> > However, the FFT itself is very amenable to parallel processing
> > techniques - on a processor with N independent compute pathways, you
> > can compute N elements in the same time that a single element would
> > take to compute just one.
> >
> How much of the mersenne-test is used for the FFT then?
> Isn't the FFT the single most time-consuming part?
>
> Wouldn't it make sense to make a parallelized version of prime95 that
> distributes the FFT over N processors?
I've been down this road before. The best parallelization you can do is to
run multiple instances of prime, one for each CPU. Otherwise, it wouldn't
be an optimal use of resources.
If one CPU takes 2 weeks to do an LL test, 2 CPU's working on it in parallel
might get it done in 1.5 weeks. On the otherhand, if you run 2 instances,
you get *2* numbers done in *slightly* over 2 weeks instead of just 1,
effectively doubling your efforts.
I say slightly over 2 weeks since Prime95/NTPrime running on multiple CPU's
will compete for memory bandwidth and I've seen it slow down to about
96%-98% of it's speed compared to when it's running on only a single CPU.
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
