Attending

Tom, Thomas N., Beau, Walls

Minutes

Round table of the recent work related to WRF+.

Walls
  • Provided Results for a Test Case for WRF TL/ADJ optimization on single proc.
  • Results are identical to the original WRF+ branch provided by Xin. The multiple proc configuration fails. The single proc configuration crashes after about 20 time steps for a medium-size test case, but it runs through for a smaller case.
  • For larger cases, the single-proc configuration fails on Bluefire but runs on Bay. This is probably due to memory limitations.
  • 5-10% speed improvement is expected from further optimization, but this work is postponed for later.
Beau
  • Documentation: improvements to flow charts and defects references.
  • Wrote a script to do a recursive grep in WRF+ code to assess the use of Namelist entries. Intermediate conclusions: about 50% of Namelist entries are never referenced in the code. Some namelist entries are only used in commented lines. Some Namelist entries have the same name as WRF (NL) code but correspond to different functionalities.
  • Variable isolation (U, V, W, T, MU, ...) nearly ready for TL/AD test
  • Noticed that the TL/AD test is not constructed for parallel code. The current strategy is to focus on serial test and then check that parallel runs produce identical results to the serial run.
ThomasN
  • Working on testing OpenAD automatic differentiation of MPAS code. Issue with arguments in subroutines:
    • derived types are not supported (yet)
    • issue with the handling of the dimensions of arrays. Various code configurations give different results, none of which is correct.
  • OpenAD not ready yet for C+, because it is harder to control the flexibility allowed by this language. C+ may never be fully supported.
Actions
  • Walls: Reproduce the test case with WRF+ "Reference" branch provided by Beau.
  • Beau:
    • Discuss with Xin to define which namelist entries are needed and a "trimming" strategy to simplify the code.
    • Run 4DVar in serial and parallel and check that results are identical.
  • ThomasN:
    • meet with Argonne and MIT folks next week
    • Document how OpenAD is breaking and provide feedback to Argonne
Next meeting

TBD, after Xin's return on 11/02.

  • No labels