>> What data rate do you expect on your 3 interfaces?
> Hmm... Along those lines - what's the maximum transfer rate anyone
> has achieved with an hme(4) or quad hme(4) in a 32-bit machine?

Well, I haven't been able to make mine run faster than about 10Mbps,
even when talking with a crossed cable to another machine which can do
full 100Mb; I see more like 5Mbps in an environment not optimized for
the maximum possible performance.

Specifically, I have two i386 machines which, when connected with a
crossed cable, transfer 100MB of /dev/zero in about 10 seconds (11
seconds according to "date ; ... ; date".

When I connect one of them to an hme/fas card in a Classic, with that
same crossed cable:

mainbus0 (root): SUNW,SPARCclassic: hostid 80722791
...
iommu0 at mainbus0 addr 0x10000000: version 0x1/0x4, page-size 4096, range 64MB
sbus0 at iommu0: clock = 25 MHz
...
hme0 at sbus0 slot 1 offset 0x8c00000 level 4 (ipl 7): Sun Happy Meal Ethernet (SUNW,hme)

then transferring 100MB of /dev/zero takes 114 seconds sparc->i386 and
112 i386->sparc, according to similar crude timing methods.  I've
played with what is in use on the various ends but have been unable to
push the time to transfer 100MB of /dev/zero below about 100 seconds,
and that good a result was with only one process on the SPARC, reading
from /dev/zero and writing to the network.  (The 114-second run had two
processes on the SPARC; the 112 had only one, reading from the net and
writing to /dev/null.  I don't know why writing was faster than
reading.)

All interfaces involved were manually forced to 100/full with ifconfig,
and the machines were otherwise idle (except the other i386 machine
used in the i386-i386 test which did achieve full 100Mb rate).  The
tests with the hme are suspiciously close to 10Mb wire rate, but I
checked, and ifconfig does assure me everything is running at 100/full.
(Unfortunately neither the hme nor the interface on the i386 machine
has a 10-vs-100 indicator light.)

The 5Mbps figure is about what I was seeing with the hme connected to
my house switch (a Baystack 350-24T 24-port 10/100 switch) and
transferring to the "other" i386 machine above.  The blinkenlights on
the switch implied that that segment was close to (but not quite) idle
otherwise.

> I'd be curious to know just how much the faster
> SuperSPARC-II/HyperSPARC/TurboSPARC-based machines can handle, both
> with and without pf(4)/ipfilter(4).

My tests were with NetBSD/sparc 2.0 on the Classic, with ipfilter in
the kernel but not enabled (no pf at all, this being 2.0).

/~\ The ASCII				der Mouse
\ / Ribbon Campaign
 X  Against HTML	       mouse@rodents.montreal.qc.ca
/ \ Email!	     7D C8 61 52 5D E7 2D 39  4E F1 31 3E E8 B3 27 4B