Re: [tsvwg] Qs on your 5G L4S slides

[Sebastian Moeller <[e-mail]>]

Hi Bob,

more below prefixed with [SM]

> On Mar 15, 2021, at 19:06, Bob Briscoe <[e-mail]> wrote:
> 
> Ruediger,
> 
> ==DOCSIS==
> Whoa! NQB is not L4S traffic. NQB is a Diffserv codepoint. L4S is identified by the ECN field. In DOCSIS the NQB Diffserv codepoint classifies into the /same queue/ as L4S traffic (renamed the Low Latency queue due to its dual role). Allowing in unresponsive traffic was only considered in DOCSIS because there was already a sufficient policing function in front of the queue (per-flow queue protection).
> 
> ==Mobile==
> If a mobile operator (or in this case a masters student), uses the ECT(1) codepoint to classify traffic into a priority bearer, then it's not L4S. It's an ECN codepoint intended for L4S but used (abused?) in a Diffserv priority scheduler. 
> 
> The problem that the DualQ Coupled AQM solved was how to isolate low latency flows without having to know how much bandwidth to set aside for them.

	[SM] Except it does not doe so robustly and reliably, at ist core it is a ~1:16 classic/LL priority scheduler that gives a massive rate advantage to LL traffic that is ameliorated by a few heuristics that under some conditions result in more equitable sharing, but under different not less realistic conditions it just behaves like the 1:16 scheduler it is at its core.

> So if there are M L4S flows and N Classic flows, M and N can take any value, including zero. That's because the coupling makes the two queues appear as one - from a bandwidth and congestion control perspective (approximately). 
> 
> So, if you have a Diffserv scheduler and no L4S mechanism, you would need to go back to using traditional Diffserv techniques like guessing what M and N might be most of the time, to decide how much bandwidth to configure for a separate priority queue, then policing it. 

[SM] Again, coupling does not solve that problem robustly and reliably either... you just keep focussing on those sets of conditions where it does and ignore the cases when it fails.

> 
> To summarize, the answers to your question:
>> The underlying question is, to which extend does the end-to-end performance of L4S depend on suitable radio schedulers coupling two congestion control algos or queuing behaviours, like L4S standardises for fixed line schedulers. And how to operate a network, if these are absent.
> 
> An operator that wants to support any technology without deploying the technology isn't going to get very far! L4S depends on using an L4S mechanism (obviously), specifically the DualQ Coupled AQM (or FQ). How to operate a network if L4S is absent - well, you go back to what you had before. But then you can't support applications that need consistently low latency /and/ the full available bandwidth, which is the point of L4S.

[SM] Might be L54S point, but so far the proof is missing that L4S can actually deliver that over the existing wider internet. I keep asking for data supporting that claim for a few years now. At this point in time I interpret the lack of data as indication that the claim is not achievable.


> 
> 
> ==WiFi==
> You say that the NQB draft "specifies mapping L4S to a priority bearer based PHB". This is because NQB is having to cope with the WiFi situation as it finds it. It's not ideal, but you'll see below how it could evolve to something better.
> I understand that the video access category (AC_VI) was the only choice that offered decent enough latency without excessive bandwidth priority.

	[SM] Erm, AC_VI will wipe the floor with AC_BE and AC_BK, since it has a considerably higher probability to gain airtime than these lower ACs... For NQB that might be acceptable IFF the ISP polices the rate of NQB ingress and egress traffic to sane levels. Assuming there are one-size-fits-all sane levels for WiFi.


> NQB just needs to be isolated from bursty traffic - it didn't choose AC_VI because of any need for /bandwidth/ priority, per se. NQB should work with quite weakly weighted priority as long as it's isolated. But that wasn't available in current WiFI.
> 
> 
> L4S is also walking into the WiFi environment as it finds it. With today's non-L4S products, I would also recommend that the L4S-ECN codepoints are mapped to the video access category, if possible. 

	[SM] since the L4S LL queue is not admission controlled that is going to be a misdesign, that makes DualQ's 1:16 priority scheduler look benign, AC_VI can not only almost starve AC_BE and AC_BK from the same station, but also all other WiFi packet on the same channel...
	I wonder whether you have actually tested your assertions about WiFi?
	


> Nokia's latest WiFi products (in the 'Beacon' range) already include an L4S DualQ Coupled AQM. And as other L4S WiFi products come out, the coupling will introduce the recommended congestion signals that can be used as back-pressure against the priority scheduler. Users don't want to abuse scheduling priority at the expense of the balance between their own applications. But they have no choice until there's a mechanism that allows their applications to balance against other apps.
> 
> Finally, once there's an L4S queue in WiFi kit, NQB traffic could be classified into it, as was done in DOCSIS.
> 
> FQ offers an alternative path for WiFi - neither precludes the other.
> 
> Does this help explain?

