2022.06.30 Meeting Notes

Agenda

(DH, FG, AJ, PG)

• Individual/group updates
• Intel update
• Developer meeting

Individual/group updates

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)

Intel meeting

• took place on Tuesday
• Parthenon(-hydro) now compiles with SYCL (llvm-parfor bug solved)
• Most unit tests pass but parthenon-hydro crashes
• PG forwarded details (warning/errors) to Intel contact and waits for feedback
• next meeting with Intel scheduled for 26 Jul

Developer meeting

postponed to next meeting

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