2024.08.08 Meeting Notes

Agenda

• Individual/group updates
• Review non-WIP PRs

Updates

LR

• quality of life improvements (e.g., TypeList) PR done and merged
• GMG currently works with CC variables only. Looking into making this more flexible (to allow for FC var in solution vector)

BP

• KHARMA is back on develop
• upstreamed small patches, which are ready for review
  • predict size of pack
  • par_reduce_inner abstraction
• two leftover downstream patches
  • one for coordinates
  • one for potential race condition between Parthenon and HDF5 MPI
• plus more quality of life improvements that are going to be upstreamed soon, e.g., 3D index ranges
• KHARMA chapter submitted

PM (also reporting more generally)

• needs to review reducing D-H-D copies for swarms
• Artemis (proplanetary disk code) is suffering memory leak (just discovered today)
  • origin unclear, might be related to sparse pools
  • will first confirm on other machine/architecture

AR

• working on PR (already online) to unify kokkos abstractions
  • using type traits to handle index ranges, integers, ...

PG

• discovered excessive memory alloc/buffer overflow for buffer pool when using few (very large) blocks per rank -- will open PR with fix soon
• also identified that boundary comms are not optimal for performance with few large blocks and a single state vector
  • fix might require additional complexity (as already done for prolongation and restriction)
• AthenaPK is now working on the JUPITER testbed JEDI

tentative next meeting 22 Aug