🚀 go-pugleaf

In article <dak.716387354@tabaqui>, dak@tabaqui.informatik.rwth-aachen.de (David Kastrup) writes:
|> daveg@synaptics.com (Dave Gillespie) writes:
|> 
|> >In article <h_8yzkk@rpi.edu> maniattb@cs.rpi.edu (Bill Maniatty) writes:
|> >> I have a large distributed stochastic model (runs on MasPar MP-1 with 2,048
|> >> processing elements).  I'm not an expert on random number generation, but
|> >> I need to simultaneously g enerate different ``Random'' sequences on each
|> >> node.
|> 
|> >If the nodes must be independent, then it would probably work pretty
|> >well to use one random number generator, say, linear congruential,
|> >to create a sequence of 2048 seeds for a different type of generator,
|> >say additive congruential, running in parallel on the various nodes.
|> 
|> This idea is really bad, because it ensures that no two generators will
|> give the same number at one moment (or even always the same). Besides, random
|> number generators are usually not designed to produce uncorrelated
|> numbers at ARBITRARY distances in their sequence, so you would probably
|> fail even simple tests terribly. As I said before, the only mathematically
|> safe way is to use ONE (and only one) generator of very high quality to
|> produce values in turn. This is because the mathematics for a good RNG
|> can only be ensured with regard to itself, not to other instances of it
|> (different seeds), and certainly not to different RNGs without further
|> information.

The problem with this solution as I see it is that it eliminates any parallelism
I might have in my program.  Please remember that there is a LARGE number of
random number generated (2048) for each time step.  It seems that making each
time step wait for all 2048 random numbers to be generated is terribly
inefficient (wait to generate 2047 random numbers never used by this processor).

Can you tell me just what sort of metrics are used to determine ``randomness'',
and suggest some solutions/references (can a sequential algorithm be extended
to operate in parallel? can a parallel algorithm be developed directly with any
reasonable degree of confidence?)

Thanks for your help,

Bill
-- 
|
|	maniattb@cs.rpi.edu - in real life Bill Maniatty
|