2024.04.25 Meeting Notes

Agenda

JM

added uuid to MeshData infrastructure
- big improvement for downstream codes (to replace string magic)
- discovered independent bug on calling Add on a Partition rather than MeshData
  - following discussion: will hard disable this as it's not used in downstream codes at the moment anyway
other teams at LANL are evaluating Parthenon for additional projects
will look at performance number on Grace-Hopper soonish with Venado becoming available

LR

working on making fluxes PR work with uuid interface
refactored mesh constructors
also working on init of more flexible meshes/forests
looking for short AthenaPK test case for quick integration test
- PG will provide a sample script

BP

had fun with threads (discovered exception now being propagating upward)
- PR is open to fix the issue (checking return status)
- found it due to race condition in profiling and debugging asserts
also wrote a short diff to disable threads
- will open PR today

BR

Added UserWorkBeforeRestart to call third party library restart functionality
PR is open for discussion
- might also discuss if there should be a hook for each output

PM

check for base container for SwamPacks
figuring what all needs to be work to enable caching (around resize pool)
- seems to work well (though potentially purging to many caches)
new SwamPacks resulted in 100x slowdown for certain test problem
- related to needing hierarchical parallelism over blocks and particles
- problem was traced back to the test problem just contain a single block (limiting exposed parallelism)
- need special care to divide block(s) and particles
- JM also came across sth similar with sparse normal packs and created a custom for

PG

Got contacted regarding parthenon(-hydro) more extended tutorial
- once we have sth in place, we could offer it to a broader audience through the "performance portability training series"
OpenPMD output basically working for cell centered fields including restarts
- will remove WIP from PR tomorrow, once the code is a little bit cleaned up and added doc
- looking for early testers

evaluate different abstraction to handle (hierarchical) parallelism that works for small and large number of blocks