Nov 04 2024

Attds: Carl, Gabriel, Stephen, Nicole, Joseph, Patrick

Note-taker

General Topics

• No scheduled meeting 11/18 (supercomputing)
  • We will set up a face-to-face once SC schedules settle down
• Sandia CI
  • A needed feature was enabled this morning, Carl will try to get it set up.
• From last week:
  • Jan and Joseph's feedback on various prior contributions

Round Table

• Carl
  • Working on CI
• Nicole
  • Finalized summer IR stuff
• Gabriel
  • NCCL implementation (copy-pasted MPI subdirectory, trying to follow along)
    • Will open a draft PR
  • NCCL's waits are stream-oriented (keep track of exec space rather than MPI request)
    • Specialize Req however you want
  • NCCL datatypes (no way to create new datatypes)
    • Maybe okay for now (not in interface)
• Patrick
  • Working with Nicole & Evan on Cabana stream-triggered API
  • Some way to batch multiple sends and receives together (e.g. NCCL group, MPI nbrhood collective), esp during GPU triggering
    • Can't kill performance when topology changes
  • Opportunity to provide APIs for what applications implement themselves
• Vivek
  • LDMS expressed some interest in Kokkos + MPI