Benchmarking wrf.exe

[PeterJamesRye]

With a little spare time on my hands, I have just compared compilations of WRF 4.1.2. using Intel fortran and Gfortran 9. I was aware that Gfortran has made a few important enhancements, starting from version 7, but the result was something of a surprise.

In a set of non-nested runs, using the SMPAR model, the difference was small - under 1%, positive or negative depending on the way the data were averaged. But using a doubly-nested run, Gfortran proved to be around 3-4% faster on average. For a free compiler, they're both pretty hot results.

If anyone wants to check the results, the Intel compilation used "-Ofast -march=native", while Gfortran used "-O3 -ffast-math -funroll-loops -march=native" (because the -Ofast compilation did not run stably).

Does anyone have comparable comparisons to report?

 

[kwerner]

Hi,
That is very interesting. The last time I ran a similar test was with GNU V6.3.0 and Intel 17.0.1, which resulted in Intel running much faster (although GNU does compile significantly faster). My tests were with dmpar builds, though. Thank you for conducting this test. If GNU is comparable (and even faster) than Intel, that is big news!

 

[alainF]

Hello PeterJamesRye,
Could you confirm your results on some other cases ?

Thank you,