Re: [cobalt-developers] Optimzing bins for archetecture

[Tim Hockin <thockin@xxxxxxx>]

Matthew Nuzum wrote:

> On Fri, 2002-03-08 at 09:40, Paul Adamson wrote:
>     Dont amd processors use the x86 instruction set already...?
>     the main difference in archetecture is the 3d/multimedia instructions on
>     amd's k6/2 processor.
>     Unless you have a 3d acelerator in your raq...?

The extra instructions are pretty fringe
 
> Obviously, things work just fine now, but part of the compiling process
> involves an "optimizer" process.  It seems like there might be some
> optimizations available.

you can recompile the whole system and get MAYBE 5% increase.  hardly worth
it.  There are two things that will give you the vast majority of your
speedup - kernel and libc.  The kernel is already optimized.  I don't know
about libc.  It really isn't worth it ;)
 
> As far as multimedia goes, one of our servers handles creating Flash and
> Gif images on the fly.  Does that count as something that can be
> multimedia-enhanced?

not without you coding in assembly.
 
> Additionally, I hear compiling PHP statically into Apache improves
> performance by 10%.  PHP has a lot of dynamic modules, is it better to
> load these modules statically into PHP, or does the memory usage start
> to become a problem?  (or possibly some other reason not to do it...)

Doubtful.  Once it is loaded it is loaded.  The overhead of an extra
dereference is really insignificant.
 

> cheap as it is, I could throw another computer in there cheaper than if
> I had spent 25 hours on the task just to find out there was a 3.2%
> improvement in speed.

This is ABSOLUTELY correct.  It just isn't worth it, except for a few core
things, and specific places where you need it.
-- 
Tim Hockin
Systems Software Engineer
Sun Microsystems, Cobalt Server Appliances
thockin@xxxxxxx