[lug] NIC Speed

[Tkil]

>>>>> "svq" == Stephen Queen <svq at peakpeak.com> writes:

svq> The advertised 100Mb/s is the maximum bit rate of the card, not
svq> the average sustained speed. If you are getting 30% of the
svq> maximum sustained speed from your NIC that would be in the
svq> acceptable or normal range.

i find 30% to be a little pessimistic.  for ethernet, i usually plan
on 60-70% of rated speed to be available (so, at 100Mbps, i'd expect
to see 12.5*.6 = 7.5MBps transfers).  

of course, i developed this heuristic back in the days of thinnet, so
i have no idea how bridges or hubs can affect things.  likewise, that
was back on 10mbps, which didn't strain the hosts nearly as much.

a few quick tests here at home showed about 8MBps from the win2k box
up to the linux box, and 7MBps down.  this is with cheap integrated
10/100 ethernet (and mii-tool is showing 10mbps/half-duplex on the
linux box, so i'm guessing that it's a big pile of crap).  netperf
reports 92Mbps throughput.

(systems are an HP pavilion w/ 1.2GHz athlon running win2k sp2,
talking through a netgear rp114 switch/nat/firewall to a beige box
running redhat 7.2 on a via c3.  both boxes have rtl8139 derived
integrated ethernet.)

i regularly saw 11MB/s transfers when i was working at the hp plant in
fort collins.  this was with pretty good 3com (905?) adapters,
connected through at least three switching boxes (pc - switch - router
- switch - host).  very nice.

hm... even the mii-diag & friends think that interface is fixed at
10Mbps -- but the switch contradicts it.  *shrug*

anyway, remember that you might be pushing the capabilities of other
components when transferring at 10MB/s: even the latest hard drives
can only do 30MB/s; PCI saturates at 133MB/s; your memory bus might be
wide and fast, but it might not be; your filesystem might need
multiple round trips to the hard drive to create a file as it's coming
in, etc.

the netperf stuff seemed to measure things reasonably well.  so, for
raw wire speed, give netperf (www.netperf.org, a google search for
"netperf win32" gave me the binaries i needed -- i used 2.1pl2 on
win2k, 2.1pl3 on linux).

failing that, ftp transfers (don't forget to put it into binary mode!)
are a not-too-bad way of estimating "real life" wire capacities.
apache should also be capable of saturating the wire, but there's a
bit more overhead there (heh, unless you want to experiment with TUX i
suppose...)

t.