[SM] The good thing is that a) airtime fairness goes a long way to fix wifi (and without L4S), and b) WiFi introduces ~4ms aggregates making the difference between L4S' 1ms target and fq-codels weaker 5ms target pretty immaterial. That is if one accepts WiFis increased worst case RTT (with lower lever retransmits RTTs can increase by 100s of ms) there is exactly zero reason to continue waiting for Godot^W L4S, just get a linux AP with airtime fairness and you solved 95% of the issue with 0% of L4S side-effects.

Best Regards
	Sebastian


> 
> 
> Bob
> 
> On 15/03/2021 11:19, [e-mail] wrote:
>> Hi Ingemar,
>>  
>> I’m not having trouble with wireless default scheduling. I’d favour the development of a DiffServ scheduler on packet layer combined with a default scheduler below. It seems to me that 3 GPP choose different approaches for 4G and 5G.
>>  
>> I wonder which scheduling was recommended for 3GPP access types, if there’s an RFC recommending a priority bearer for L4S at WiFi interfaces.
>>  
>> Regards,
>>  
>> Ruediger
>>  
>> Von: Ingemar Johansson S <[e-mail]> 
>> Gesendet: Montag, 15. März 2021 12:08
>> An: Geib, Rüdiger <[e-mail]>
>> Cc: [e-mail]; [e-mail]; [e-mail]; Ingemar Johansson S <[e-mail]>
>> Betreff: RE: [tsvwg] Fwd: Qs on your 5G L4S slides
>>  
>> Hi Ruediger
>>  
>> I can’t really comment on how this is handled for WiFi. But I also notice that DOCSIS has a mechanism that demotes misbehaving L4S flows into a classic queue.
>>  
>> For 3GPP access already L4S with default bearers gives quite some improvement. 
>> The use of L4S with priority scheduling can enhance performance even more but poses some additional concerns, where the use of a DBS scheduler is one extreme in this context. There are other alternatives such as increased scheduling weight that has a more limited impact on other traffic that runs on default bearers. 
>> But this problem is not unique to L4S. You would face the same issue with e.g., GBR bearers for the cases where an endpoint gets in bad coverage. Additional methods can be needed here to avoid that one bearer gets unduly large share of the radio resources.   
>>  
>> /Ingemar
>>  
>> From: [e-mail] <[e-mail]> 
>> Sent: den 15 mars 2021 11:48
>> To: Ingemar Johansson S <[e-mail]>
>> Cc: [e-mail]; [e-mail]; [e-mail]
>> Subject: AW: [tsvwg] Fwd: Qs on your 5G L4S slides
>>  
>> Hi Ingemar,
>>  
>> That depends. For WiFi, draft-ietf-tsvwg-nqb-05 specifies mapping L4S to a priority bearer based PHB. Then this stops to be an L4S problem. I’d like to be clear about that issue and the question is, whether there will be a recommendation to assign L4S traffic to a 4G or 5G priority bearer. If your answer is no, why is there a draft specifying a priority bearer for WiFi L4S traffic?
>>  
>> The underlying question is, to which extend does the end-to-end performance of L4S depend on suitable radio schedulers coupling two congestion control algos or queuing behaviours, like L4S standardises for fixed line schedulers. And how to operate a network, if these are absent.
>>  
>> Regards,
>>  
>> Ruediger
>>  
>> Von: tsvwg <[e-mail]> Im Auftrag von Ingemar Johansson S
>> Gesendet: Montag, 15. März 2021 10:55
>> An: Smith, Kevin, Vodafone Group <Kevin.Smith=[e-mail]>; Bob Briscoe <[e-mail]>; tsvwg IETF list <[e-mail]>
>> Betreff: Re: [tsvwg] Fwd: Qs on your 5G L4S slides
>>  
>> Hi Kevin, Bob + others
>> CC Davide (thesis author)
>>  
>> Yes, there was a test with the use of the dedicated bearer (DBS) and no-L4S. This is exemplified in section 5.3.6 in the thesis report. In short the outcome is that the background traffic will be severely affected. The reason is that the DBS scheduler (originally devised for e.g. VoLTE) prioritizes a bearer when the queue delay exceeds a given low threshold (e.g 10ms). And because SCReAM without L4S targets larger queue delay, the outcome is that it will hog an unreasonable share of the available resourses. 
>> What this means is that it is necessary to use some extra guard mechanism when prioritized bearers are used, but this is of course not only an L4S problem.
