[Haskell-cafe] Fannkuch timings

Donald Bruce Stewart dons at cse.unsw.edu.au
Tue Jan 10 18:13:11 EST 2006


daniel.is.fischer:
> Am Dienstag, 10. Januar 2006 19:11 schrieben Sie:
> > Hello Daniel,
> >
> > Tuesday, January 10, 2006, 7:40:24 PM, you wrote:
> >
> > DF> These are user/MUT times, at the moment, my machine is busy, so that
> > elapsed DF> time is about double that, otherwise these times are rather
> > consistently DF> reproduced (between 8.4 and 8.9 for pure, 1.7 and 1.9 for
> > impure, clean DF> imperative I've done only twice, second run took 4.82s
> > MUT time).
> >
> > on the busy machine even MUT times may be (and even should be)
> > inaccurate because other threads will throw out data in CPU caches
> 
> So that explains (not surprisingly) that MUT times on the busy machine are 
> slightly higher (8.9 resp. 1.9), while on an otherwise idle machine, I 
> consistently got values in the range of 8.4 - 8.7 (I'm not absolutely sure, 
> but I think actually the range was 8.55 - 8.7) resp. 1.7 - 1.78 secs.
> 
> The problem remains, why is the speed-ratio so different for the different 
> algorithms?

Some things I can think of:
    * OpenBSD vs Linux. I've noticed on some benchmarks large (2x) 
      differences between the same code running on the two OSs. Code
      that runs 2x faster on openbsd, compared to and older version of
      the code, will show no speedup on linux.
    * Cpu arch. I'm testing on a Penitium M. This 1.6Ghz Pentium M
      outperforms a 2.4Ghz P4 sitting on my desk.
    * Cache size.
Etc.

-- Don



More information about the Haskell-Cafe mailing list