how to further improve the running/outputs speed of real.exe ?

[esmwg2020]

I am now conducting the real.exe with the WRF ARW V4.0, and the details can be seem in the attached file namelist.input. It is at 15.0km grids space with 345X310 grids.

Now I have to conducet a longterm simulation (about 30 years), and I need to processe a laege number of time steps in a single lateral boundary file (the issue of limited 10000-maximum number of time periods have been adress, see my former post thread. Thanks WRF&MPAS-Aforum and kwerner again).

When I run the ./real.exe (mpirun -np 24 ./real.exe &), I found the output speed is much lower than I expected. It need almost 3.0 seconds to deal with each loop:
...
d01 1993-03-23_06:00:00 Timing for processing 0 s.
d01 1993-03-23_06:00:00 Timing for output 0 s.
d01 1993-03-23_06:00:00 Timing for loop # 4710 = 3 s.
...
I have tried to use mpi and with more nodes, more tasks and more cpus (> 120) to run the real.exe, but it seems that it still use 3-seconds for each loop.

I also tried to modify the &namelist_quilt options here:

&namelist_quilt
nio_tasks_per_group = 1,
nio_groups = 6,

but it didn't work.

So I hope you could give some advice to improve the speed of running real.exe except adding more cpus ?

Thanks a lot.

 

[Ming Chen]

Hi,
REAL usually runs quickly and the grid number of your cases (345 x 310) is not large. In this case, I won't suggest you run with the quilt options. Instead, you can ru this case using 72 processors, which I suppose should be sufficient enough and ensure a quick run.
One issue I would like to discuss is that, because this is a long-term run, probably you should run with the 'restart' capability of WRF. You may create separate wrfbdy files, each covers the specific period of your simulation. This will avoid taking long time to read/write a large-size wrfbdy file.
Hope this is helpful for you.

 

[esmwg2020]

Thank you so much for giving these helpful suggestions.