[Wrf-users] Unexpected "hot-spot" like differences with WRF runs using Intel compiler

[Craig Mattocks]

Hi Ligia,

I wonder what would happen if you define two custom precision Fortran kinds:

!----------------------------------------
! Define a KIND with 6 digits of accuracy
! and exponent range of 10^-37 to 10^37
!----------------------------------------
integer, parameter :: sp = selected_real_kind(p=6, r=37)

!-----------------------------------------
! Define a KIND with 15 digits of accuracy
! and exponent range of 10^-307 to 10^307
!-----------------------------------------
integer, parameter :: dp = selected_real_kind(p=15, r=307)

and use them in your pendulum simulation.

Granted, you will still face issues with intrinsic functions and MPI implementations, but it might help improve consistency across platforms.

Best,
Craig

On Wed, 13 Jun 2012 16:32:39 -0600 Ligia Bernardet <ligia.bernardet at noaa.gov> wrote:

> Yongtao,
> 
> Forecasts will be different when the same code is run on different computational platforms. Even different compilers, or different compiler flags in the same platform will yield different results. 
> 
> Chris Harrop and I did a study to quantify the differences in forecast when several models are run in various platforms (http://www.mmm.ucar.edu/wrf/users/workshops/WS2011/Extended%20Abstracts%202011/P52_Harrop_ExtendedAbstract_11.pdf).
> 
> It is very difficult to know when the differences one encounters are "normal", just caused by changing platform, or when they are the result of a bug. Different compilers will treat un-initialized variables in different ways: setting them to zero or to a random number. But since you used ifort v10.1 in both platforms, this should not be the issue you are facing.
> 
> Regards,
> Ligia