Attending

Tom, Thomas N., Beau, Xin, Don

Minutes
  • Thomas ran WRF v2 through Tapenade but it fails at the solve_em() level. Is trying a similar test through OpenAD
  • Discuss priorities and agenda for 7/23 mid-term review. Try to focus on some items that are demonstrable.
  • Discuss strategies for fixing wrf+
    • Zaizhong has run multiple single time steps successfully. (??)
    • We have "consolidated" on the https://svn-wrf-var.cgd.ucar.edu/branches/wrfplus2.0/ wrf+ branch. That is will be attempting all fixes and going forward with this branch.
    • Investigate optimization and warning flag options
    • Priorities:
      1. Get multiple time steps working on bluefire
      2. Get multiple processors working on bluefire
      3. Port to other machines, i.e. PGI (but trying PGI compilers early might provide some diagnostic feedback.
  • Next week, i.e. 7/9 meeting will be at 11:00 not 10:00
Action Items:
  • Tom: Will circulate ACAPS statement of work
  • Tom: Circulate PowerPoint template for those that may be presenting at mid-term review
  • Tom: Investigate phone system to enable multiple call ins.
  • Beau (with Thomas assistance): Run a version of MPAS through OpenAD for mid-term demo
  • No labels

2 Comments

  1. Thomas Nehrkorn

    There is an extensive discussion of why the use of source transformation tools is more appropriate for complex numerical codes (like WRF) than operator overloading in Section 1.4.1 of the OpenAD User's Manual (available from www.mcs.anl.gov/OpenAD).  Of all the reasons listed there, the two most relevant/important are:

     - activity analysis: source transformation will identify "active" variables from an activity analysis of the nonlinear code, and subsequently generate derivtaive codes for only those variables.  Operator overloading would apply overloaded operators to all variables (unless active variables are identified and typed as such manually).

     - reverse mode: while overloading operators with corresponding derivative operators is relatively straightforward in the forward mode (TL), in the reverse mode a simple overloading approach would require a naive store strategy ("store all"), which quickly balloons memory requirements.  Source transformation allows judicious choices of storage, checkpointing, and recomputation by the AD tools and/or the tool user.

  2. Thomas Auligne

    Minutes:

    • Short-term milestone = ACAPS mid-term review 07/23 in Boulder. Each person will do a presentation to describe progress on assigned tasks and future work.
    • ThomasN will survey other AD tools such as operator overloading techniques (Vivtech?, ...).
    • MPAS can be used in the future as benchmark code. Mike Dudha is working on parallelizing the prototype but is on PTO until 7/7. We should start with the serial version provided to Beau.
    • "Lightning" is a linux cluster at CISL with several compilers, which could be used for debugging/testing.

    Actions items:

    • Don/Xin: check with "consulting people" about IBM compiler options performing more checking for debugging purposes.