Much of the problem with parallel algorithms is that they are machine
specific. There have been attempts at standardization of the parallel
languages (F90, Fortran D, CMFortran), but communication calls are
far from standard. If you are porting code between very different
architectures, often the datasets will need to be restructured to insure
reasonable performance. When you ask about code already written, you
should also inquire as to which platform it was developed on.
PCN , developed between Caltech and Argonne Labs allows for the same
program to run with NO modification on numerous uniprocessor platforms as
well as a number of popular hypercubes. It is available for anonymous ftp
from: info.mcs.anl.edu in pub/pcn
> 2. Is parallel computing on campus something we would use?
In general, if you have a problem that requires a significant amount
of interprocessor communication, a serial machine might give you better
performance.
Hmm, I beg to disagree. I spent a summer 2 years ago parallelizing an
fcc-lattice algorithm to predict tertiary protein structures. My parallel
version (developed with PCN) ran worlds faster.
--
Terry Brannon tbrannon at lion.eecs.lehigh.edu
medical biology via acupunctural bioelectrics
primitive reigns supreme over nearly everyone
