Gentoo Linux faster app-load than (Mandrake|Fedora)

[Jean Francois Martinez]

About Gentoo and the processor optimizations.   I used the nbench-byte
benchamrk on a PIII.  Compiling with -march=i686 instead of -mcpu=I686
(the parm used by RedHat) gave me a 35% speedup... on one of the integer
tests, the other six integer tests remained the same so overall
improvement was 5%, in the floating
point tests I got a 3% overall improvement. 

Restrictions:

1) I could have gone farther with exotic alignment options and sse
(in fact I tried, got a little additional improvement but  less than
above)

2) On an athlon I get more boost from the additional instructions
allowed by -march=i686. I get still an additiional (slight) increase if
I compile with -march=i686

3) The increase is somewhat flattenned because the test spends time
on glibc and this didn't change.  On a PIII the glibc of gentoo would
provide no benefit since the glibc of Redhat is fully optimized for the
PIII (with -march=i686).  On a PIV or Athlon RedHat will run slower than
Gentoo due to its glibc who is not optimized for them.


About loading times:  Too many people mix loading speed with "cruise
speed".  It is very easy to have a program load very fast: the configure
script will look if some libraries needed for additional functionality
are installed, if they aren't it will continue but will build a 
crippled binary: a samba who cannot use certain authentification
methods, a gimp who will be unable to read anything but Jpegs and so
on.  You can do additional curtailing with parms to configure.  Since
there is less to load, and less initiliazation to do the crippled binary
will load blindingly fast and that way an unscrupulous/demagogic distrib
vendor can give the impression its distro is ultra fast.  Of course one
day you will miss that functionality who was left out.

Linus loading times versus Window 98 loading times.  I have found that
Word/Excel/Explorer load much faster on a PIII/667 than OpenOffice and
Galeon on machines who are two or three times more powerful (a
PIII/1000, an Athlon 2000).  In part this is due to the loading of
dynamic libraries being much simpler on W98.  I have tried some
tricks to impprove loading times on Linux like reordering members
of the libraries so the ones who are more frequently used are grouped
together.  It didn't provide an appreciable improvement.  But something
has to be done about these loading times.

On Thu, 2003-10-09 at 19:25, Rahul Karnik wrote:
> Mark Mielke wrote:
> > I guess this issue is important to me because I find it a little
> > embarassing that my wonderful 'linux desktop' at work takes 3X as long
> > to start my web browser, and 10X as long to start my word
> > processor.
> 
> Actually, look at the Gentoo forums. There are plenty of cases where the 
> binaries provided by upstream prohects are faster than the "optimized" 
> ones compiles by Gentoo users (mozilla and openoffice.org, IIRC). For 
> large programs, app load times are *probably* best when compiled with 
> -Os, something that is rarely done by Gentoo users. Typical Gentoo 
> CFLAGS are "-O2 blah" or "-O3 blah". In my experience, -O2 yields the 
> best results, but I would go with upstream guidance every time.
> 
> As for processor optimizations, I am willing to bet that they do not 
> affect app load times -- places where they might matter are 
> computationally intense applications like MPlayer or scientific computation.
> 
> Pardon the ramblings,
> Rahul
> --
> Rahul Karnik
> rahul at genebrew.com
> http://www.genebrew.com/
> 
> 
> 
> 
> --
> fedora-devel-list mailing list
> fedora-devel-list at redhat.com
> http://www.redhat.com/mailman/listinfo/fedora-devel-list
-- 
Jean Francois Martinez <jfm512 at free.fr>