Beowulf & FFT

[Josip Loncaric]

Martin Siegert wrote:
> 
> The authors of FFTW claim that this is the fastest FFT you can get. 

I second that.  The FFTW library runs great on a wide range of
machines.  However, FFT is a global transform, and it requires
all-to-all communication at some point.  If you can break your problem
into spectral "bricks" then you can reduce the amount of communication
to some boundary condition between these "bricks".  This gives you a win
thanks to a favorable surface/volume ratio, but the internal boundary
conditions are typically of lower order of accuracy.  While you will not
get spectral accuracy, in most applications going beyond order 6-10 does
not help much.  I believe that A. Patera (MIT) had some papers on such
methods...

Sincerely,
Josip


-- 
Dr. Josip Loncaric, Senior Staff Scientist        mailto:josip at icase.edu
ICASE, Mail Stop 132C           PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center             mailto:j.loncaric at larc.nasa.gov
Hampton, VA 23681-2199, USA    Tel. +1 757 864-2192  Fax +1 757 864-6134