Re: [tsvwg] start of WGLC on L4S drafts

[[e-mail]]

Jonathan,

several IP flows pass a single access shaper at fixed network access and some may briefly fill a queue there. Per flow treatment doesn't scale well - up to now at least and from those dealing with the issue inside Deutsche Telekom, I didn't yet hear the opposite. There should be Best Effort flows which indicate that they respond to ECN and see low jitter then. The other Best Effort flows may see a queue, as they'd expect. I don't yet get the point what's the benefit balancing traffic of different accesses against each other - the link to DSL gear will likely not be congested. What I expect is, that the operator access router is configured for ECN AQM (and maybe for one new classifier and one new Queue), then things need to work in the way the customers expect (for all flows). I'm not sure I understood conditions and requirements for this to work with your proposal, if scale is regarded as an important criterium.

I've discussed L4S with colleagues in fixed and mobile R&D branch and both look at it as feasible from technical and scalability point of view. I wouldn't communicate here otherwise. If I was aware of alternatives pursued within Deutsche Telekom and requiring support in IETF WG, I'd do so (which is not to say, there are none - there may be some, that I'm not aware of). 

Regards,

Ruediger

-----Ursprüngliche Nachricht-----
Von: Jonathan Morton <[e-mail]> 
Gesendet: Mittwoch, 22. September 2021 12:40
An: Geib, Rüdiger <[e-mail]>
Cc: [e-mail]; [e-mail]
Betreff: Re: [tsvwg] start of WGLC on L4S drafts

> On 22 Sep, 2021, at 1:31 pm, <[e-mail]> <[e-mail]> wrote:
> 
> e) no IP flow based mechanism for the resource to be shared (doesn't scale well with available network provider HW to the extent I'm aware) - that means bulk traffic treatment only

Okay, then just use CoDel instead of FQ-CoDel.  All you lose is the isolation of flows' delay influences from each other, and the max-min fairness.  I would however suggest that you at least treat traffic to a subscriber as "a flow", and thus perform max-min fairness and delay isolation between subscribers.

With that said, I'm aware of production hardware that implements Approximate Fairness AQM at 100Gbps, and an implementation of FQ-CoDel in P4.  The latter, as far as I'm aware, should scale up to the point where you start using optical switching anyway, at which point the right strategy is probably over-provisioning.  Even software implementations of FQ-CoDel have been tested successfully at 40Gbps.

 - Jonathan Morton