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

[Lucas Pardue <[e-mail]>]

Hi Oliver, all,

Taking a very brief look over the 03 draft, if I grokked it properly I have
an observation that may be interesting in the broader thinking about
extending QUIC.

In draft-deconinck-quic-multipath-03 [1] it states that because there are
new path-related IDs, some frames need to be modified and then lists them:
NEW_CONNECTION_ID, RETIRE_CONNECTION_ID and ACK. IIUC this would keep the
type value but modify the frame contents. My model of thinking has been to
assume that frame definitions are immutable, and that extensions that
require tweaks to frames would define new ones.

For example, draft-huitema-quic-1wd-00 [2] proposes a timestamped ack and
so requires a "Time Stamp" field in the ACK frame. It does so by defining
two new ACK frame types (vanilla, and one to contain ECN marking).

So my observation is not about the proposals themselves but about the
composability of extensions. Say for example someone wanted to combine
mpquic with 1wd, would they modify the ACK frame or define a new one?

>From an implementation perspective, making frame parsing condition on
connection-level characteristics such as extensions seems like a bad idea.
Frame types are cheap so lets use them.

[1] https://tools.ietf.org/html/draft-deconinck-quic-multipath-03
[2] https://tools.ietf.org/html/draft-huitema-quic-1wd-00


On Fri, Jan 10, 2020 at 9:51 AM Olivier Bonaventure <
[e-mail]> wrote:

> Dear Mirja,
>
>
> As Tommy said, migration is already the first half of multipath support
> for QUIC. So I expect it will be easier to add it to QUIC than it was for
> TCP.
>
>
> The main benefit of QUIC is that we do not need to take middleboxes into
> account and can reason about the transport layer directly.
>
> I would at least be curious to spend some time on figuring out how much
> (or how less) effort it would be. Maybe there is also another step before
> going to full multipath support that could support the handover use case
> better, e.g. sending redundant data on multi paths or something. I think it
> worth think about it.
>
>
>
> Please don’t do that. With QUIC,  have a unique opportunity to design an
> architecturally clean and easily deployable multipath transport protocol.
> We should not pollute this design with hacks and short term solutions that
> have been included in too many Internet protocols for various reasons.
>
> However, while I would like to see this discussion starting now, I think
> it’s fine to wait for adoption of any draft until v1 is done.
>
>
> As already explained in previous emails, Quentin De Coninck has leveraged
> all the lessons learned from Multipath TCP in a Multipath QUIC design that
> has evolved in parallel with the standardisation work of QUIC v1. The
> initial design was described in
>
> https://multipath-quic.org/conext17-deconinck.pdf  and in
> https://www.ietf.org/archive/id/draft-deconinck-multipath-quic-00.txt
>
> There was a running implementation on top of quic-go that is available
> from https://multipath-quic..org <https://multipath-quic.org>
>
> The draft has evolved several times on
> https://datatracker.ietf.org/doc/draft-deconinck-quic-multipath/
>
> The recent draft has been implemented on top of picoquic and was released
> on pquic.org
>
> Quentin has studied these designs in details with many experiments, both
> in the lab and in the wild on smartphones. He’d be ready to present this in
> details to the working group if if eventually decides to work on the
> multipath item of its charter. I would encourage the working group members
> to start to read the last version of Quentin’s draft and provide comments.
>
>
> Best regards,
>
>
> Olivier
>
>
> ------------------------------
>
> Disclaimer: https://www.tessares.net/mail-disclaimer/
>
>
>