[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Poor performance of G5 on GNOME's CPU benchmarks.

On Mon, 2008-08-04 at 10:24 +0900, Charles Plessy wrote:
> Le Sun, Aug 03, 2008 at 05:32:26PM +0200, Michel Dänzer a écrit :
> > On Sun, 2008-08-03 at 11:53 +0900, Charles Plessy wrote:
> > > 
> > > The result is that on some particular types of computations, the G5
> > > performs extremely bad: something like twice slower as an old 1.5 Ghz
> > > Celeron machine. For some other tests, the performances of the
> > > processors scale with the frequency.
> > 
> > Most likely, the x86 binaries tend to have optimizations for SSE/MMX
> > whereas the powerpc binaries tend not to have VMX/Altivec optimizations.
> Hi Michel,
> Is it somthing that can be turned on, or does it require modifications
> to the sources? Quoting a GCC webpage:
>   "The interface is made available by including <altivec.h> and using
>   -maltivec and -mabi=altivec."
> http://gcc.gnu.org/onlinedocs/gcc/PowerPC-AltiVec-Built_002din-Functions.html

These flags are prerequisites for allowing the resulting binaries to
contain Altivec code at all, but on their own they don't cause any such
code to be generated (except maybe for some simple cases like memory
copies). Current versions of GCC can automatically vectorize some code
with -ftree-vectorize, but the results probably still can't compete with
hand-optimized code.

Also note that these flags will make the resulting binaries unusable on
systems without Altivec support.

> The case in which I saw the biggest difference if for the building of
> the 'exonerate' package: it took 9 h on a bi-G5 2GHz, compared to 1 h 30
> on the i386 buildd, 30 min on the alpha buildd, and > 30 h on the MIPS
> buildds.

Not sure that's directly related... I suspect this could be rather due
to different amounts of RAM. (Even so though, it seems weird the alpha
buildd could be so fast...)

Case in point:


says 04:17:30, and voltaire is an old dual G4 500 MHz... Though


also says it only has 320 MB of RAM...

BTW, did you pass -j2 to dpkg-buildpackage and/or set
DEB_BUILD_OPTIONS="parallel=2" to try and make the build use both cores?

> I am starting to wonder if some of the binaries we ship in Debian Med
> for bioinformatic computation are useful on PowerPC. It must obviously
> depend on what kind of calculation they are doing. Is there a way to
> guess other than by benchmark?

Not sure, but I suspect not - in general, the only way to be sure if a
given setup is suitable for XYZ is to try XYZ on the given setup. :)

Earthling Michel Dänzer           |          http://tungstengraphics.com
Libre software enthusiast         |          Debian, X and DRI developer

Reply to: