2023.03.30 Meeting Notes

Agenda

JM

uid registry for fetching vars at higher perf is ready for review. Caching uses unique ids (numbers versus strings originally)
integrator PR waiting for reviews (includes bugfix for storage buffers independent from stages)
will revisit particle output early next week. First focus on O and then I.

BP

wrangling with freeing MPI (after use in reduction objects)
bugfix already merged but not a full fix as some free after use still remain (for specific combinations of MPI libs, HDF libs, ...)
- there seem to be intermittent issue with race conditions between freeing a communicator (which seems waits for all processes to actually decallocated). An HDF output (requesting a new communicator) may trigger this.
if someone still sees this issue and can create a reproducer, we should revisit.

LR

finished fixing uid PR (see above)
found issue with std::filesystem backwards compatibility, could be fixed with two lines of CMake voodoo
revived timing based load balancing in Riot (but required pack_size=1 right now). Seems to work nice. Should be backported to upstream Parthenon, soonish.

PM

fighting with yt output, we need to double check that vector based outputs works

FG

PG

finally Ascent PR merged, found bug in all application (driver needs to be scoped for proper cleanup before Kokkos finalize)
CUG23 paper almost done, people who're interested in feedback, please contact PG next week
yt frontend for Parthenon close to final review (first tests added)
- important: please send sample output from downstream codes so that it can be included in the yt test infrastructure.

if it's 100% remote and transparent, then why not
need to ensure/figure out what this mean for the developer (e.g., does dev need to pull own changes)