Skip to content

2023.07.27 Meeting Notes

Jump to bottom
Philipp Grete edited this page Jul 27, 2023 · 2 revisions

Agenda

  • Individual/group updates
  • IO design for non-cell centered fields (deferred)
  • MeshBlockTree
  • review non-WIP PRs

Individual/group updates

JM

  • attended workshop and advertised Parthenon
  • will setup separate call to exchange experience with Parthenon and similar codes on Frontier

BR

  • still looking at swarm packing
  • trying to separate/opensource non-rel MC rad transport as stand-alone Parthenon package (on 6-9 months timescale)

FG

  • working on coordinate systems (sph/cyl)
  • will not affect codes that just use standard logically Cartesian based ones
  • we might have missed a couple of coordinate abstraction when refactoring (to be confirmed)
  • bigger question: are all the prolong/restriction ops (and others) really correct for non-Cartesian coordinates as they're implemented/refactored right now (e.g., Xc and Dxc are identical for Cartesian, but are not for other coordinate systems)?
  • will put example (non-working) PR up today
  • keeps INCITE sims running

BP

  • working on cleaning up open (and to be opened) PRs backporting methods from KHARMA

LR

  • working on block level geometric multigrid
  • prolong/restrict already working
  • uses recursive task list
  • still missing smoother
  • current approach will not go coarser than single block (to reduce comm)
  • looking for feedback on what to do with non-cubic root grid

BW

  • working on finishing PR on UserMeshWorkBeforeOutput (add test)
  • got AthenaPK compile with nvc++ (with old version), but didn't work correctly

PG

  • IO issue on Frontier seem to be filesystem related (nothing in Parthenon)
    • disabling progressive file layout, increasing number of storage targets, and disable collective buffering at the MPI IO layer results in up to 7x increase in write performance
    • will fix chunking PR (staying at original defaults) and document usage on Frontier for larger outputs
  • performance regression on Frontier with latest software stack, should be updated with Slingshot update (soon). For now, use old stack.

IO design for non-cell centered fields (deferred)

  • to be revisited once more downstream codes (particularity AthenaPK) have POC implementation

MeshBlockTree

  • see discussion https://github.com/parthenon-hpc-lab/parthenon/issues/912
  • with the updated LogicalLocation there's an explicit, natural (hashable) ordering of blocks
  • that'd allow to put the blocks in a std::(unordered_)map versus a search in a tree
  • might simplify/enable neighbor finding, boundary comm, distributed mesh structure, ...
  • but current implementation is working as is
  • Geometric MG (see above) already does this (replicating the tree)
  • thus, path forward (for now), to keep both in place and then potentially refactor later (based on performance numbers and opportunities to cleanup code)

review non-WIP PRs

Clone this wiki locally