Help: About Intel Fortran Compiler:

Jeff Layton jeffrey.b.layton at lmco.com
Mon Oct 6 14:02:39 EDT 2003


Donald Becker wrote:

> On Mon, 6 Oct 2003, Franz Marini wrote:
>
> > On Sat, 4 Oct 2003, Ao Jiang wrote:
> >
> > > ifc -I /opt/mpich-1.2.5/include -Lmpi -w -Lm -o p_wg3 
> p_fdtd3dwg3_pml.f90
> >
> > Try with :
> >
> >   ifc -I/opt/mpich-1.2.5/include -lmpi -w -o p_wg3 p_fdtd3dwg3_pml.f90
> >
> > Btw, a cleaner way to compile mpi programs is to use the mpif90
> > (mpif77 for fortran77) command (which is a wrapper for the real
> > compiler).
>
> Acckkk!!
>
> This is one of the horribly broken things about most MPI
> implementations.  It's not reasonable to say
>     "to use this library you must use our magic compile script"
> A MPI library should be just that -- a library conforming to system
> standards.  You should be able to link it with just "-lmpi".
>

I don't like the mpi compiler helper scripts much either. I just
want a simple makefile or a list of the libraries to link in
in the correct order. I usually end up reading the helper scripts
and pulling out the library order and putting it in my makefiles
anyway (no offense to anyone).

However, in defense of the different MPI implementations,
they have somewhat different philosophies on how to get
the best performance and ease of use. Sometimes this involves
other libraries. Just telling the user to add '-lmpi' to the
end of their link command may not tell them everything
(e.g. they may need to add the pthreads library, or libdl or
whatever).

> One element of a high-quality library is ease of use, and in the long
> run that matters more than a few percent faster for a specific function
> call.
>

One piece of data. While we haven't looked at specific MPI
calls, we have noticed up to about a 30% difference in wall
clock time with our codes between the various MPI
implementations using the same system (same nodes, same
code, same input, same network, same nodes, etc.). I'm all for
that kind of performance boost even if it's a little more
cumbersome to compile/link/run (although one's mileage may
vary depending upon the code)

Jeff


-- 
Dr. Jeff Layton
Aerodynamics and CFD
Lockheed-Martin Aeronautical Company - Marietta


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list