Newsgroups: comp.parallel
From: nmm1@cus.cam.ac.uk (Nick Maclaren)
Subject: Re: Random numbers on Paragon
Organization: University of Cambridge, England
Date: 23 Oct 1995 14:48:42 GMT
Message-ID: <46ga0a$375@usenet.srv.cis.pitt.edu>

In article <4663n7$4ve@usenet.srv.cis.pitt.edu>, Paal Skjetne <skjetnep@stud.unit.no> writes:
|> 
|> My question is simple, is there a parallel random number out
|> there? I need my numbers to be random, on each node and between
|> nodes. Please no dirty tricks such as picking Nproc random seeds.
|> I need to know that the sequences really are random, and independent.

Yes, there is, but remember that 90% of the papers on this topic are
just plain nonsense.  Most of them use 'independence' to mean
'disjointness'.  De Matteis and a few other people have done some
serious analysis of the cross-processor interactions.  For a better
approach, look at the following papers:

N.M. Maclaren,
The Generation of Multiple Independent Sequences of Pseudo-Random
Numbers (Journal of the Royal Statistical Society (1989), Series C
(Applied Statistics), Vol. 38, no. 2, pp 351-359).

TI: RANDOM NUMBER GENERATORS IN SUPPORT OF MONTE-CARLO PROBLEMS IN 
    PHYSICS
AU: DYADKIN_IG
NA: VNIGIK,RUSSIAN FED LOG RES INST,CHAYKOVSKOGO 28-2,TVER 
    170034,RUSSIA
JN: NUCLEAR GEOPHYSICS 1993 Vol.7 No.3 pp.419-429

There is an implementation of my method in one of the NAG parallel
libraries, and I am currently (if slowly) working on a hybrid version
that will achieve excellent efficiencies on both vector and parallel
systems.  I have been able to get it up to 35,000,000 numbers a second
on a (single vector processor) Hitachi S3600/180.


Nick Maclaren,
University of Cambridge Computer Laboratory,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email:  nmm1@cam.ac.uk
Tel.:  +44 1223 334761    Fax:  +44 1223 334679

