Attending
Tom, Thomas N., Beau, Walls
Minutes
Round table of work related to WRF+ in the past 4 weeks.
ThomasN
- Extended Tapenade test to OpenAD code. OpenAD did not create correct code.
- Contact with MIT people about OpenAD. They have bi-monthly meetings with Argonne to resolve pending issues with OpenAD and ThomasN will attend them as NCAR collaborator (cf. MoU between NCAR and Argonne).
Beau
- MM5 test approach extended to WRF+: isolate test by unit (e.g. winds, temperature, etc). Not complete, partly hard coded, for one time step and the whole solve_em.
- Bluefire: AD test is Ok for serial but fails in parallel and error increases with the number of procs. Explosion for multiple nodes.
Will test individual routines, ThomasN expects the test to work for each routine. Will discuss results and debugging strategy with John Mikalakes.
Walls
- Optimization of AD code nearly finished
- 50-60% speed-up for single proc (tested on Bluefire, Bay, Linux, Mac). 40% speed-up on 32 Procs (Bluefire)
- Memory usage cut by 50% (e.g. for grid=154x100x15 the memory went from 2.1Gb to 1Gb)
- Optimization produces exactly the same results as original code. However, the original code (from Xin) crashed after approx 20 steps.
Will coordinate with Beau to port & test optimization modifs to the "reference" code.
Actions
Tom:
- Send MoU to ThomasN
- Send Don's Wiki link to ThomasN
- Ask Don for a QnA session with ThomasN and Beau about single executable
Next meeting
To limit the number of meetings, it was decided to have WRF+ telecons on a bi-weekly basis.
Next meeting is planned for 11/17 at 10:00AM (MT).