2021.05.27 Meeting Notes

Agenda

• Individual/group updates
• Paper
• Review non-WIP PRs

Individual/group updates

LANL/External CS

Joshua Brown

• Fixed some CI issues
• Fixed CI issue that wouldn't run after a branch was rebased
• Working on open code coverage PR
• Looking into performance impact from HDF5 compression: https://github.com/lanl/parthenon/issues/470
• Setting up training with HDF5 Group
• GitHub application backend for performance regression is available standalone now for other projects, working on adding documentation
• Looking for suggestions for additional performance regression tests. Only running MPI advection right now

Jonas Lippuner

• Working on sparse, added SparsePool and made StateDescriptor aware of sparse fields/pools
• Question: Is it ok to make a first sparse PR that introduces new interface but not sparse functionality in the sense of allocating/deallocating sparse fields on the fly (all sparse fields would have to be allocated on all blocks, like now). Consensus: this is a good plan to keep PR size manageable

Tanner Broaddus

• Will be looking into some MPI communication patterns
• Managed to build Parthenon with external Kokkos library

LANL Physics

Ben Ryan

• Working on downstream astro codes
• Doing some tasks on particle PR. One down one two to go

Jonah Miller

• Unified restriction PR finally ready to go in, fixed two issues:
  • Slower than before change, caused by old machinery was still called, removing that resulted in 10x speed up
  • Test failed, solution was to add restriction in some places that were missed (Mesh initialize and restrict before prolongation), all tests pass now

AthenaPK

Philipp Grete

• Busy writing proposal => probably/hopefully more folks working on Parthenon
• Mentored GPU Hackathon team in France working on astro code with Kokkos
• Looked into shared memory on GPU, team policy with shared memory might not work very well, or needs some extra care to make work well
• Getting performance metrics for unified restriction
• Reviewing particle PRs

Forrest Glines

• Submitted PR to add workaround for GCC 7.3 bug
• Submitted PR to fix component name order
• Also working on PR to make Python tools more efficient

Agenda

Paper

• Once Hackathon improvements are merged in, we can start writing paper showing good CPU/GPU performance
• Philipp will start outlining paper and assign tasks, will open an Overleaf project

TODO

• Philipp to submit PR with additional performance regression tests (the ones we used during Hackathon)