2022.04.21 Meeting Notes

Agenda

BR

2/3 big particle PRs got merged, last one waiting for input from PG
now back to downstream work
found bug (? or design assumption) with implicit flattening of vars, will open a PR for further discussion. Could potentially get around by changing indices order.

JMM

discovered bug for variables that are not aligned with MeshBlock::shape when writing data
already opened issue

LR

AJ

wrote benchmarking tool to understand what impacts performance in ghost exchanges, see https://github.com/anku94/amr/tree/main/tools
current focus on topology, i.e., neighbor relations
would be happy to receive feedback. Will also share link on Matrix

PG

MPI Comms and bvals
- seem this also fixes bug on Spock (need to confirm that is was this change and not the SW update)
- see more comments further down
Texascale
- excellent weak and strong scaling up to 460k ranks
- separate comm by distance didn't change behavior. Need more testing also on GPUs.

problem with different use of data pointers for fluxes, prolong/restrict, phys boundaries, and ghost exchanges for container at different stages and "in-one" and "per-block" machinery
Generally, still we need comm for non-base containers.
Overall bvar/bval design would benefit from a clean overhaul
Potential issue with Phoebus for using non-base container for comm when removing ResetBoundary
Should keep ResetBoundary for now. PG will open issue and revert changes in open PR.