On Jun 1, 2005, at 3:29 AM, Iago Vilar Neves wrote:
> I am using lam-mpi and fftw 2.1.5 to convert in parallel an aplication
> wich uses normal fftw (monoprocessor). The problem is that the results
> of fftwnd_mpi() are diferents on each number of process. Only for one
> processor is equal than the original monoprocessor version.
How different are the results? Due to floating point round-off and
ordering differences, it's not uncommon to see slight variations in
results as number of processors is varied. This isn't an MPI problem,
but the nature of working with floating point numbers. If you're
really unlucky, you can even have ordering changes in floating point
operations cause huge differences.
So, if you are seeing small variations, it's probably just ordering
issues. You should verify your results, but are probably ok. If you
are seeing larger variations, outside of the tolerances of your system,
you should really talk to the developers of the MPI version of fftw -
they would be better suited to answer your questions.
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
