2024.04.11 Meeting Notes

Tentative agenda

JM

IO update for non-cell centered vars done!!!
- includes cleanup/refactoring
Added support for coordinates fields (threaded through to xdmf)
- support arb. coordinates transformation in Visit and Paraview
- should be useful for GR and Lagrangian codes

BP

discovered that our current single thread task list can cause issues because threads can currently die silently
now investigating how to fix this, so that the code can die gracefully -- more complex than initially anticipated
- current design is potentially complicating things as there's no main thread and threads queue their own tasks
- similarly, tasks rely on returning (i.e., making a simple check for futures complicated)
- quick fix is to keep check on void and TaskStatus futures (so not fully supporting arb return types)

PM

SwarmPacks works and interface seems reasonable
basic design info right now
- particles only work with base register
- still some code duplication (so could potentially be merged with SparsePacks in the long run)
- Pack caching currently disabled/not working as particles views internally can be resized
Discovered multi-d par_for inner and added functionality for reductions
- will share with BP and then upstreams

LR

Forest got merged
Waiting for review on multiple PRs
Unified flux correction (and fluxes are their own variables)
Updating logical boundary communication to support multiple trees
Investigating logical, topological coordinate transformations between trees (as i,j,k indices do not need to be aligned)

PG

Fighting with OpenPMD restarts (attributes and metadata work, reading blocks is still missing)
encountered issue with h5py when trying to copy the Info group
- code just segfaults, unclear why, not going to fix/investigate as copying contents from that group (e.g., Attributes works fine)

GS found issue on Slingshot interconnect (likely related to number of concurrent sends which hits an internal MPI lib limit)
- currently unclear if its a real issue or sth at the library
slightly related, JM is contact with local postdoc looking at alternative MPI approach (for dense variables)
we'll wait for additional info from Cray/HPE

best practice performance doc
- default pack sizes and loop patterns
investigate performance of loop patterns on various platforms (and new par for patterns, e.g., MDRange in hierarchical parallelism versus index split for vectorization)
related to MPI/Slingshot issue: combine message between ranks