AMD64 impressions (Re: AMD64 box)

Sander Vesik sander at haldjas.folklore.ee
Fri Nov 7 15:20:52 PST 2003


Matthew Dillon <dillon at xxxxxxxxxxxxxxxxxxxx> wrote:
>    I was able to boot the AMD box with APMBIOS commented out (FreeBSD-4.x
>    also can't boot the box with APMBIOS turned on either.  I'll test
>    FreeBSD-current later today).
> 
>    I ran some preliminary buildworld benchmarks.  I don't have equivalent
>    machines but a few things are obvious from the results:
> 
>    (1) The AMD box is very fast.  Very, very fast.  It blows away everything
>        else hands down.  It is so fast that it spends nearly 1/3 of its
>        time waiting on I/O despite the fact that the I/O requirements
>        for a buildworld are fairly minimal.  It is so fast it makes my P4
>        look like a snail, and even puts my DELL2550 (which is no slouch)
>        to shame.
> 
>        The best comparison is the AMD box verse the P4 box.  Both are UP
>        boxes.  The AMD box is 2.1x faster then the P4 box.  If clock speeds
>        scaled, that would make the AMD box equivalent to 2.6GHz P4.  Since
>        clock speeds don't scale, the AMD box is more like a 3.3GHz P4.

Opteron is easily the fastest commodity (that is, i'm discounting the huge-cache
Xeon variants) 32 bit processor. Compiling programs with icc (ironicly the best
Athlon compiler) makes it even more mindboggling fast.

> 
>    (2) The AMD's syscall overhead is very low compared to the P4.
>        I ran the 'sc1' test in /usr/src/test/sysperf.  The AMD box
>        comes back at 242nS/loop while the P4 takes 1300nS/loop and
>        the DELL box (LEAF/SMP) takes 617nS/loop.

ouch on teh p4 figure

> 
>    (3) I need to work on the SYS time for DragonFly SMP.  LEAF's SYS
>        time should be around 400s, not 961s.  It's the SMP lwkt_token code
>        being used for the vnode locks that is the likely culprit.  LEAF
>        is a 2xCPU DELL-2550 and no slouch (not like VALinux 2xCPU box which
>        is a piece of junk).  LEAF should have done better then it did.
> 
>                                                -Matt
> 

-- 
	Sander

+++ Out of cheese error +++





More information about the Kernel mailing list