On Fri, 5 Dec 1997, Henry Zongaro wrote:
> XLF uses the multiplicative congruential method to generator random
> numbers. The algorithm wasn't mentioned in our manuals for XLF 3.1, but it
> has appeared in the manuals (under RANDOM_SEED) since XLF 3.2.
>
> Two generators have been available since XLF 3.2, both well-tested.
> Which is used can be selected by specifying the non-standard GENERATOR dummy
> argument to RANDOM_SEED - the values permitted are 1 and 2. The first
> generator has a period of 2**31-2; the second has a period of 2**46.
> If you need any more details, please let me know.
I have two questions, one concerning the XLF generator, the other
about the f90 subroutine RANDOM_NUMBER in general.
1) Has the _double precision_ version of the generator been tested as
well? I am asking this because I knew so far only one generator in
double precision, and the availability of such a generator is
mandatory when sampling very rare events. That implies a good
"randomness" of all digits, even the least significant ones, and
testing at this level is probably not straightforward.
2) Has anybody thought of the portability problem between different
platforms? Here, unlike for other intrinsic functions, different
algorithms may give _very_ different results, and not only
different numerically but with a very different level of randomness.
For this reason, up to now I prefer to use an external generator
(in F77), whose source I can transfer between different machines.
Is there any hope of standardization? (I guess not)
A. Fasso'
--------------------------------------------------
Alberto Fasso'
Stanford Linear Accelerator Center
Radiation Physics Dept. ms 48, P.O. Box 4349
Stanford CA 94309 (USA)
Tel. (1 650) 926 4062 Fax: (1 650) 926 3569
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|