[Click] Load balancing - packet reordering

Beyers Cronje bcronje at gmail.com
Wed Mar 1 14:35:55 EST 2006


Hi Torquato,

For Tcp have you enabled Selective Acks (SACK) ? This should help with
excessive retransmissions due to out of order packets.

Another idea that may possibly help is to split the IP flows to two seperate
queues, i.e. by using a hash classifier to output all even IP Address flows
in one queue, and all odd IP Addess flows to the second queue. Each outgoing
inteface can then pull from their respective queues, ensuring that each
flow's packets will always end up in the same queue. Obviously this can
affect full utilization of the load balanced link, but I'm sure there are
better hash functions than the odd/even ip address way.

As you noted, maybe an ordered TCP queue at the receiving router might do
the trick aswell.

On a side note, have you done any test on SMP click with your configuration?
Theoretically these 'dual path' configurations should benefit from SMP
Click.

Beyers



On 3/1/06, Torquato Bertani <torquato at gmail.com> wrote:
>
> > Out of order packets are going to be a fact of life.
>
> :)
>
> > Question:  Why is reordering such a problem for you?
> > Is it intra-stream or inter-stream reordering thats the problem?
>
> The problem is intra-stream, obviously. If I send a UDP stream (real
> time audio or video for example) I need it reordered before passing it
> to superior layers. Otherwise I have many malfunctions because
> superior layers don't like very much out of order data...also tcp
> throughput decrease due to out of order packets.
>
> At the moment I'm working on a ethernet configuration but my idea is
> to load balance different wireless interfaces. Wifi transmission
> problems make the % of out of order packets very high (20-30% in some
> rapid tests).
>
> I forgot to write another idea in previos post: marking packets at the
> tx side and use the number to reorder them at rx side.
> I got some ideas from the MPLS protocol but I don't want to implement
> a real complex protocol.
>
> Torquato
>
> _______________________________________________
> click mailing list
> click at amsterdam.lcs.mit.edu
> https://amsterdam.lcs.mit.edu/mailman/listinfo/click
>


More information about the click mailing list