Beowulf Questions

[Bryce Bockman]

 From a theoretical standpoint I can't add too much to this debate, but
when I was writing an extension to Mathew Walls GALib (a genetic
algorithms library) I used MPI to implement a collection of "Island
Model" genetic algorithms.  Initially, I took no steps to make sure that
the random number pools didn't overlap other than using different seeds
(node id along with time) on each node.  Mathew's library implements the
famous ran2 from numerical recipes.  However, when I finally did
implement a leapfrog method for guaranteeing disjoint sets, my solutions
began to converge much more quickly.  So my guess is that making sure
that I had unique pools did help in my case.  Clearly GAs are one of the
more RNG intensive applications, but I think there is a case for using a
leapfrog method on small clusters or if your random number generator
permits it, a look ahead technique that allows you to partition your
pseudo random number space on a per node basis.  Or use of the SPRNG
library which was mentioned earlier.

Cheers,
Bryce

>I was assuming a sophisticated RNG.  With such, the likelyhood of
>identical seeds is very low, exactly the same as correlation within the
>number stream.  Anyone that needs a cluster to generate random numbers
>will be far beyond using a LFSR with a small seed.
>
>I'll even put forth a hand-waving argument that multiple machines will
>be working from a much richer entropy pool, and thus generate better
>quality numbers.
>
>
>  
>