Re: [tsvwg] WGLC: draft-ietf-tsvwg-nqb-14 - coexistence with L4S

[Sebastian Moeller <[e-mail]>]

> On Nov 3, 2022, at 00:34, Greg White <[e-mail]> wrote:
> 
> Hi all,
> 
> See my responses to the points that relate to the NQB draft below [GW].
> 
> -Greg
> 
> On 11/2/22, 10:24 AM, "tsvwg on behalf of Sebastian Moeller" <[e-mail] on behalf of [e-mail]> wrote:
> 
>    Hi Roland,
> 
> 
>> On Nov 2, 2022, at 16:32, Bless, Roland (TM) <[e-mail]> wrote:
>> 
>> Hi Rüdiger,
>> 
>> see below.
>> 
>> On 02.11.22 at 15:47 [e-mail] wrote:
>>> 
>>> current draft text informs about a coexistence of NQB and L4S in the same LowDelay queue. Does this include that
>>> 	• NQB traffic should or could be marked by a DSCP and have ECN 1 set?
>>> 	• L4S traffic should or could have ECN 1 set and be marked by NQB DSCP?
>>> 
>>> The current text is not clear on that issue. As all marking is to be done by application programmers and this is a standards track doc, I’m interested for the draft to be crisp and clear to that point. Include or exclude simultaneous NQB&L4S  DSCP&ECN setting?
>> As far as I understand the draft, it is just an implementation detail:
>> NQB packets do not need to have ECT(1) set to be put into the L4S low-latency queue.
>> It is a node local mapping from the NQB DSCP(s) to the L4Slow-latency queue, i.e.,
>> they just use the already existing low-latency queue.
>> If I understand section 3.1 correctly, L4S flows should not be marked
>> as NQB traffic since they violate the following condition:
>> 
>> " In contrast, Queue-Building (QB) flows include those that use TCP or QUIC, with 
>> Cubic, Reno or other TCP congestion control algorithms that probe for the link 
>> capacity and induce latency and loss as a result."
>> 
>> and especially L4S flows may also violate the (not so strict) condition of a 1Mbit/s limit.
> 
>    	[SM] That is IMHO incorrect at least generally. Just because TCP allows a flow to search for the capacity limit applications can (and do) decide to send less than possible. 
> 
> [GW] Sebastian's clarification here is correct.  An application could mark its packets with both ECT(1) and NQB if it meets the requirements for both.  This might be a relatively unusual case, but it isn't precluded.  To Ruediger's original request that the draft be crisp and clear on this point, we could add a statement in Section 3.3 that explains this. 

	[SM] Let's see how unusual that is going to be, after all NQB will result in a noticeable better low-latency performance over the ubiquitous WiFi networks (with the cost of that choice invisible to the application employing NQB marking unless the same application also establishes a CS0 parallel flow to asses whether the WiFi network is NQB-aware and whether NQB usage does not result in an overall throughput hit for the WiFi network (less aggregation will result in a total throughput loss)).


> 
> 
>              [SM]But I guess the bigger issue is that the L4S drafts have some words to say when and when not to set ECT(1):
> 
>    https://datatracker.ietf.org/doc/html/draft-ietf-tsvwg-ecn-l4s-id-29#page-14:
> 
> 
>    In order to coexist safely with other Internet traffic, a scalable
>       congestion control is not allowed to tag its packets with the ECT(1)
>       codepoint unless it complies with the following numbered requirements
>       and recommendations: [list skipped]
> 
>    So NQB flows would need a scalable congestion control to be allowed to set ECT(1)
> 
> 
>    However usage of a "scalable congestion control" is not mandatory for setting ECT(1):
> 
>    As a condition for a host to send packets with the L4S identifier
>       (ECT(1)), it SHOULD implement a congestion control behaviour that
>       ensures that, in steady state, the average duration between induced
>       ECN marks does not increase as flow rate scales up, all other factors
>       being equal.  This is termed a scalable congestion control.
> 
> 
>    I would respectfully argue that that SHOULD is mighty weak...
> 
> 
>    BUT that L4S draft hasd the following ot say about NQB:
> 
>    To identify packets for just the scheduling treatment, it would be
>       inappropriate to use the L4S ECT(1) identifier, because such traffic
>       is unresponsive to ECN marking.  Examples of relevant non-ECN
>       identifiers are:
> 
>       *  address ranges of specific applications or hosts configured to be,
>          or known to be, safe, e.g. hard-coded IoT devices sending low
>          intensity traffic;
> 
>       *  certain low data-volume applications or protocols (e.g. ARP, DNS);
> 
>       *  specific Diffserv codepoints that indicate traffic with limited
>          burstiness such as the EF (Expedited Forwarding [RFC3246]), Voice-
>          Admit [RFC5865] or proposed NQB (Non-Queue-
>          Building [I-D.ietf-tsvwg-nqb]) service classes or equivalent
>          local-use DSCPs (see [I-D.briscoe-tsvwg-l4s-diffserv]).
> 
>    It makes a ton of sense to mirror this in the NQB draft explicitly to
>    at least keep these two consistent.
> 
> [GW] I'd argue that mirroring this text in NQB isn't warranted, but I would agree that referencing it in NQB section 3.3 would be appropriate (and it's easier to ensure consistency that way).

