Anybody here ever messed around with the Linux
network bonding stuff, particularly in conjunction
with a Cisco Catalyst switch and its EtherChannel
capabilities?  I have the bonding stuff mostly
sorta kinda working but the behavior is a little
strange in that the throughput numbers exhibit
variations that I don't understand.  Note that
I'm in unfamiliar territory here so if it seems
like I don't know what I'm talking about it's only
because I don't...

This appears to be a switch issue because if I
declare (say) 5 pairs of Catalyst ports to be 5
"channels" and I connect machineA's two NICs to
portPairA and machineB's NICs to portPairB (with
nothing else connected to the switch) all will
be well for a time and I will get the expected
aggregate throughput numbers in both directions,
as measured from either host by saying "netperf
-H otherHost".  Then, sometimes, after a while
netperf on one host will start reporting lower and
lower numbers until we're at (or even below) the
numbers obtained for a single un-bonded connection,
while the other host seems perfectly happy and
continues to obtain the higher aggregate numbers
(ie. ~165 Mbits/S versus ~95)

Strangely, if I swap the connections (ie.
machineA's two NICs into portPairB and machineB's
NICs into portPairA) the "slow" machine will be
fast and vice versa.  Or if I move the "slow"
machine's connections to a previously idle channel
like portPairC then both machines will once again
start reporting the higher numbers.

When I ask the Catalyst to tell me everything about
the vlans and channels and statistics and every
other dang thing I can think of I see nothing
that's different or remarkable for one port or
channel compared to any other.

FYI, I also tried this with an Intel switch (their
term is "port aggregation") and the numbers are
rock steady at ~95Mbits/S, showing no variations
but also no throughput numbers higher than I can
get with an unbonded NIC.

_______________________________________________
gnhlug-discuss mailing list
[EMAIL PROTECTED]
http://mail.gnhlug.org/mailman/listinfo/gnhlug-discuss

Reply via email to