On Mon, Feb 12, 2007 at 03:09:27PM +0100, Reyk Floeter wrote: > On Mon, Feb 12, 2007 at 01:30:18PM +0000, Jon Morby wrote: > > On 12 Feb 2007, at 13:18, Stuart Henderson wrote: > > > > >On 2007/02/12 12:44, Jon Morby wrote: > > >>My problem is that graphs of the 2 cisco ports show traffic is only > > >>going via the 1 port and not being balanced across both ports as I > > >>would have expected. > > > > > >loadbalance hashes the header to determine which link to use; you > > >might > > >want round-robin instead? check trunk(4) for descriptions. > > > > > > I should have said, I have also tried with roundrobin > > > > and also removing the channel-group from the switch. > > > > the default cisco port/etherchannel link aggregation works similar to > the loadbalance mode. some switches/ios versions allow to use other > protocols, like roundrobin. > > > The only real performance increase I've seen is with the channel- > > group removed in which case we do see some traffic across both ports, > > but we still only get about 1.4MB/sec and not the 1.8MB/s-2.2MB/s I > > would have expected to see from scp transfers. (graphs show 8Mbit > > which matches what I'm seeing from scp) > > > > With the ports set to GigE we see a major speed increase, so it's not > > a bottle neck on the sending machines as far as I can ascertain. > > > > again, the roundrobin mode will distribute every single packet over > the ports and you may get a speed increase with single connections. > > the loadbalance mode will hash the packet headers (src/dst ethernet, > src/dst ip, vlan) and distribute the connections over the ports. you > may get an overall bandwidth increase with many connections from > different addresses/vlans. > > by default, all the known vendors do a hash-based loadbalancing > (cizzco-eeh etherchannel/FEC, hp trunk, ...). it is a marketing lie > that it will increase the performance by the number of ports, it > heavilly depends on your individual network traffic and the number of > different connections, but it will never exceed the maximum link speed > for a single connection. > > as is said, roundrobin mode may increase the speed, but it also > increases the interrupt load and many other factors. and it doesn't > work very well with non-openbsd systems on the other side. i have seen > it only once, that i got ~166Mbit/s with a crosslink trunk between 2x2 > rl(4) nics. >
Roundrobin may increase packet reordering which in turn reduces the tcp window size because tcp thinks it is a network congestion. In the worst case one connection may run slower over two link trunk than over a single link. You need a real multilink capable L2 portocol (like ppp) to fully use the bandwith of the additional link. Ethernet was not designed for that and so bonding/trunking of interfaces give you a sub-optimal performance improvement. -- :wq Claudio
