[Haskell-cafe] Fannkuch timings

[Cale Gibbard]

On 10/01/06, Daniel Fischer <daniel.is.fischer at web.de> wrote:
> On my 1.2 GHz Duron (SuSE linux), I get significantly different timings than
> those on the wiki. Sebastian Sylvan's, Kimberley Burchett's and Bertram
> Felgenhauer's all take roughly thrice as long as posted (that's rather
> consistent, but the factor is surprisingly large). Cale Gibbard's takes 18.44
> secs for N = 9, which is rather bad, I think.

Well, yeah, it's known not to be as efficient as the others -- I threw
it together rather quickly from some parts I had laying around. I
unfortunately hadn't tried the naive implementation of permutations
since it's actually faster. (My version ought to still be better than
the original though.)

>
> Really surprising (to me) are the following (N = 10)
>
> Algo                       Wiki          Here
> Clean imperative   2.1 s         4.54 s
> Fastest Pure       1.8 s         8.65 s
> Fastest Impure     1.4 s         1.77 s.
>
> So the ratio for the fast impure version is approximately 1.6GHz/1.2GHz, which
> is natural, but the ratio for clean impure is 2.2 and for the fast pure
> version, it's even 4.8!
> All are compiled with -O2 -optc-O3.
> Does that mean my system is really poor in producing binaries from pure
> Haskell?
> And why would that be so?

Hmm... that's really odd. Which version of ghc/gcc are you using and
how are you measuring the times? Perhaps other processes on your
system are interfering? Do you get these numbers consistently?

 - Cale