Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Thursday 20 May 2010 21:26:05 Ben Timby wrote: > My point was that balance-rr is the ONLY mode to provide multiple link > aggregation for a single tcp stream. That said, the switch's > cooperation is only necessary when a switch is present. The > documentation spells out why. > > As both interfaces share the same MAC address, the switch will > recognize only one of them, thus only forwarding packets to it. > Trunking/etherchannel will allow the switch to direct traffic to both > ports. It would seem this problem would not exist if a switch is not > present, however, I have never tested that scenario personally. > > The documentation also mentions that with this mode, packets can > arrive out-of-order, requiring reassembly which will eat into any > throughput gain otherwise realized. Again, whether this problem exists > without a switch is unknown to me. > > You can use tcpdump to sniff the underlying interfaces. Pay attention > to the sequence numbers and timestamps, IIRC, if you view the dump in > wireshark, it will helpfully show you out-of-order segments. > > The documentation I linked discusses this condition along with some > possible mitigating strategies. I tested this configuration a few > years back, and was able to get around 1.2Gbps for a single stream out > of two bonded 1Gbps links. Of course I was using a switch for my > testing, as I had more that two machines involved. > Since testing with a switch resulted in even poorer results, I'll shy away from that for now. Will have a go at investigating with Wireshark though seeing the problem will probably not indicate a solution right of the bat. Cheers, B.