[computer-go] Library of Effective GO routines v 0.106

[Nick Apperson]

The other thought (kind of silly) but just a pssing thought is that the
threads that are benchmarking are keeping a local copy of their count and
writing it back at the end (which is allowed) so you are actually getting
good performance, but you don't know it because the processors aren't
talking to each other properly.  Exactly 2x is hard to imagine happening by
chance.

On 2/22/07, Sylvain Gelly <sylvain.gelly at m4x.org> wrote:
>
> Hello,
>
> > I do not understand it. Maybe someone does?
> > I've made some tests on 2 core processors, and I have strange results.
> > Some of 2 core processors got results exactly 2x times worse than they
> should.
> > Why?
> > I have no idea.
> > But 2.8 Ghz 2 core works exactly like my 1.4 laptop.
> > Also version of g++ does matter.
>
> Here, from my experience, the following can matter a lot:
> - version of g++ (g++ 4.1 gave me +50% against g++ 4.0 on an opteron!)
> - version of the libc: even compiled with a modern compiler, a program
> running on a machine with an old version of the libc can be very
> significantly slower (-30% observed!).
> - exact version of the processor: today the frequency means nothing,
> nor the name of the processor. You have to check the exact numbers.
> And I also observed that even the small numbers as the "stepping
> number", matters. MoGo runs faster on my P4 3.4 Ghz  than on a 3 years
> newer P4 3.8 Ghz, which has also more cache. I ran test on other P4,
> and the slower had a different stepping, that's all (all are dell
> computers using same hardware).
> - Measure of time: if you take the CPU time, on multiprocessor
> machine, while using multithreading, sometimes the reported time is
> not what you expect (generaly reported for all the threads), so when
> you calculate the speed of playouts, you can have a 2x/4x factor.
>
> Hope that can help,
> Sylvain
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20070222/8bbee4b6/attachment.htm