null character without reason in written files

[N.M. Maclaren]

On Oct 31 2013, Janne Blomqvist wrote:
>>
>> Try Marsaglia's DIEHARD tests.  I have some that are rather more
>> stringent, though they test different aspects, and it doesn't fail
>> them.  I haven't run DIEHARD on it, but I assume that somebody has.
>
>I haven't tested myself, but I recall skimming through some paper
>testing various PRNG's, and IIRC KISS (which is what gfortran uses)
>passed DIEHARD, DIEHARDER(?), and TEST_U1(sp?).

Oh, yes, it would.  He wasn't sloppy that way and wouldn't publish anything
that failed his own tests.  The only questions are whether gfortran uses
the right version and has coded it well.  Marsaglias's main flaws were that
his coding style was out of the ark (think Fortran 66), and he was very
late in realising the importance of higher precisions.

I didn't really publish the results, though they are implicit in another
paper, but a precision of N bits at ANY stage (e.g. converting the
composite seed to the result) means that the generator is suitable for
simulation sizes of only (say) 2^N/N.  The failure mode is that the
adjacency (spacing) properties fail, which is really serious for some
applications.

I have some tests for the effect (please ask if you want a copy), and
have tested gfortran up to 10^7.  I might kick off 10^8 tonight and go
to bed, but it may not finish before I get bored and kill it - it's a
cache thrasher :-)


Regards,
Nick Maclaren.