Re: Gfortran vs Ifort when using MPI

[Janne Blomqvist <blomqvist dot janne at gmail dot com>]

Johan Seland wrote:
> A follow up to my earlier posts:
>
> Profiling shows that the time difference between "very slow" and "full
> speed" is all spent in MPI synchronization routines.
>
> Let me first summarize my findings:
>
>     * Simulator runs at full speed on Ubuntu Hardy 32-bit using
> repository MPICH and gfortran.
>     * Simulator runs at full speed on Gentoo 32-big using repository
> versions of MPICH and gfortran.
>     * Simulator runs very slow on Ubuntu Hardy 64-bit using repository
> MPICH and gfortran
>     * Simulator runs at full speed on Hardy 64-bit when MPICH (not the
> simulator itself!) is compiled with Intel Fortran (ifort).
>        It does not matter if the simulator itself is compiled with
> ifort or gfortran.
>     * Simulator runs very slow on Ubuntu Hardy 64-bit when MPICH is
> compiled using gfortran.
>        It does not matter if the simulator itself is compiled with
> ifort or gfortran.
>
> These results are reproducible for several machines, both dual and quad core.
>
> I am not yet certain if this is due to a bug in Gfortran, MPICH or Linux.

Interesting.. When you say the "fast 64-bit MPICH" is compiled with 
ifort, do you mean that the C parts of MPICH are compiled with gcc, and 
only the Fortran interface with ifort, or do you mean it's compiled with 
icc + ifort? AFAICT there's not much Fortran code in any MPI 
implementation, the actual code tends to be C with only a very thin 
Fortran wrapper.

Have you tried a different MPI implementation, such as Open MPI?

Or does "use mpi" vs. "include 'mpif.h'" make a difference?

> Is there any significant difference between 64-bit and 32-bit versions
> of Gfortran? (I would not think so?)

As far as the frontend is concerned, not really. Some differences in 
sizes of some types, e.g. array indexes are internally 64-bit in the 
64-bit version.

--
Janne Blomqvist