Re: What to do about multipath in QUIC

[Christian Huitema <[e-mail]>]

On 11/12/2020 3:10 AM, Mikkel Fahnøe Jørgensen wrote:
>   1. We think QUICv1 has already laid down the foundation to build a general-purpose multi-path since migration can be viewed as a special type of multi-path. Therefore, we think one should reuse the design of migration in QUIC-v1 as much as possible, along with the features such as PATH_CHALLENGE/PATH_RESPONSE for path challenge and address validation, and NEW_CONNECTION_ID/RETIRE_CONNECTION_ID for CID renegotiation of new path(which is called Sub-connection in our draft). Reusing these features of QUIC-v1 with small extensions has enabled us to get general-purpose multi-path features with very little efforts in
> Alibaba’s XQUIC(an implementation of QUIC-v1).

Yes. That's a major investment in QUIC V1, and we should keep it.

>   2. We find that the simplest way to add a second path is to use a sub-connection. The concept of sub-connection is directly inherited from connection in QUIC-transport, defined as the logic session of each physical path. Similar to a connection, each sub-connection has its own packet number space for the purpose of loss detection and recovery.
>

I did not want to do that in my own draft for a couple of reasons. The
main one is the interaction with encryption.

The use of AEAD is only safe if the same packet number is not reused
twice with the same key. If we use multiple packet number contexts, AEAD
is only safe if these contexts use different encryption keys. This
requires adding a key derivation procedure for the "sub connection", and
also adding ways to identify the relevant key in the incoming packets.
This gets complicated very quickly, especially if we want to keep the
possibility of using zero-length connection identifiers on the client side.

I use a concept very similar to the sub-connection, but only as a way to
manipulate paths, so the client can instruct the server when paths ought
to be abandoned. Otherwise, I just keep track of which PN maps to which
path.

>   3. To merge the gap between migration and the general-purpose multi-path, several new features need to be supported: 
>   - (1) how to manage the lifecycles of individual sub-connections. 
>   - (2) how to distinguish packets coming from different sub-connections.  
>   - (3) how to co-operate with a multi-path scheduler.
>
> We would appreciate hearing any thoughts, comments and suggestions.


I think we need more work on the "multi-path scheduler". We have heard
of three application scenarios: maintaining the lowest RTT when sending
voice segments (Apple Siri), avoiding buffering delays when playing
music (Apple Music), and using two available links with equal preference
(Ali Baba "high speed train"). I wish that we could distillate that into
a couple of simple concepts.

-- Christian Huitema