On Tue, 2002-11-05 at 17:38, Thomas Dodd wrote:
>
>
> Jean Francois Martinez wrote:
>
> > I wrote a numerical analysis who was published in february 1999 and
>
> Is that available online?
>
> > basically you don't gain 5%. It is closer to 1.5% to 2.5% and this
> > when you are running kernel code (when you are running user mode code
> > it is unaffected). And since 2002 machines are much bigger than the
> > 32 meg box I used as an example the memory savings brought by a kernel
> > compile are still more irrelevant
>
> What compiler? gcc-3 is supposed to have better optimization.
> Perhaps the numbers would be different now?
>
Yesterday I told some things from memory about compiler settings for K6
and it turned that my memory failed me.
I reran the tests with gcc 3.2 and basically at -O2 optimization
-march=k6 is overall a little better than -mcpu=i686 despite losing for
the integer tests and losing heavily at bit operations (not as heavily
as I told). But -march=k6 provides better memory bandwidth. It is the
floating point tests who end giving a small edge to -march=k6.
At -O3 optimization -mcpu=i686 beats -march=k6.
Now kernel does not use floating point and it should not be compiled
with -O3. :-)
JFM
