2024.05.16 Meeting Notes

Agenda

PM

Swarm packs are in
Debugging xdmf spec to fix Visit viz
Adding particle data view for Visit (via xdmf)
- parts work, but some issue wrt to geometry specs remain
  - can be fixed by separate data structure containing coords
- need unique xdmf file for particles and fields at the moment
  - can be fixed by using separate xdmf files

PG

Had a long meeting with ADIOS2 and OpenPMD devs
Current approach for defining mesh through OpenPMD seems fine
- dumping data to a host buffer is waste of memory
- in the mid-term could be updated to directly write from device memory (though support for "negative" offsets/ghost needs to be added to OpenPMD, see next item)
- while we have a host buffer, recommendation is to increase the ADIOS buffer to it's max (2GB)
Dumping ghost zones is currently not supported through OpenPMD due to missing exposed ADIOS2 API
- other group also interested in this, so this will be supported sooner than later
Exciting news about data compression/reduction (MGUARD)
- ADIOS2 devs are happy to analyze some sample data blocks wrt to optimizations

LR

Bugfix for restarts (to enable bitwise exact restarts due to missing dealloc counter for sparse and blocks)
- one edge case may remain for sparse variables https://github.com/parthenon-hpc-lab/parthenon/issues/1077
Bugfix for static meshes (introduced with trees)
Added option for having "fine" fields (double resolution than standard blocks)
- will add AMR support
Refactor restarts ready for merge
General 2D forests are ready for review
Continued development on forest infrastructure

BR

fixed boundary conditions for swarms to work without RDC
fixed setting neighour index so that it's not required in downstream codes any more
MPI reqeust may become invalid with swarms
- current work around is to update MPI tag def
- suggested approach is to use separate MPI Comm

BP