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

[Mark Handley <[e-mail]>]

2009/11/3 Joe Touch <[e-mail]>:
> Fighting could occur if the subflows send packets at the same time, even
> if the aggregate is the same as a single TCP would have sent. A single
> TCP never sends two packets out at once (it serializes them), but
> separate TCPs could easily do this (even if the OS serializes them, they
> could go out simultaneously from different NICs).
>
> Keeping in mind how TCPs like to sync, this sounds like an area to keep
> an eye on.

Good point.  I'd hope that buffers at the queue feeding the common
bottleneck would cope with this, but it's possible it makes the
behaviour of the aggregate at the bottleneck more bursty.  We've not
seen any evidence for this, but it's not something we've explicily
looked for.  Regular TCP is pretty busty though, sending a lot of
back-to-back packets unless explicit spacing is done, so I doubt
MP-TCP would be any worse in most topologies.  Still, definitely
another thing to look out for for in our experiments.

 - Mark