🚀 go-pugleaf

maniattb@cs.rpi.edu (Bill Maniatty) writes:

>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?)

The most important test for randomness is the spectral test. In essence,
ina p-dimensional spectral test, you take all possible sequences of
p consecutive numbers, and regard them as points in p-dimensional space.
The maximum distance of two p-1-dimensional planes in any direction of
the space is a measure for the goodness of the generator. All good
algorithms have been shown to pass this test reasonaby, all bad
generators seem to fail it. p usually runs about to 1-8 when testing.

For more specifications and some good generators, try Seminumerical
Algorithms (The Art of Computer Programming) by Knuth.

Letting ONE generator run on all machines is no problem, because
applying a linear congruential generator 2048 times can be replaced by
applying a different one only one time. So you start by giving
each generator the next value in sequence of the basic generator
as a seed, then proceed by letting the individual generators
advance 2048 steps at a time. But in 2048-dimensional space a
single LNG is pretty probable to be bad. Maybe the best idea would
be to choose a set of different LNG with relatively prime period,
and use them instead. This would certainly imply having to program
a spectral test to check the generators, or else bad ones will be
there (I know no precompiled tables of 2048 RNGs)