4DVAR minutes 11/30/06
10:30 am 11/30/2006, room 2072
Participants: Dale, Hans, John M, Michael, Qingnong, Tom, Xiaoyan, Jimy, Jenny, Yongrun, Chris
Contents |
Progress review
Merge 4D-Var branch to the main trunk
4D-Var 2.2, the "3 exes" (wrftrunk, wrfvartrunk and wrfplus) system also works in parallel mode.
Veritical diffusion and large scale condensation
The tl and ad of the vdiff scheme have been developed and passed both gradient test and adjoint test. Testing in 4D-Var has started and will be reported next time. Hans asked for the vdiff related code. Qingnong agreed to commit them in, in about two weeks.
A nonlinear routine of the large-scale condensation has been coded and is under testing. Work will soon start on the tl and ad.
TAF
The source to source converter has been completed and tested. The problem now is the generation of adjoints using TAF. Tom made an agreement with FastOpt for speeding up the communication and problem solving process. He is now preparing a complete testing case, which will be sent to FastOpt.
John will contact Thomas Nehrforn and arrange for his visit. We will discuss the optimization issue with him and the possibility of collaboration.
Meteorological testing
The Haitang case (4 cycles per day for 3 days) has been re-run with both "centered assimilation window" and "one-sided assimilation window". The preliminary results indicate the centered window setup may still have problems.
In the KMA project, we have a task to compare 3DVAR and 4DVAR. A KMA case, over a 5 day period, has been tested. Verification is ongoing.
Dale will soon start 4D-Var tests using CONUS configuration.
Jenny and Yongrun will also start 4D-Var tests on convective scale domains.
JcDFI
Xin started another round of tests, using 2.2.
General discussion about 4DVAR cost
We have had many email exchanges during the month on how to reduce the 4D-Var cost. The aim is a factor of 10 in cost reduction.
Using Dale's proposal as a strategy, we will work in 3 areas.
1) Numerics. Hans agreed to take the lead. The current thinking is simplify the Runge-Kutta steps by removing the nonlinear re-computations. This is expected to give a factor of 2-5.
2) Optimization. John agreed to take the lead. Through reducing recomputation and communication, he hopes to get a factor of 2.
3) Var optimization. Dale agreed to take the lead. Through tuning obs and bg errors, and improving the preconditioning, he also hopes to get a factor of 2.
AOB
There is no comments on the last meeting memo. Dale suggested sending the memo out in pdf format, in addition to the link. Hans agreed.
The next meeting: 10:00am, 12/29/2006.