Re: [tsvwg] L4S drafts: Next Steps

[Sebastian Moeller <[e-mail]>]

Hi David,

> On Dec 9, 2020, at 23:28, Black, David <[e-mail]> wrote:
> 
> Bob,
>  
> Reminder of the goal for [1] (Rework the “Prague Requirements” into the following two entities ...):
>      >> Success of this exercise will be judged by the WG reaching “rough consensus” that multiple transport protocol implementations
>      >> currently meet and/or are reasonably expected to be able to meet the normative requirements (a).
> If the L4S folks believe that this has already been done, then it’s time to start bringing forward the multiple transport protocols that demonstrate success of that exercise. 
>  
> >Nonetheless, if there are specific items in S.4 that someone believes should or shouldn't be there, or should be worded differently, let's have those discussions now on the list - I'm all ears
>  
> Speaking only for myself as an individual beyond this point, these two normative requirements struck me as ones that might be better stated as design and implementation guidelines:
>  
>    o  A scalable congestion control MUST eliminate RTT bias as much as
>       possible in the range between the minimum likely RTT and typical
>       RTTs expected in the intended deployment scenario (see
>       Appendix A.1.5 for rationale).

	[SM] As argued before, the comparison here should be against the current state of affairs and not against an unspecific potential and the requirement should be something actionable like:

RTT bias MUST not be significantly worse than the RTT bias observed in a reference transport protocol (e.g. Reno, or to better reflect the current internet CUBIC?) over a reference queue (e.g.  dumb FIFO or reference AQM, like RED or PIE). As written it is quite confusing, including the question why the bias removal magnitude seems is contingent upon deployment scenario. How should a transport protocol designer be able to predict where the protocol is going to be used? Most protocols will aim for use over the internet, so what are minimum and typical RTTs over the internet? 
	I have noted before that L4S introduces a specific new meaning to the term "typical RTT", so does the typical RTT in this requirement denote that specific L4S definition or the "intuitive" interpretation of "the mean/median RTT ne expects on a specific link"?



The really really odd thing here is, that TCP Prague plus DualPI2 actually introduce a considerably HIGHER RTT bias compared to the status quo (see e.g. https://camo.githubusercontent.com/0ca81a2fabe48e8fce0f98f8b8347c79d27340684fe0791a3ee6685cf4cdb02e/687474703a2f2f7363652e646e736d67722e6e65742f726573756c74732f6c34732d323032302d31312d3131543132303030302d66696e616c2f73312d6368617274732f727474666169725f63635f71646973635f31306d735f3136306d732e737667 and compare the first 4 pairs of bars with the fifth, L4S clearly increases RTT bias by a lot) and hence both fail completely to meet this requirement.  Comparing pair 1 with pair 5 shows that DualPI2 increases RTT bias even for flows in the cassic queue over a FIFO, and comparing pair 4 with pair 5 shows that TCP Prague takes this to a new starvation-class extreme. Hence both the reference L4S AQM (Dual Queue coupled) and the L4S reference transport protocol fail hard in this requirement.


