Random Number Generators: Performance

[Lars Segerlund]

 Would you you mind having a look at the Mersenne twister in libgfortran ? Do you think it's up to the job ?

 I choose the MT since it has a VERY big period and a not too big seed, coupled with a performance which is right 'up there'.

 You could also try 'ent' ( a program to calculate the entropy of random bit sequences ) on the different algorithms, when I did this it did suggest that MT had a good entropy measure for any size bitstream.

 / Lars Segerlund.


On Sat, 24 Apr 2004 10:10:05 -0400
Scott Robert Ladd <coyote@coyotegulch.com> wrote:

> As someone whose work is reliant on reliable, long-period random number 
> generators, I've collected the several top-flight algorithms, including 
> the Mersenne Twister and several suggested by George Marsaglia. You can 
> find the implementations of these in my Coyotl library, available from 
> the website in my sig below.
> 
> On a Pentium 4 (Northwood), I get the following performance numbers:
> 
> mtwister =   13.87 (72105154 rn/sec)
>   mwc1038 =   15.65 (63913964 rn/sec)
>   kissrng =   17.64 (56692771 rn/sec)
>    mwc256 =   16.84 (59386090 rn/sec)
> cmwc4096 =   13.51 (74011471 rn/sec)
> 
> On an pure 64-bit Opteron system, the numbers are quite different, due 
> to the availability of 64-bit processor instructions:
> 
> mtwister =   15.89 (62952145 rn/sec)
>   mwc1038 =    9.73 (102786532 rn/sec)
>   kissrng =   10.36 (96536781 rn/sec)
>    mwc256 =   14.97 (66808050 rn/sec)
> cmwc4096 =   10.08 (99243210 rn/sec)
> 
> I suspect the twister could be improved on a 64-bit platform by 
> converting the array of 32-bit seeds into one using 64-bit values.
> 
> Relative performance is, as shown above, highly platform-specific. The 
> C++ code was compiled with -O3 -march={pentium4|opteron}, using GCC 
> 3.4.1 from CVS. (Tree-ssa does not bootstrap this morning due to a bad 
> patch.)
> 
> In terms of general overall performance, cmwc4096 looks to be a good 
> overall choice, save for it's use of a 4096-element array of 32-bit 
> integers. KISS is a short period generator, but requires only four 
> 32-bit state variables.
> 
> I'm working on parallel random number gereators, but that is another 
> tale. My feeling is that RANDOM_NUMBER should a general algorithm that 
> performs well across platforms, and we ignore thread-safety, providing 
> an extension random number generator suitable for parallel applications.
> 
> -- 
> Scott Robert Ladd
> Coyote Gulch Productions (http://www.coyotegulch.com)
> Software Invention for High-Performance Computing
>