>Parallel FFTs typically have three steps: 1) crunch like mad on local
>data, 2) global scatter-gather (everybody trades data with everybody
>else), 3) crunch like mad on local data. In the GIMPS case, you want
>an inverse FFT too, so that the order is 12321. With the real FFTs that
>Prime95 uses it's worse...step 3 has some global communication too.
>The communicaton is much more a problem for systems with distributed
>memory, but since speed is of the essence here the communication means
>cache thrashing. Running two processors in parallel will speed things up,
>but I would think that the speedup wouldn't be as great as running two
>copies of Prime95. You could probably get some cool single machine
>statistics though :)
Wouldn't this help tremendously the problem of having to do enormous 10-million
digit primes?
I'm guessing, a client hasn't been attempted, but I'd certainly love to have
one. I have
way too many machines that are Dual and Quad Processor Intels that could
seriously take
advantage of this.
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
