Re: [multipathtcp] High-level design decisions /architecture

[Mark Handley <[e-mail]>]

2009/11/3 Joe Touch <[e-mail]>:
> Costin Raiciu wrote:
>> Hi Joe,
>>
>> The proposed congestion controller gives the aggregate multipath
>> connection the throughput a TCP would get on the best (fastest) of the
>> paths available (to the multipath connection). It does so only by
>> looking at loss rates and rtts on the different paths. It does not need
>> to know whether there is a shared bottleneck or not.
>
> If it doesn't, then you will create the potential for a single multipath
> group to fight with itself, which seems less than productive.

Well, so long as the total window increase per RTT of all the subflows
in a connection is no more than a single vanilla TCP would get, then
you don't fight with yourself any more than one TCP would ("fighting"
in this context is pretty much equivalent to how much unacked traffic
per RTT a flow can inject - the ack-clocked traffic already had its
space last RTT, so isn't "fighting" for extra space).

This is exactly how the proposed congestion control algorithm works
through a common bottleneck.

 - Mark