(j3.2006) (SC22WG5.4107) [ukfortran] Standard intrinsics and coarrays
Thu Oct 22 02:23:33 EDT 2009
>>> It is processor dependent whether all images use a common
>>> generator or whether each image uses a separate one.
>> Perhaps we should require that it should be one or the other.
> If we require a single generator,
I meant exactly that we should require it to be "one or the other", not that we
should pick which one ourselves.
> If we require a single generator, it's difficult to get the same sequence on a
> particular image on consecutive runs, which is useful (some would say
> indispensible) for debogging,
Unless the generation of random numbers on each image is independent of the
results of the calculations or the speeds thereof on other images, it's
impossible anyway. So you're going to be stuck in the bog.
> It looks like Hobson's choice, but we ought either to choose, to provide a
> mechanism for a program to choose, or require the processor to provide a
> mechanism, outside the language, to choose.
If the user desperately needs one or the other, he probably also needs various
guarantees which we don't make about RANDOM_NUMBER (e.g. a lack of correlation
between the sequences on each image) so will be using his own anyway. Depending
on what he is doing, the qualities he wants from a PRNG will be different, and
some of those qualities are mutually incompatible.
I see no reason to punish the ordinary user by requiring a suboptimal choice of
RANDOM_NUMBER on one subset of machines. Processor dependent is the right
choice for the ordinary user, and the expert (or the person who needs an expert)
will just have to read the manual or use a third-party product.
................................Malcolm Cohen, Nihon NAG, Tokyo.
More information about the J3