icsphy(4) for nfe(4) - better Microsoft Xbox support

[Yar Tikhiy]

On Wed, Feb 07, 2007 at 02:53:10PM -0600, Brooks Davis wrote:
> On Wed, Feb 07, 2007 at 09:46:36PM +0100, Rink Springer wrote:
> > On Wed, Feb 07, 2007 at 09:39:38PM +0100, Ed Schouten wrote:
> > > I just compiled and installed a kernel with the new nfe(4) driver and
> > > DEVICE_POLLING enabled. Below are the results of some scp(1) transfers:
> > 
> > scp(1) isn't really the best way to test this; the XBOX CPU cannot
> > satisfy a 100mbit link using scp(1). FTP would be a much better idea,
> > preferably with a large file.
> 
> Much better (unless you want to see the effect of disk I/O) would be
> something like benchmarks/netpipe.

For simple benchmarking, I ftp /dev/zero to /dev/null with stock
ftp(1) and ftpd(8), and it works.  It can saturate FastEthernet on
a sub-GHz CPU, giving the theoretical maximum of 11.2 Mbyte/s, if
all the software and hardware involved DTRT.  As to its sensitivity,
once I could notice using the benchmark in NetBSD that ex(4) sucked
by 20% against fxp(4).  And the same benchmark via loopback in
FreeBSD-CURRENT w/o any debug stuff showed a funny thing: just
leaving BPF out from the kernel _reduced_ throughput by 1.41+-0.36%
at 95% confidence.

-- 
Yar