Skip to content

2023.01.12 Meeting Notes

Philipp Grete edited this page Feb 1, 2023 · 3 revisions

Agenda

  • Individual/group updates
  • OpenMP parallelism (outer inner)
  • Nested parallelism
  • Status of face centered fields
  • Support for Ascent
  • Check stale branches
  • non-WIP PRs

Individual/group updates

JD/LR

  • working on downstream codes discovering "interesting" Parthenon behavior
    • "packing by list" could become "expensive" for 2D problems (with little data), but not really relevant for 3D sims
    • call vec.size() rather than caching size
    • creating variable labels by creating new string rather than using cached ones
    • will be easy to fix (simple tuning for infrastructure calls)
  • working combining sending "null" data multiple sparse vars between blocks
    • originally sending quite a lot of messages
    • coalescing sometimes helped (but strongly dependent on MPI lib)

BP

  • working on getting KHAMRA from diverged "forward" (LANL) branch to develop (100 commits behind and ahead)
  • redesign KHAMRA to subclass packages
    • what about the "refinement" packages. Is it useful at this place or out of place?
    • will write a spec ("Parthenon Enhancement Proposal") to discuss

JM

  • started implementing experimental DG method on top of Parthenon
  • big Q: how to translate DG "language" to Parthenon vocabulary

FG

  • working on AthenaPK shakedown tests
  • updated coords class merged
  • looking at options to change primitive prolongation

PM

  • got access to Polaris for INCITE

JS

  • working on LB in AthenaK for AMR

PG

  • fixed Crusher/Frontier configuration (PR ready for review)
  • updated AthenaPK and parthenon-hydro to run on Crusher/Frontier
  • started WIP to fix compiler warnings in the entire code base

OpenMP parallelism (outer inner)/Nested parallelism

  • JD discovered really bad performance OpenMP at some point but seem to behave predictable (and as expected) right now
  • Packing of blocks has impact on performance in combination with environment variables on OpenMP thread placement

Status of face centered fields

  • JD will cleanup "CC->Var" PR as soon as feasible as other works depend on it (and it's 95ish% done)

Support for Ascent

  • should be optional dependency
  • can this be done as a package?
  • -> implement minimal version for now and update/separate later incrementally

Next meeting Feb 2

Clone this wiki locally