[ Date Index ] [ Thread Index ] [ <= Previous by date / thread ] [ Next by date / thread => ]
On Tue, 5 Nov 2013, Paul Sutton wrote:
Hi A while back it was suggested we try and build a raspberry pi cluster and test it by cacluating Pi. The topic seems to have gone quiet, maybe we can do this at a Pi jam, or something.
This might be a fun project to do - but don't be under the impression that you're making some sort of supercomputer here... This is purely an academic excercise.
My Desktop can compute Pi faster than a small cluster of Pi's My desktop takes 60 watts - that's about 15 Pi's worth of electrical power - would 15 Pi's be faster? Unlikely. The program would also be more complex, although the problem of solving Pi on parallel processors have been "solved", trying to do something similar for other calculations may be somewhat more complex.
The "classic" parallel solver is for raytracing, or something like a fractal where you can use a SIMD type scenario - each node gets the same data, runs the same code, but is told which part of the data set to work on. This works well for data sets like images, etc. but not so well for things like calculating Pi - unless there is an algorithm that says: "Give me digits A to B of Pi" that's efficient...
And the physical wiring will be intersting but not difficult - a single 24-port Ethernet switch will do, and you'll need a big number of power sockets too.
Good luck! Gordon -- The Mailing List for the Devon & Cornwall LUG http://mailman.dclug.org.uk/listinfo/list FAQ: http://www.dcglug.org.uk/listfaq