Compiling for CPU (was Re: Performance testing (pass 1))

[Stephen Smoogen]

On Fri, 2003-09-26 at 10:42, Rick Johnson wrote:
> ByteEnable wrote:
> 
> >>1. If Open Source writes better software; then why is it fatter and 
> >>slower?
> >>
> >>2. Is any one at RHL working on desktop performance?
> >>
> >>These are some of the sugestione that I may need to answe to when I 
> >>present this and further data to my managers. Gys... what should I say?
> 
> One question I've always had is why Red Hat still compiles for i386 
> (march=i386 mcpu=i686 - ok, so they get i686 instruction ordering)? Once 
> upon a time, I rebuilt my entire RH 7.3 tree and a good portion of my RH 
> 8.0 tree for i686 and/or athlon. I've also run Gentoo, compiled for 
> Athlon XP, which isn't as solid of a comparison, but is still a "Linux" 
> general comparison. In both Red Hat cases, there was an appreciable 

I have also put Gentoo banners on a Red Hat machine and had some
students say it felt so much faster than that Red Hat machine they were
using last week. It is hard to find out where psychological feelings
(red cars are faster than green cars) and real factors come into play.

> Not to pit distros, but Mandrake compiles for i586 by default, though 
> they don't really give out any i686 or athlon specific packages (last I 
> checked). Given that Red Hat's minimum specd hardware is a Pentium class 
> machine with 64MB for text, and a PII 400 with 128MB more more for GUI 
> use, what would we have to lose by changing the default to compile as 
> march=i586 and mcpu=i686, or providing i686 versions of core components

You would be losing as much as 40% performance on non i586 machines with
certain operations. One of the things that you have to be careful with
optimizations is that you may also find that chipset runs may effect
your optimizations as anything. 

Anecdotal story from around here, compiling for the cpu found that what
was thought to be identical PIII had radical differences in speeds. The
PIII turned out to be from different runs or something and that
optimizing for the CPU on the earlier PIII resulted in much slower code
but slightly faster on the newer PIII. The cluster people decided to
just stick with i386 compiles and found that the code ran well
throughout the system. Again this is an anecdotal story from around the
lunch table... could be bs.

>  
> of XFree86? I'm sure it "throws out" the few who like to build the 
> distro for 386 and 486 machines, but I feel the majority would benefit 
> and the minority could simply rebuild the SRPMs for i386 for their 
> specific application (would be a handful of packages at best since those 
> types of machiens probably wouldn't be running full-blown Gnome or KDE 
> desktops, but perhaps browser kiosks at best.
> 
> Thoughts?
> -Rick
-- 
Stephen John Smoogen		smoogen at lanl.gov
Los Alamos National Labrador  CCN-5 Sched 5/40  PH: 4-0645 (note new #)
Ta-03 SM-1498 MailStop B255 DP 10S  Los Alamos, NM 87545
-- So shines a good deed in a weary world. = Willy Wonka --