Channel bonding problems with 1950s and 2748 switch

[Stephen Childs]

Hi,

I have spent a few fruitless days trying to get channel bonding to work on 
  our new cluster. The machines are the latest revision PE 1950s and the 
switch is the Dell PowerConnect 2748 switch. The network driver in the 
1950s is as follows:

eth0: Broadcom NetXtreme II BCM5708 1000Base-T (B2) PCI-X 64-bit 133MHz

I have tried every channel bonding mode by now I think, both with link 
aggregation groups set up on the switch and without. I never seem to get 
more than the throughput a single card would provide (~940 Mbps or 990 
with jumbo frames and other optimisations). When I look at the output of 
netstat, I see that the transmit traffic seems to be evenly distributed, 
but the receive traffic is mainly being processed by one card.

It's my understanding that bonding modes 0, 1, and 2 require link 
aggregation on the switch, the 2748 doesn't seem to support the LACP 
protocol needed for mode 4, and modes 5 and 6 shouldn't require switch 
support.

As a sanity check I connected the two machines with two crossover cables 
and in this configuration I was able to get ~1.96 Gbps using the 
round-robin mode. (However, balance-alb didn't seem to provide any 
increase in throughput.)

Does anyone else have experience with this switch and NIC? Any tips?

Stephen
-- 
Dr. Stephen Childs,
Research Fellow, EGEE Project,    phone:                    +353-1-8961797
Computer Architecture Group,      email:        Stephen.Childs @ cs.tcd.ie
Trinity College Dublin, Ireland   web: http://www.cs.tcd.ie/Stephen.Childs