2026 06 18 meeting minutes

Attendees: Carl, Evan, Kaelyn

General Topics

Changelog (post 0.1)

• Cedric will give this a try, follow up after his absence
• changelog.md, modified when we merge a PR
  • When you open a PR, you get a number -> add a commit to your PR that adds the PR to the changelog

• Planning for 0.2 (end of June, PASC):
  • static_assert on GPU buffers in MPICH
  • fix MPI detection in cray-mpich (Carl)
    • Workaround at configure time to say that MPI is MPICH
  • Non-GPU-aware MPI host staging (Gabriel)
    • staging / packing / etc logic is repeated everywhere, working on abstraction
  • Spack package recipe (Cedric)
    • draft PR open in spack repo
  • NCCL and MPI results from the same code (Gabriel)
    • OSU-style latency benchmark (MPI + NCCL enabled at the same time)
      • Gabriel: some overhead from fence in MPI backend (hopefully not present in stream-triggered backend)
      • Open draft PR so Nicole can use (Gabriel)
    • Heat3D benchmark in her fork -> open a PR (Nicole)
  • Blog post (Carl)
  • Fixing documentation (Carl)
    • Cedric opened #201
    • Things have been partially addressed in multiple PRs that have been merged recently
  • miniFE port (Carl)
  • Lulesh port (Gabriel + minions)
    • port is working, but not 1-1 of existing code (some refactor, etc)
  • Stream Triggering (Evan)
    • presentation in 2wks or 4wks, get feedback on interface
  • MPICH + CUDA test runner (Carl)
  • Known Issues in documentation

• Initialize NCCL once per test binary rather than once per test. Improves run time a lot.

Carried forward from last time

• Updates on CI dev container? (Cedric)
  • draft PR open
• Ron Brightwell gave Kokkos Comm talk at CEA/NNSA meeting 7/16.

Round Table

• Gabriel

  • Working on draft design document for packing / staging abstraction
  • const / non-const Views in e.g. send / recv APIs
  • MemorySpace concept

• Evan

  • Have a plan for the Stream triggering interface
  • In 4wks, 20 minutes on stream triggering