WeeklyTelcon_20230829

Attendees (on Web-ex)

• Luke Robison [AWS]
• Jeff Squyres [Cisco]
• Quincey Koziol [AWS]
• David Bernholdt [ORNL]
• Edgar Gabriel [AMD]
• Joseph Schuart
• Matthew Dosanjh
• Pedram Alidazeh [AMD]
• Thomas Huber [Cornelis]
• Thomas Naughton
• Todd Kordenbrock
• Wenduo Wang [AWS]
• Tommy Janjusic [NVIDIA]

v4.1

• Merged a pair of Fortran related PRs
  • Another RC yesterday, hopefully released by end of week. Low impact release.

v5.0

• mpirun manpage - 90% done

  • Build infra stuff, pretty close to be done
  • ETA later this week (Friday likely)

• Issue 7668: ORTE -> PRRTE users educated on changes

  • Need to establish the bar for acceptance criteria that satisfies the needed changes.
  • Building list, going through source, will produce an RST doc, or table with mappings.

• Accelerator framework delayed initialization

  • Is delayed initialization - If Cuda Init wasn't initialized before MPI_Init, then components can't get some info from Cuda, and doesn't try later. So order of initalization between components (MPI_Init) and Cuda is important.
    • In v4.1, we had to initalize Cuda_Init before MPI_Init()
      • Delay was supposed to fix this, but doesn't.
    • One thought to fix this is to have MPI Initialize Cuda.
      • Always, even if user doesn't use Cuda
  • Not a blocker for v5.0.0, to be addressed in a future fix/patch.

• ABI telecon follow up

  • Fortran , how to handle ABI compatibility in Fortran and C and what does it mean foundational place we got to, C knows about Fortran, and MPI Fortran knows about C stuff
  • C/Fortran ABI compatibility defined as a platform between the System OS and a particular Fortran Compiler