Re: Multi-path QUIC Extension Experiments

[Yunfei Ma <[e-mail]>]

Dear Spencer,

Thanks for your comments. To achieve real performance gain, it might take
you to design some "smart" algorithms depending on what you are working on,
but what we were hoping to understand in the earlier experiment was what
basic features we should include to enable a larger design space. In other
words, the algorithm may be more complex and flexible, but the protocol
feature that enables it is simple and independent of the algorithm, like
the QoE_Control_Signal frame defined in draft-liu. Just like what has been
pointed out by Christian, it solves the problem of how to close the
asymmetric knowledge gap between a client and a server. The "knowledge
sharing" mechanism can definitely go beyond our early experiments. I think
the power-saving use case you just mentioned serves as another example
where such a knowledge sharing mechanism can help.

Cheers,
Yunfei

On Tue, Sep 21, 2021 at 12:14 PM Spencer Dawkins at IETF <
[e-mail]> wrote:

> Hi, Christian,
>
> On Tue, Sep 21, 2021 at 1:31 PM Christian Huitema <[e-mail]>
> wrote:
>
>> On 9/21/2021 9:37 AM, Spencer Dawkins at IETF wrote:
>>
>> > I have been maintaining a collection of various goals for path selection
>> > (most recently, in
>> >
>> https://datatracker.ietf.org/doc/html/draft-dawkins-quic-multipath-selection-01#section-3
>> ).
>> > It's not a short list.
>> >
>> > If we're also including user preferences that would include knowing the
>> > difference between cellular connections that are metered, cellular
>> > connections that are unmetered but being throttled, and cellular
>> > connections that are unmetered and unthrottled, and cross-matching them
>> > with wifi connections that are likely to work better than cellular
>> > connections vs. wifi connections that are only intermittently performing
>> > well, that's not a small amount of complexity.
>>
>> This leads to the problem of asymmetric knowledge between client and
>> server. On the client, applications can use system APIs and user
>> preferences to identify which connections are "metered, cellular
>> connections that are unmetered but being throttled, and cellular
>> connections that are unmetered and unthrottled." But the server side
>> does not know that. Yet in many scenarios most of the traffic volume
>> originates from the server, and the scheduling choices of the server may
>> cause increased bills or exhaustion of quotas for the client. Which
>> implies that there should be some way for the client to signal how the
>> server shall consider connection.
>>
>
> Yes, very much so. My point was that this is complex, and your point about
> asymmetric knowledge is well-taken. The idea that "clients" (UEs) know
> things that "servers" (in this case, UPFs) don't know is also popping up in
> 3GPP ATSSS discussions (for example, in clause 5.32.8 of 3GPP TS 23.501
> V17.1.1):
>
> *UE-assistance indicator: *This indicator may be provided only when the
> Steering Mode is Load-Balancing. When provided by the network, it indicates
> that (a) the UE may decide how to distribute the UL traffic of the matching
> SDF based on the UE's internal state (e.g. when the UE is in the special
> internal state, e.g. lower battery level), and that (b) the UE may inform
> the UPF how it decided to distribute the UL traffic of the matching SDF. In
> the normal cases, although with this indicator provided, the UE shall
> distribute the UL traffic as indicated by the network.
>
>
>
>> I think this "knowledge sharing" is independent from "application
>> preferences". The work at Ali Baba shows that there is a lot of benefits
>> in incorporating application states in scheduling decisions. For
>> example, if the video streaming application notices that the video
>> buffers are emptying too fast, it might tell the server to start using
>> more bandwidth, even if that means using metered connections. But that
>> application state, "video buffers are emptying", is not the same as the
>> network state, "path 3 is on a metered connection".
>>
>
> I like this (especially because what you're talking about is meaningful to
> applications and independent of changes in underlying network technologies
> along a path).
>
>
>> I don't think that we will be able to fully standardize the way
>> applications pass their preferences. That's why in draft-liu the
>> "QOE_CONTROL_SIGNALS" is mostly used to pass application-defined
>> signals. But we might be able to have a reasonable set of codes (or
>> priorities) explaining the "state of a path", and use that in the
>> "PATH_STATUS" frame.
>
>
> You and I have shared the hope that we would be able to describe building
> blocks that applications can use, rather than define new codes every time
> someone comes up with a new way of distributing packets across multiple
> paths, for a while now -
> https://datatracker.ietf.org/doc/html/draft-dawkins-quic-what-to-do-with-multipath-03#section-2.4
> was based on a discussion we had, and that was submitted last January.
>
> I hope the QUIC working group's return to multipath discussions gives us a
> chance to make progress on that.
>
> Best,
>
> Spencer
>