	[SM2] I agree, mirroring was bad phrasing, what I meant is describe the same recommendation (only use ECT(1) for "scalable protocols", otherwise use the NQB DSCP for other flows fulfilling the L4S requirements) and add a reference. I see that "mirroring" sounds like I wanted to propose to copy the text verbatim, which I agree would be too much.

> 
> 
>> 
>> So my interpretation of the draft is that the answer to both questions is no, i.e.,
>> either NQB DSCP is set or ECT(1) is set. However, there should be a hint
>> what to do in case both are set (following the L4S argument of letting DiffServ being 
>> orthogonal, ECT(1) should take precedence and the DSCP should be kept nevertheless).
> 
> [GW] While both the NQB draft and the RFC-to-be on ECN-L4S-ID talk about NQB and L4S traffic sharing a low latency queue, it isn't precluded that they be queued separately (i.e. an NQB queue and an L4S queue).  If they were to be queued separately in a node that supports both, I think I agree with you that ECT(1) should take precedence.  The only situation where I can imagine a debate about that would be if a node implemented Traffic Protection for the NQB queue, but didn't implement such a feature for the L4S queue. That said, I'm not super comfortable with starting down the path of making recommendations for this separation of L4S and NQB, since I think it opens a can of worms. For example, is the NQB queue given equal forwarding preference compared to the "Classic" queue in the L4S dual-queue, or is it given equal forwarding preference compared to the L4S/Classic dual-queue combination?  

	[SM2] I read "preference" as a synonym for the more usual "priority" here... why call it preference? (Genuine question: is this a usual nomenclature used in the field?)

> And, would we then want to also describe the use of 4 queues (instead of 3) to handle all four combinations independently (in which case there is no precedence between the two markings)?

	[SM2] I would respectfully propose that until we implement and test such a beast our description will be hardly better than what an implementer will be able to figure out in actual testing/use, so why try to describe something where at best we can offer theoretical musings? Fine if the observations are either trivial or obvious (but in which case not making them should cause little harm), but what if we recommend something on purely theoretical grounds which then in practice fails to meet its goals? IMHO RFCs should try to not speculate too much and stick to what has been supported by data and facts. 

Regards
	Sebastian


> 
>    	[SM] The bigger issue is that L4S AQMs are not mandated to check at all whether any marked traffic actually fulfills the requirements. So honestly it matters little whether a flow uses NQB of L4S to gain access to L4S high-priority queue. Without actual policing/enforcement any requirements in RFC are reduced to "requirements".
> 
>    Kind Regards
>    	Sebastian
> 
>    P.S.: I am convinced that this draft is not (yet) ready and should not pass last call right now.
> 
> 
>> 
>> Regards,
>> 
>> Roland
>> 
>>> 
>>> Von: tsvwg <[e-mail]> Im Auftrag von Black, David
>>> Gesendet: Mittwoch, 2. November 2022 04:47
>>> An: tsvwg IETF list <[e-mail]>
>>> Betreff: [tsvwg] WGLC: draft-ietf-tsvwg-nqb-14, closes 21 November 2022
>>> 
>>> This email announces a TSVWG Working Group Last Call (WGLC) on:
>>> 
>>> A Non-Queue-Building Per-Hop Behavior (NQB PHB) for Differentiated Services
>>>                        draft-ietf-tsvwg-nqb-14
>>> https://datatracker.ietf.org/doc/draft-ietf-tsvwg-nqb/
>>> 
>>> This draft is intended to become a Proposed Standard RFC.
>>> 
>>> This WGLC will run through the end of the day on Monday, November 21.
>>> The WG will meet in London on Monday, November 7, while this WGLC
>>> is in progress.
>>> 
>>> Comments should be sent to the [e-mail] list, although purely
>>> editorial comments may be sent directly to the authors. Please cc: the
>>> WG chairs at [e-mail]  if you would like the chairs to
>>> track such editorial comments as part of the WGLC process.
>>> 
>>> No IPR disclosures have been submitted directly on this draft.
>>> 
>>> Thanks,
>>> David, Gorry and Marten
>>> (TSVWG Co-Chairs)