>>  
>> /Ingemar
>>  
>> * http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A1484466&dswid=-2512
>>  
>>  
>> From: tsvwg <[e-mail]> On Behalf Of Smith, Kevin, Vodafone Group
>> Sent: den 12 mars 2021 14:56
>> To: Bob Briscoe <[e-mail]>; tsvwg IETF list <[e-mail]>
>> Subject: Re: [tsvwg] Fwd: Qs on your 5G L4S slides
>>  
>> Hi Ingemar,
>>  
>> Just to ask, was there also a variant of the test with no L4S but with the dedicated bearer? I’d be interested to see that comparison.
>>  
>> @Bob, regarding UPF placement: the ability to virtualise network functions in 5G Core allows easier scaling of UPFs as required. 
>>  
>> All best,
>> Kevin
>>  
>>  
>>  
>>  
>> C2 General
>> From: tsvwg <[e-mail]> On Behalf Of Bob Briscoe
>> Sent: 10 March 2021 17:41
>> To: tsvwg IETF list <[e-mail]>
>> Subject: [tsvwg] Fwd: Qs on your 5G L4S slides
>>  
>> CYBER SECURITY WARNING: This email is from an external source - be careful of attachments and links. Please follow the Cyber Code and report suspicious emails.
>> tsvwg,
>> 
>> Fwd'ing to list, with permission...
>> In case anyone else had the same questions 
>> 
>> 
>> -------- Forwarded Message -------- 
>> Subject: 
>> RE: Qs on your 5G L4S slides
>> Date: 
>> Wed, 10 Mar 2021 14:33:42 +0000
>> From: 
>> Ingemar Johansson S <[e-mail]>
>> To: 
>> Bob Briscoe <[e-mail]>
>> CC: 
>> Ingemar Johansson S <[e-mail]>
>> 
>> 
>> Hi
>> Please see inline [IJ]
>> 
>> /Ingemar
>> 
>> -----Original Message-----
>> From: Bob Briscoe <[e-mail]>
>> Sent: den 10 mars 2021 14:46
>> To: Ingemar Johansson S <[e-mail]>
>> Subject: Qs on your 5G L4S slides
>> 
>> Ingemar,
>> 
>> #5 "Dedicated bearer / QoS flow for L4S traffic"
>> Is this a per-app microflow or a per-user flow?
>> [IJ] It is per-user flows, i.e each bearer can handle many flows
>> 
>> 
>> And I think you'll need to explain where the UPF is typically located. I believe
>> it's close to the edge, isn't i?
>> Further into the network (beyond the UPF) these flows just become an
>> aggregate of all the users.
>> [IJ] The UPF is close to the edge somehow, it is hard to say for certain where they are located, they can be real close to the base stations or >100km away.
>> 
>> 
>> #6 Question:
>> Do you have any feel for qDelay & throughput if a "Classic ECN AQM" like PIE
>> or CoDel was used?
>> [IJ] No, it was not studied in the master thesis work.
>> 
>> 
>> #6 - #11:
>> Is the DBS scheduler between users, or between flows?
>> [IJ] Per user (bearer)
>> 
>> 
>> #12: L4S is meant to greatly reduce the throughput-delay tradeoff, and in our
>> results it did.
>> Any idea why not here? I guess, with video, it's the 'getting up to speed' fast
>> problem (that I'm working on with Joakim).
>> [IJ] One reason is the large variation in frame sizes that video coders generate.
>> Another is that SCReAM paces out the video frames as 50% higher rate than the nominal video target bitrate. This pacing overhead can be configured lower but then the video frames (RTP packets) are more likely to become queued up in the sender instead. I really believe that it can be done better, was hoping to have time to improve SCReAM in this respect but the work hours fly in other directions .
>> With that said. Also a DCTCP flow (with L4S) marking will get a reduced throughput compared to e.g a Cubic flow (without L4S) over cellular. The reason is that the large buffers with Cubic absorb the fast fading dips in LTE and NR. With DCTCP + L4S some extra headroom is needed to avoid queue build up.
>> 
>> 
>> 
>> Bob
>> 
>> --
>> __________________________________________________________
>> ______
>> Bob Briscoe https://protect2.fireeye.com/v1/url?k=828d3ddc-
>> dd1604f1-828d7d47-8692dc8284cb-1ab58b5eb7943901&q=1&e=b0160f51-
>> 6418-41ea-9221-efaca6b7cec8&u=http%3A%2F%2Fbobbriscoe.net%2F
>>  
> 
> -- 
> ________________________________________________________________
> Bob Briscoe                               
> http://bobbriscoe.net/