Re: [Fwd: Re: random variate from power exponential distribution:continue]

[Brian Gough <bjg at network-theory dot co dot uk>]

Giulio Bottazzi writes:
 > Do you think that, in a hypothetical rewriting of the actual
 > implementation of the power exponential random variable generator in GSL
 > (randist/exppow.c), one should take in consideration the different
 > architectures, maybe via architecture-specific pieces of code toggled by
 > preprocessor directives? Or it would be better to try to cook up a
 > "reasonable" routine valid on "the average machine" (whatever that
 > means... probably nothing)
 > 
 > I would be happy if someone can provide me with some opinion/suggestion
 > on this point.

Have you quantified how much of the variation is due to the underlying
RNG, and how much is due to the randist algorithm?

I would suggest doing profiling with gprof or instruction-level
profiling, in addition to measuring total runtimes.

For GSL, the goal is to have simple generic algorithms that perform
reasonably on any machines.  Optimising for one platform tends to make
things worse on others, as they have different strengths and
weaknesses.

-- 
Brian Gough

Network Theory Ltd,
Commercial support for GSL --- http://www.network-theory.co.uk/gsl/