On Mon, 22 Mar 1999, Aaron Blosser wrote:
> Subject: Mersenne: Can NTPrime be coded for MP?
>
> I have a few quad processor PPro machines that I run NTPrime on.
>
> Currently, they're set up to run 4 instances of NTPrime with affinities from
> 0-3, and this works just fine.
>
> Now, I'm sure George has thought of this, maybe, but wouldn't it be fun to
> have a version of NTPrime that was capable of multi-processor computations?
>
> Does the algorithm lend any good ways to do this?
>
> The reason I thought of this was that if this is possible, would it not also
> be possible to use multiple computers on a network to work on the same
> number? With Windows OS', you could use DCOM or even just RPC to get many
> machines working on the same number.
>
> I thought perhaps it might look something like you assign a "group" name in
> prime.ini, identifying that PC as belonging to a group of others all working
> in tandem.
>
> It might be too slow for the Internet since the dataset we're talking about
> can be big, but for a LAN, I think it's reasonable, and certainly a good
> idea for multiprocessing on the same machine.
>
> I just don't know enough about the LL algorithm to see for myself whether it
> can be scaled to MP in any useful way. Basically I think it'd be fun to
> hook my 32 current PC's together and have them crank out a single LL test on
> an exponent above 7M in less than a day. It'd make a great double-checking
> "engine" when we find the next prime. We could verify it in hours rather
> than weeks. :-)
>
> Just a thought, maybe someone has more thoughts on this.
>From what I know, FFT can be effectively handled by multiple processors
with a shared memory abstraction, which should make it possible to greatly
accelerate the LL test in a single SMP machine.
On the other hand, the shared memory abstraction is really crappy
speedwise when implemented in a clustered situation (memory access gets
something like 4 orders of magnitude slower even if you use dedicated
100base-t links between the machines), so that wouldn't lend itself to
parallellation nearly as well.
My suggestion would be to forget about clustering the machines, and
concentrate on either redoing the FFT for SMP or automating multiple
simultaneous tests on the same machine.
--
Henrik Olsen, Dawn Solutions I/S
URL=http://www.iaeste.dk/~henrik/
Get the rest there.
________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
