Linux Benchmarking

[Andrew Cowie]

On Tue, 2005-23-08 at 06:05 -0400, Scott Allen wrote:
> You have to watch Gentoo. Being a source based system that makes it 
> easy to choose the optimisation flags, [but] a "best-tweaked" system
> can also become unstable.

Big time.

Conversely, having been conservative in both my CFLAGS and in choosing
when to do upgrades (ie only when I need them, allowing dependencies to
evolve slowly as a result, and only running bleeding edge packages when
I know what I'm getting into) I have had a number of extremely stable
systems.

You really don't want anything more than:

CFLAGS=-02 -pipe -march=pentium3

O2 is more than a sufficient optimizations (not to mention generally
common in other distros so quite safe) whereas I *have* found that the
ability to specify the processor architecture results a fairly
significant gain.

[That's subjective, but compared against the Debian installation that
proceeded the Gentoo installation on one particular system, switching
from generic x86 machine code to machine code that included Pentium3
specific instructions gave me a 5% or so boost on the tasks for which I
had reasonably accurate time experience. <shrug>]

Trying to optimize any further on a system wide basis is sheer
foolishness. I once ran into an idiot who had built his X server (along
with everything else) with -ffast-math. Uh, mate, fast-math means that
floating point libraries are approximated away with integer operations.
Good for crypto. Bad for measuring distance. Needless to say his machine
was crashing left and right.

AfC
Sydney

-- 
Andrew Frederick Cowie
Managing Director

Toronto:
   (416) 848 6072

http://www.operationaldynamics.com/
Management Consultants specializing in strategy,
organizational architecture, procedures to survive
change, and performance hardening for the people
and systems behind the mission critical enterprise.

Sydney  New York  Toronto  London

--
The Toronto Linux Users Group.      Meetings: http://tlug.ss.org
TLUG requests: Linux topics, No HTML, wrap text below 80 columns
How to UNSUBSCRIBE: http://tlug.ss.org/subscribe.shtml