2022.06.30 Meeting Notes

Agenda

(DH, FG, AJ, PG)

FG

making progress on face centered fields (including edge centered fields)
implements "backwards" example (EM wave) within Parthenon to figure out what needs to go where
will open PR soon to get feedback on design
had separate meeting with LANL

AJ

working on performance profiling
shared plots for feedback
- run on 512 ranks (CPU)
- 15-40% of time is spent on UpdateMeshBlockTree (across 30k cycles) using tau for tracing
- potential issues in WaitForAll, but Q is whether this is for stragglers (and, if so, if the stragglers are the same between ranks)
- according to heatmaps (cycles versus time/rank) shows that same ranks are typically slowest/fastest
suggestions
- check if results are consistent across runs/restarts (to rule out hardware/network issues)
- so far running with pack_size=-1, try also pack_size=1 -- especially as on host

PG

added/fixes several small pieces (e.g., Mesh level pgen, output component labels, MeshBlockUserWorkBeforeOutput)
reviewed/merged parallel scan PR
looking at storing (and output of) generic Mesh level data
trying to get Parthenon running with SYCL backend on Intel hardware (see next item)

postponed to next meeting

Next meeting: tbd, need to check avail of LANL folks first