Random number generators

[Steve Kargl]

On Mon, Mar 16, 2015 at 11:22:31PM -0500, Pedro Giffuni wrote:
> Hi;
> 
> FreeBSD libc random functions are not too bad but in general I was having some thoughts about how the random generator functions in libc are slow and predictable and how just about every application nowadays is including "Mersenne Twister"  or similar algorithms (which are fast and better in every way but can?t be adapted for the C API) in their applications.
> 
> OpenBSD did something drastic about it [1], breaking standards and compatibility and whatnot.
> I wouldn?t go there and I don?t think there is any real ?solution? for this. The musl libc guys tried something interesting though. They took the tempering function from MT:
> 
> http://git.musl-libc.org/cgit/musl/commit/?id=20d01d83b5a13c77805976e7c520f566244ba3ff <http://git.musl-libc.org/cgit/musl/commit/?id=20d01d83b5a13c77805976e7c520f566244ba3ff>
> 
> It should be something relatively easy to try on our implementation too, if someone feels like running the tests and measuring if there is a difference.
> 
> Pedro.
> 
> [1[ http://www.tedunangst.com/flak/post/random-in-the-wild
> 
> 

I suppose it depends on what you want to accomplish.  MT
can be a horrible thing to use.  See the history of 
libgfortran/intrinsics/random.c (svn r82443) where I ripped
MT out many years ago in favor of George Marsaglia's KISS generator.
The KISS generator that I used was his 32-bit version.  GM has
a 64-bit generator as well.  The 32-bit version passed all
of GM's diehard tests.  I haven't read a report on the
64-bit generator's diehard result. 

One big issue is saving internal state.  IIRC, MT requires 623-bit
of internal state.  KISS requires 4 32-bit int.  Thus, if
you want to reseed the generator, KISS requires far less effort.

-- 
Steve