Participants

Tom Auligne
Don Stark
Xin Zhang
Zaizhong Ma
Beau Paisley
Hans Huang
Thomas Nehrkorn

Agenda

Telecon w/ Thomas Nehrkorn. Objective = to discuss his progress and ideas on WRF+ code.

Minutes (thanks to Don)
  • Issues:
    - Performance: storing vs. recomputing options
    - Testing: need a standard testing infrastructure in place for effortless unit testing. WRF Plus already has something, but may need to be revisited.
    - Parallelism: further tests need to focus on parallel code examples, rather than the serial examples examined so far. A major decision as to how to go forward will depend on the answer to this!
    - Simplify forward prior to forming tangent linear code and adjoint by replacing the RK3 forward integration with something that requires fewer function calls (e.g. Leap-Frog).
    - Initial discussion with Bill Skamarock (post meeting):
    • Leap-Frog used to be available as "option 1" in WRF code. It was not supported for a long time and has finally been removed from the code. It might work (or not...) in WRF+ code generated from WRF version 2.0.
    • The Leap-Frog scheme may give dodgy scientific results for two particular aspects: moist physics and damping. Needs testing on idealized cases.
  • Automatic Adjoint creation tools:
    - TAF: ThomasN felt that not much as changed with the TAF tool since the initial WRF Plus development and is therefore focusing on newer products.
    - Tapenade: ThomasN has processed serial versions of some of the lower level solver routines. When it fails, there isn't much feedback as to what the problem is, so it has been a time consuming task. This tool is designed to store all intermediate information in the stack. This differs from the strategy employed in OpenAD.
    - OpenAD - http://www.mcs.anl.gov/OpenAD/, is an open source product from Argon NL and others. It looks promising since MIT has replaced TAF as the tool for  GCM adjoint creation. ThomasN hasn't had a chance to examine it closely yet, but will do so next. A survey of OpenAD presentations: http://www.mcs.anl.gov/OpenAD/. OpenAD is designed to allow the user to specify how much of the intermediate information is vs. recomputed.
    - NAG F95 compiler based adjoint creation. On first pass it appears that this project is no longer active, but ThomasN will check into it.
    - All other tools looked less promising because they rely on operator overloading or are of limited scope.
  • Thomas's work so far has consisted of testing serial versions of the low level solver routines in solve. The performance bottleneck has more to do with the upper level routines and parallelism, so he needs to refocus work in those directions.
Actions
  • Thomas N. to examine the available automatic adjoint tools out there and report findings by July 1, 2009 in time for AFWA report
  • ThomasN to contact Tom Henderson about his previous attempts at automatic adjoint creation, pay particular attention to a review he wrote while still at MMM concerning what he had found concerning the compilers
  • Beau to get spun up with WRF+ testing package for future support to test ThomasN developments
  • Xiaoyan (+Beau+Don???) to test "Leap-Frog" solver (option "1") in WRF+ code.
  • No labels