Re: Silly question

[Simon Burge <simonb%NetBSD.org@localhost>]

Jason Thorpe wrote:

> On Sep 6, 2008, at 3:49 PM, Perry E. Metzger wrote:
> 
> > Vincent <10.50%free.fr@localhost> writes:
> >> If true, what are the benefits (and drawbacks) of using AMD64 instead
> >> of I386? Increased memory space, but is there any performance  
> >> upgrade?
> >
> 
> You also get a much faster system call interface.
> 
> In general, I've found "64-bit user code is a total win" for almost  
> everything.

On a single core opteron system from earlier this year, a build.sh test
shows a slight benefit using 32-bit userland/kernel over 64-bit:

i386 building i386 release
     7157.711u 1477.441s 2:28:11.86 97.1%    0+0k  3268+288792io 58119pf+177w
     7162.819u 1475.024s 2:28:21.42 97.0%    0+0k 10427+288785io 58130pf+157w
amd64 building i386 release
     7064.045u 1823.976s 2:34:21.40 95.9%    0+0k 13196+295870io 68022pf+81w
     7065.928u 1842.092s 2:34:32.37 96.0%    0+0k 12758+294870io 67918pf+72w
 
i386 building amd64 release
     5529.617u 1149.725s 1:56:21.37 95.6%    0+0k 13885+226200io 39807pf+193w
     5528.191u 1151.156s 1:56:16.65 95.7%    0+0k 13477+227283io 39808pf+210w
amd64 building amd64 release
     5409.935u 1418.328s 2:00:08.56 94.7%    0+0k 13632+232673io 48524pf+110w
     5410.603u 1421.017s 2:00:06.11 94.8%    0+0k 13338+231756io 50470pf+115w

> > It is very application dependent. The AMD64 architecture has more
> > registers, and the registers are wider, so for some applications you
> > can get significant speedup. On the other hand, you also get wider
> > pointers and such, and those consume more cache and memory. You should
> > benchmark if you're doing any sort of performance critical application
> > to figure out which is better for your app.

Exactly!

My current rule of thumb is "if you need to access more that 3.5G of RAM
use amd64, otherwise use i386."

Simon.