Re: Portable random number generator

From:

"osmium" <r124c4u102@comcast.net>

Newsgroups:

comp.lang.c++

Date:

Wed, 10 Nov 2010 12:19:12 -0600

Message-ID:

<8k0610Foc3U1@mid.individual.net>

"Victor Bazarov" wrote:

On 11/10/2010 10:11 AM, osmium wrote:

"Victor Bazarov" wrote:

On 11/9/2010 11:12 PM, Gus Gassmann wrote:

I am collaborating on a rather large project with complicated XML
files (some objects nest ten levels deep) and corresponding data
handling challenges. I want to do proper testing of the (C++) code and
decided that the only way to go is random tests. So now I am looking
for a random number generator with the following properties:

1. Portability.
2. Random starting points.
3. Replicability on demand.

I presume this means that I would seed the RNG based on the clock, but
keep a copy of the seed that I could optionally use at the start in
case I found a problem on a previous run.

Statistical properties are of lesser importance.

I presume I am not the first person to attempt this and am hoping to
find some guidance here. Both C and C++ would be OK for the RNG, hence
the cross-post.

Thanks for any hints.

I am afraid to ask... Why not use 'srand/rand' pair of functions? You
can always do

      unsigned seed = (unsigned)time(0); // for keeping
      srand(seed);
      ...

Of course, it's so damn obvious that I expect some kind of a trick...

I assume by portable he means the same sequence would be provided from a
given seed regardless of the target computer.

Ah... I didn't assume that. But it might be valid, if he wants to run the
failed tests on a different platform than the one on which they fail.

rand() gets an F on that.

In that case he just needs to take the source of the 'rand' from any
standard library implementation that he can lay his hands on, and stick it
into his source. From what I gathered of the requirements the evenness of
the distribution or the length of the cycle aren't high on his list.

I don't think that works. Many of the generators that I have looked at have
built in assumptions; computations are assumed to be mod 16 or mod 32, for
example based on the (PRNG) programmer's presuming a certain platform.
These are implicit assumptions and don't show up in the code. For example,
I wouldn't be happy if I saw a simple variable of type int in the code,
since int has no particular meaning.

From the general tenor of the original request, I think he wants testing on
various platforms to exhibit consistent results.