Best Regards
	Sebastian


>  
>    o  A scalable congestion control SHOULD remain responsive to
>       congestion when typical RTTs over the public Internet are
>       significantly smaller because they are no longer inflated by
>       queuing delay (see Appendix A.1.6 for rationale).
>  
> The use of “as much as possible” in the first bullet makes it hard to figure out whether a specific congestion control meets that mandatory requirement for participation in the L4S experiment.  The second bullet appears to involve a prediction of the future as currently written.
>  
> Thanks, --David
>  
> From: Bob Briscoe <[e-mail]> 
> Sent: Tuesday, December 8, 2020 6:01 PM
> To: Black, David; [e-mail]
> Subject: Re: [tsvwg] L4S drafts: Next Steps
>  
> [EXTERNAL EMAIL] 
> 
> David and chairs,
> 
> On 04/12/2020 15:54, Black, David wrote:
> The WG chairs have consulted among ourselves and with our AD (Martin Duke).
> Our guidance to the WG and the L4S draft authors is that at least the following two things need to be done before WGLC will become possible on the L4S drafts:
> [1] Rework the “Prague Requirements” into the following two entities, which will overlap:
> a.       Transport-protocol-agnostic normative requirements for all congestion-controlled traffic that uses the L4S low latency queue.
> ·         The DualQ AQM design relies upon traffic adhering to these requirements.
> b.       Design and implementation guidelines for new congestion controls that meet the normative requirements (a).
> ·         Ultimately, the choice of whether a particular aspect is a guideline (b) or requirement (a) is a WG “rough consensus” decision.
> 
> [BB] I believe draft-ietf-tsvwg-ecn-l4s-id is already structured for this division between normative requirements and design and implementation guidelines. 
> * [1]a) is Section 4 "Prerequisite Transport Layer Behaviour". Nearly every paragraph also refers off to subsections of Appendix A.1 for extra non-normative info, headed "Requirements for Scalable Transport Protocols". [Actually, it might be useful to swap these two headings]
> * [1]b) is Appendix A.2: "Scalable Transport Protocol Optimizations"
> 
> The rule that determined which aspect went in which was:
> [1]a) if it's about an aspect of one transport's behaviour that has potential to impact/harm others
> [1]b) if it's solely about how a transport can improve it's own performance.
> They have been divided like that from the very first ad hoc meeting in Prague that formulated them (in 2015) 
> 
> Code to address all the requirements and most of the optimizations has been implemented. Over the last year, 3 or 4 of the later requirements in a) have become SHOULDs not MUSTs. Personally I would have left most as MUSTs, but I've tried to reflect what the WG wanted. Those demoted to SHOULD have generally been considered a lower risk of harm (either low likelihood of occuring or minor harm if they do) - taking into account the complexity of implementing them.
> 
> We could certainly flag the point where the transition from MUSTs to SHOULDs occurs within section 4. But I think we should still group all the items with potential for harm together in the same section. Because assessment of risk will change as the Internet landscape changes (possibly over the duration of the experiment).
> 
> Now I've explained, I hope the chairs will all agree that "rework the Prague Requirements into two entities" (normative requirements and design and implementation guidelines) is already done and dusted.
> 
> Nonetheless, if there are specific items in S.4 that someone believes should or shouldn't be there, or should be worded differently, let's have those discussions now on the list - I'm all ears. 
> 
> 
> 
> Bob
> 
> 
>      Success of this exercise will be judged by the WG reaching “rough consensus” that multiple transport protocol implementations
>      currently meet and/or are reasonably expected to be able to meet the normative requirements (a).
> [2] Build the safety case for L4S experimental Internet-wide deployment in the L4S Ops draft.
> a.       This safety case does not rely on endpoints running TCP Prague, though it can assume endpoints are meeting the reworked Prague requirements
> b.       Ops draft will provide guidance on how to detect L4S problems on their RFC 3168 network, and how to mitigate them.
> c.       Ops draft must consider these scenarios:
> ·          An unsophisticated user purchases an L4S endpoint and runs it on a service provider's single-queue RFC 3168 network.
> ·          L4S traffic from another domain enters an RFC 3168 single-queue network (e.g. a peer-to-peer application)
>      Success of this exercise will be judged by the WG reaching “rough consensus” that deployment of the L4S experiment
>      is unlikely to cause significant damage to the Internet, and that any damage that results is expected to be tolerable.
> Thanks, --David [as TSVWG WG co-chair]
> ----------------------------------------------------------------
> David L. Black, Senior Distinguished Engineer
> Dell Technologies, Infrastructure Systems Group
> 176 South St., Hopkinton, MA  01748
> [phone]           Mobile: [phone]
> [e-mail]
> ----------------------------------------------------------------
>  
> 
> 
> -- 
> ________________________________________________________________
> Bob Briscoe                               http://bobbriscoe.net/