Re: The new scheduler and x86 CPUs

[Vladimir Makarov <vmakarov at redhat dot com>]

Toon Moene wrote:
> 
> Vladimir Makarov wrote:
> 
> > As for the data, I see the following
> 
> > http://www.spec.org/osg/cpu2000/results/res2001q2/cpu2000-20010604-00681.html
> 
> > Alpha 833Mhz: base : 643
> 
> > http://www.spec.org/osg/cpu2000/results/res2001q2/cpu2000-20010522-00660.html
> 
> > Itanium 800 Mhz: base: 701
> 
> > If 342 for Itanium were for gcc, I would not be surprised.
> 
> <Shrug>  Look again.  The 342 for Itanium vs. 544 for Alpha is for the
> _apsi_ program.
> 
> This makes it (again) clear why it is so dangerous to *only* look at the
> bulk number for SPEC, something that SPEC keeps pointing out.
> 
> The reason why it is interesting to look at a weather forecasting
> program (actually, apsi is a dispersion model based on a weather
> forecasting model) is that these programs tend to *not* have a single
> hotspot that you can easily design your compiler for.
> 
> Of course, this doesn't rule out that OOO implementations wouldn't run
> out of steam eventually against newer paradigms, but the Itanium is
> definitely not proof of that.

  I don't know why this test is better on Alpha.  But I think this is
not merit of OOO.  May be because Alpha has a bigger L2 cache (8Mb vs.
4Mb of Itanium).  Or may be because Compaq used KAP optimizer.  Intel is
a novice in high performance computing.  Compaq (former DEC) are one of
the most experienced in this field.  They was the first which adopt
Fortran90 (actually they bought a company for this).  I agree that if
you are going to buy computer for given task, you need a solution not a
potential.

  Actually I did not want to start a war about what architecture is
better.  As probably you, I like diversity and do not like a monopoly.