4DVAR minutes 08/25/06

Participants: Dale, Hans, John B, John M, Qingnong, Tom

Contents 1 Review 1.1 Hans 1.2 John M 1.3 Tom 2 Discussion 2.1 Strategy on code 2.2 DFI 2.3 nup 3 AFWA/AFRL mid-term review 4 AOB

Review

Hans

• A script for running 4dvar with observation windows centered at analysis time has been developed and tested.

• More observation operators for 4dvar have been coded. Now the only SSMI and radiance are left.

• A few bugfixes have been added.

• John M's parallel code has been tested by Hans and Xin. The adjoint code seems to be the bottle neck.

• Xin has made the parallel code working for serial mode.

• Hans and Xin have merged all the new developments and updated the 4dvar branch under subversion with John B's help.

• Xin has managed to compile the 4dvar branch using the Linux Intel compiler and g95 compiler.

• Hans and Xin have made the initial development of JcDFI. The code seems run correctly, but further experiments and tuning are necessary. Noise level, typhoon tracks and observation verification scores will be computed to assess the meteorological impact of 4dvar and JcDFI.

John M

• The first version of parallel 4dvar has been made and a tarfile created for the group to make test. The problem associated with the adjoint model is known and will be investigated next. We may need another assessment on whether we can avoid to many re-computations in the adjoint model.

• John agreed to check the code Xin made and may get Xin started with optimize the adjoint model.

• The current parallel code is based on some features of the load leveler, which should be fine for bluesky and AFWA's IBM. However, we need to make the 4dvar code running on bluevista as bluesky will stop by the end of the year.

Tom

• Got some information about Harvard-FastOpt collaboration. There are still problems using TAF on some F90 code. The approach the Harvard group takes is to maintain ~15 problematic subroutines manually. [Dale: I think Tom meant "15 problematic areas", which could be in many more individual subroutines. Tom - can you clarify? [Tom: Dale, you are correct.]] To get additional support from FastOpt needs to pay extra on top of the lisence fee. We can also sub-contract FastOpt some tasks. [Tom: FastOpt has thus far refused any contracting that involves actual devliverables. Only time-and-effort contracts are accepted.]

• John commented that in WRF, F90 features are mainly above the solver level. We should be able to use TAF without major problems below the solver level.

• Tom will give a review at the AFWA/AFRL mid-term review next week. He will continue the investigation and make recommendations before the end of the current TAF lisence.

Discussion

Strategy on code

• Hans and Dale proposed a strategy for the 4dvar code. As it is not clear whether we can maintain TL and AD, using TAF, in the same WRF repository, we may construct the 4dvar with 3 executables: VAR, WRF and TL+AD. VAR and WRF will be in one repository. TL+AD can, at least for now, remain on the old branch.

• John B wants to put an end to the var component in the 4dvar branch once the suggested system made.

• John M explained the current 2-executable MPP setup. He will discuss with SCD on how to contruct the 3-executable MPP setup.

DFI

• Hans also has difficulties to put all DFI related code back to WRF repository. The group had a long discussion on how to proceed, but no clear agreement has been reached.

nup

• Hans pointed out to the group the last bit of our multi-incremental 4dvar relies on the further development of nup. John agreed to work on it later with Hans.

AFWA/AFRL mid-term review

• Hans agreed to prepare an overview presentation. As he will be on travel next week, Qingnong will give the presentation.

• John M will present his work on parallelization.

• Tom will present an overview on the TAF applications.

• Xin will present the JcDFI progress.

AOB

The next meeting: 9/29/2006.