[Wrf-users] gcc and gfortran builds?

[Jan Ploski]

Davide Cittaro <daweonline at gmail.com> schrieb am 02/29/2008 03:59:45 PM:

> Hi Jan
> 
> On Feb 29, 2008, at 2:48 PM, Jan Ploski wrote:
> 
> This page may be helpful:
> 
> https://bi.offis.de/wisent/tiki-index.php?page=WRF-gFortran

> 
> Great! Thanks so much!
> 
> However, note that I've had some significant differences in the output 
> between gfortran and PGI, so be careful to check the results after 
you've 
> built the code. I think g75 is officially supported (may be better), but 

> don't know about OpenMP in it. Based on my experience, the OpenMP 
version 
> of WRF seems to scale up poorly compared to MPI. For example:

> https://srvgrid01.offis.uni-oldenburg.de/wrf-benchmark/case.pl?
> case=conus12km_2001
> 
> Very interesting benchmark, I will keep it in mind. 
> Since MPI seems to perform better, are you aware of different MPI 
> implementations performance with WRF? Supposing I can use MPICH, 
> LAM/MPI or OpenMPI (I'm not sure if I can do this), is there a 
> recommended one?

I have not tested different MPI implementations on exactly the same 
hardware so far. However, the benchmarks on our web page do compare 
ParaStation's MPI + InfiniPath (variant 24) with MVAPICH 0.9.7 + 
InfiniBand (variant 19) and the speedups are almost exactly the same. 
Based on this little data I'd risk saying that the MPI implementation does 
not play a big role (except when you compare Gigabit Ethernet vs. Infini*, 
but that's hardware-related).

Regards,
Jan Ploski