Progress review

Physics (Qingnong)

Jimy has made the technical outline of what should be done in equations. Qingnong needs the nonlinear code this month for evaluation, in order to meet the TL and AD deadline.

Meteorological testing (Hans and Xin)

Fix bugs in transfer between wrfvar and wrfplus. This makes the adjoint check passed. Re-runs are ongoing with the bug-fixed code.

Hans has also started to look for new cases.

Multi-incremental (Xin and Hans)

The fixed 1/3 ratio program from high-resolution to low resolution, nup, and its inverse, nup-inverse, are done.

In the first multi-incremental formulation, wrfvar will be called twice. The first time is for the innovation calculation using the high-resolution background and observations. The second call is for the minimization of the inner-loop on the low-resolution grid. The write and read routines for innovation are finished, now in tuning and test stage.

Optimization (Tom)

I continued working on optimization of 4DVAR adjoint via trade-off of re-computation for memory. The first two trade-off options are now complete and tested for a serial build. The first reduces the run time of serial solve_em_ad() by about 5% and the second by an additional 7%. These numbers are rough estimates because the test platform has been overloaded with other jobs and performance varies from one run to the next due to contention. This problem should be lessened once I begin testing parallel configurations.

I analyzed dependencies in detail in the dynamics section of solve_em_ad and discovered additional opportunities for memory vs. re-computation trade-offs. Some of the pre-existing store-restore code may also be unnecessary. Also, many of the halo updates inside nested loops appear to be unnecessary. I will remove these where possible in a second round of optimizations to improve scaling.

Thomas Nehrkorn and I drafted a proposal for tasks he might perform to contribute to the 4DVAR optimization effort. Hans and Dale reviewed the proposal and decided not to fund it since many of the tasks are already covered by other efforts.

Other projects

Xiaoyan and Qingnong are testing 4dvar over a new domain and on the Katrina case: 101x101x31, 6km, 1h assimiliation window.

Yongrun and Jenny will test 4dvar on the IHOP case.

AOB

Yongsheng will start to work on 4D-Var from tomorrow (May 1).

From this year, AFWA requires a monthly report. Hand did the first one based on the March "meeting" memo. Hans asked Qingnong and Tom to write short summary on the convection and optimization tasks. Agreed.

The next meeting will be 10:00 May 31, room 2072.