[Soekris] net4801 bridging firewalls benchmarks results

[Giovanni Faglioni]

2005/10/9, Chris Cappuccio <chris at nmedia.net>:
> So, what in the world is "maxxed out irq's" ?

      My understanding is that "maxxed out irq's means that
      the irq processing is the bottleneck.

> Anyways, the OpenBSD driver is very similar to the FreeBSD one, as that's
> where it's derived from.

     I'll check if it's possible to recompile an OpenBSD kernel
     with DEVICE_POLLING defined for the natsemi chipset.
     (And maybe HZ=1000 or 2000).

> Although, I agree I've never seen numbers for FreeBSD on Soekris as high as
> this before.  (Then again, most of my memory is probably of net45xx performance
> and not net48xx.)  But even with polling, the idea that a net4801 uses only
> 4-8% CPU while pushing nearly full speed is interesting and seems
> inaccurate.

    Maybe we measure it in a wrong way. Our methodology is to
    keep a "top" window open on the 4801 box, and look at the
    "system" and "user" numbers. They go up (expecially "system")
    immediately as the test begins, then drop quickly after a couple
    of seconds to 7-8% levels. (user is <2% system is ~6-7%)

--

   --Giovanni Faglioni