[Python-3000] C API for ints and strings

[Nicholas Bastin]

On 9/12/07, Gregory P. Smith <greg at krypto.org> wrote:
> [cpu rant]
> Off topic: not true.  The Pentium D is the final Pentium 4 netburst
> architecture based design.  It is not at all close to the Pentium M.   The M
> is much more a derivative of the pentium pro,ii,iii, & iii-m before it as
> core and more distantly core2 are follow ons to the M.  Yes the D (50xx) and
> Woodcrest core2s (51xx) shared the same socket and front side bus but
> internally they are unrelated.
> [/cpu rant]

Yeah, my mistake, I misread intel's NetBurst page.  I should have
stuck with Wikipedia (who knew).

> Regardless comparing between different cpus doesn't matter, only the
> difference between runs on the same cpu.

I agree.

> for instance on a 1.4Ghz efficeon:
>
> python2.5:
> 10 loops, best of 3: 932 msec per loop
> python 3.0a1 svn trunk:
> 10 loops, best of 3: 1.54 sec per loop
>
> (both compiled with gcc 4.1.2 -O3)
>
> which falls right smack in the middle of the measurements others were
> reporting in this thread. ;)

I should look at a comparison of 2.5 and 2.6 at some point, for better
reference.

> > Without looking into it at a much lower level,
> > it's hard to tell, but the difference between a 1MB and 2MB L2 cache
> > might make all the difference in 3.0 performance.
>
> doubtful, python's ceval core and the data representing the code being
> executed are both tiny.

Makes me miss the G4/5 version of Shark on MacOS X, which would show
you the pipelining in the processor and cache utilization, so you
could actually see what was going on - the x86 Shark doesn't seem to
have this capability.  It's suitably interesting on my windows xp
machine to turn processor affinity off and see the performance go to
hell in a handbasket.  (Why, oh why, does windows insist on moving
processes between CPUs all the time?)

--
Nick