Re: Multipath (was: Re: Re-chartering for extension work)

[Jana Iyengar <[e-mail]>]

I'm not sure that this is the time for this conversation, but I'll share my
thoughts since the question has been raised.

There will be quite some heavy lifting to do if we go down the path of
designing a multipath extension for QUIC. We've seen this before. I worked
on multipath for SCTP, which was quite some work. I remember when we
started work on MPTCP at the IETF in 2009, and it was a lot more effort
than most people anticipated. When we started work on connection migration
for QUIC, we thought it might get done relatively quickly, but it was a
feature that kept on giving.

None of this is to say that we shouldn't take on a hard problem, but before
we start sliding into it, I would like to be convinced of its importance.
We already have connection migration in the core drafts, and I'd like to
see an articulation of the problems that multipath would solve in addition
to basic migration. I would also like to see implementations that are keen
to build it if an extension were to be developed, and applications that
will use it. Without those, we would be putting the cart before the horse.

- jana

On Mon, Dec 16, 2019 at 12:52 PM Christian Huitema <[e-mail]>
wrote:

>
> On 12/16/2019 9:20 AM, Mikkel Fahnøe Jørgensen wrote:
>
> If you can observe encrypted traffic on both paths you have an opportunity
> to correlate traffic if an endpoint consistently acks on one path and sends
> on another. On path has small packets a fixed interval after larger packets
> on the other paths. That could happen in a migration scenario where one
> path is not yet fully committed so ACKs stay on the main path.
>
> I don’t recall how it works today with single path migration, but I think
> that it is required that ACKs happen on the same path as the packets on
> which they are sent.
>
>
> No, that's not required. In fact it cannot be required, because one of the
> scenarios is "transmit some data on path X, then migrate to path Y". If you
> don't ack on path Y for packets sent on path X, you end up having to repeat
> these packets on path Y.
>
> But your general point is correct, Mikkel. If application traffic is
> spread over multiple paths, it is probably hard to hide the correlation
> between these multiple paths. Not necessarily impossible, but if this is a
> goal we have better design for it.
>
> -- Christian Huitema
>