open MPI + CUDA
      • Carl will make a container
    • Not sure if it's an MPICH problem or something else we're doing wrong
    • Gabriel is splitting up the unit test

  • Message in mpi-interop Slack channel

    • Hydro code, saw Gabriel's presentation, may have a simple use

  • External contribution for the RCCL backend

    • Need to review
    • Centralizes NCCL + RCCL shared patterns so we implement bindings only once

Round Table

Clone this wiki locally