Skip to content

2022.07.28 Meeting Notes

Philipp Grete edited this page Jul 28, 2022 · 1 revision

Agenda

  • Individual/group updates
  • Face/edge centered fields

Individual/group updates

JM

  • still figuring out performance issue for prolongation/restriction
  • different approaches are good/bad for small/large pack sizes and small/large block sizes
  • new abstraction for prolong/restriction should also help in designing face/edge fields

LR

  • sparse bound done (also removed old machinery) at same or better performance
  • also allows to tune order of communication
  • new way for individual alloc/decalloc and init of sparse vars

AJ

  • calculating chi^2 for various parameters like numbers of block per rank
  • collecting traces for messages, too (in addition to the task based ones)

FG

  • working on face and edge centered fields, see below

PG

  • got scaling data on Frontier, results should go in the paper
    • packing issue above also relevant for performance here
  • updated CI environment to use GitHub Actions rather than Gitlab

Face/edge centered fields

  • mock-up downstream app for fake-loop advection for illustration
  • agreement that new "sparse" pack should be used (and old packing machinery potentially go away entirely at some point)
  • for indexing we should stick to current convention (block first,then variables, then kji)
  • big question: just CellVariables with new distinguishing MetaData (and nx#+1) or different variables types (or using substructures inside CellVariables)
    • CellVariables approach seems like least complicated
  • will schedule online hackathon (likely in Sep)

Long term ideas

  • autotuning (of buffer filling routines, comm paths, pack_size, ...)
  • set pack size by number of packs per rank
  • make sure GetOrAdd is only called with default number of packs
  • in-situ/inline analysis and post-processing
  • separate output "thread"
  • remove "non-pack" machinery (enforce index b) to reduce interface

Next meeting in tbd.

Clone this wiki locally