From: John Beckett 

"Gregg N"  wrote in message
news::
> -b 1000m -l 1024 ==> 250 Mbps
> -b 1000m -l 1025 ==> 200 Mbps

This looks interesting. I'd like to try it out when I get some time.
However, would you please spell out a little more what the test setup is.

I infer that for your loopback test, I use a single Windows computer and run:

# Following makes computer act as a server to receive test.
iperf -s [-o outputfile]

# Following sends the test.
iperf -u -c 127.0.0.1 -i 1 -b 1000m -l 1024

I have heard that XP and W2003 have a QoS service running that reserves a
few percent of the bandwidth. It can be turned off. I don't think it
applies to a default W2000. No doubt the tuning links on the iperf web site
have this info and more.

My first reaction would be to run a network sniffer and find out exactly
what is happening. I wonder how the tool measures errors. Is there an ACK
packet? Does the switch (when not using loopback) get overwhelmed? Is there
a full-duplex issue (the server send packets back, and the full-duplex
traffic break something?).

Of course, using loopback makes these questions pretty redundant. Also, it
is mega strange how Rich could use a totally different tool and get similar
weird results.

I wonder if some memory alignment requirement could penalise the 1025-byte
packet (which is presumably 20+8+1025 = 1053 bytes for the IP+UDP+data).

Actually, I just read your link:
http://www.chch.demon.co.uk/wintest/wintest.html
and his theory that >1024 bytes makes Windows move data around, perhaps
with dynamically-assigned buffers sounds very plausible. In which case, all
our speculation and testing is pointless .

John

--- BBBS/NT v4.01 Flag-5