Re: [tsvwg] L4S vs SCE (Evolvability)

[Bob Briscoe <[e-mail]>]

Roland,

As I just said, I noticed that Koen's useful response snipped some of 
your responses about evolvability that I ought to have addressed.
Sorry, you'll need to reload state from November...

On 22/11/2019 02:32, Roland Bless wrote:
> Hi Bob,
>
> see inline.
>
> On 21.11.19 at 15:44 Bob Briscoe wrote:
>> Roland,
>>
>> I'm not getting it....
>>
>> On 21/11/2019 09:32, Roland Bless wrote:
>>> Hi Bob,
>>>
>>> see inline.
>>>
>>> On 21.11.19 at 19:34 Bob Briscoe wrote:
>>>> On 20/11/2019 21:22, Roland Bless wrote:
>>>>> Yes, but as I also expressed my concerns w.r.t. the L4S codepoint
>>>>> earlier, at the cost of binding this to a quite fixed set of L4S
>>>>> behaviors and "burning" the last ECT codepoint. Personally, I like
>>>>> concepts with a little bit more potential to be useful for future
>>>>> development (evolvability) of congestion controls, e.g., BBRv2 and
>>>>> LoLa could also benefit from an SCE-like marking...
>>>>>
[BB] BBRv2 has not used SCE, but it has used L4S marking. And the 
authors of the LoLa draft have not used SCE but they have merged with 
the NQB draft, so that they can work with L4S as well. So perhaps your 
belief that SCE is "more evolvable" is just that - a belief.


>> This last sentence... please really spell it out what you could possibly
>> be meaning here. What has made you suddenly think this particular
>> marking scheme has got magical properties that bestow evolvability? I'm
>> serious, if you can't explain why you've said a sentence like this, it
>> implies you are under the spell of some cult or fad.
> Not sure what you are trying to indicate by your last sentence, but
> sure, I can elaborate a bit on my last sentence.
> I find it architecturally cleaner to have an additional ECN codepoint
> used for indicating "SCE" rather than (ab)using it as classifier for the
> dualQ AQM.
[BB] I just forked another thread 'ECN as a classifier' to address that 
assertion.
> The SCE markings can be used independently whether you have multiple
> queues or just a single queue.
[BB] This seems to be the opposite of reality. Perhaps you've used the 
word 'can' because you think it might be possible. But there's good 
reason to doubt that. As follows...

Packets from non-SCE and from SCE senders are indistinguishable by just 
looking at them. So, if there's a single queue, they have to be mixed 
together in it {Note 1}. But a single queue can only have one length. 
Any transports that don't understand SCE drive the single queue to a 
deeper operating point (defined by CE or drop). This overruns the SCE 
operating point, and causes SCE marking to approach 100% {Note 2}. Then 
those transports that do understand SCE will starve.

In contrast, L4S works in dual queues or multiple queues (where 'works' 
means it gives very low latency without losing utilization).

Also, there is considerable scope for evolvability with both:
* the dualQ is a framework that is intended to encompass different AQMs 
in future.
* and there is considerable scope for developing the ways FQ uses L4S.

{Note 1}: Unless L4 identifiers are used to separate out microflows, and 
identify low latency flows by their behaviour. But then it's not really 
a single queue, unless you make it look like a single queue, but still 
treat it like multiple queues, as in LFQ. However, schemes like LFQ are 
"single-queue" in name only. They still have all the downsides of 
multiple queues.

{Note 2}: Unless the SCE operating point is hardly any shallower than 
that for CE, in which case SCE gives hardly any improvement.

> If one leaves the coupling aside the
> algorithm for marking will probably not differ so much from what is
> proposed in L4S.
[BB] You still haven't explained (rather than asserted a belief) why the 
choice between SCE and L4S has anything to do with evolvability.

>>>> My whole purpose in solving the problem of deploying scalable CCs over
>>>> the Internet was to re-juvenate evolution (to widen the range of
>>>> applications that could be supported by different transport behaviours,
>>>> particularly for real-time with low latency and high throughput at the
>>>> same time). One of the main things that has stopped CCs evolving so far
>>>> is the need for friendliness with the Reno behaviour that was not
>>>> scaling over the years.
>>> FWIW, our delay-based approach has deployment problems at the
>>> other end of the spectrum: it gets suppressed by loss-based CCs...
>>> (and we don't want to sacrifice our low delay goal).
>>>
>>>> If SCE is primarily supported in FQ AQMs, that will constrain flows to
>>>> be capped at the rate that FQ gives them. How is that doing anything
>>> I was solely referring to the marking behavior of SCE, not a particular
>>> implementation based on FQ-AQMs.
>> This implies you believe all that silliness about a shallower SCE
>> threshold not starving an SCE flow in a single queue, because it makes
>> SCE flows not want to use the queue.
> I do not understand what you are saying here.
[BB] I've now explained this above (after "But there's good reason to 
doubt that...").

The silliness I'm referring to is the statement by Jonathan M in his 
tsvwg talk at IETF-106 that CNQ is backward compatible with RFC3168 
traffic, because its performance for SCE traffic when mixed with 3168 is 
bad enough that it discourages use of SCE alongside 3168.

>>>> other than massively constraining future evolution of CCs, especially
>>>> real-time ones? See Per-Flow Scheduling and the End-to-End Argument
>>>> <http://bobbriscoe.net/projects/latency/per-flow_tr.pdf>. I don't need
>>>> to tell you that the e2e argument is all about giving end systems the
>>>> power to innovate without permission.
>>>> Anyway, what are you imagining would stop CCs evolving alongside other
>>>> scalable CCs? In much the same way CCs have always evolved. With L4S you
>>>> have a clean slate that seems just like a FIFO with a shallow ECN-only
>>>> immediate AQM. And other flows are causing you hardly any delay and very
>>>> rarely any loss. Think of all the things you can do with that. Go
>>>> evolve, Roland!
>>> I already said that the Dual Queue AQM is nice, but comes with the
>>> problem due to its built-in dropping law. Other CCs may react differently
>>> and BBR was one example of newer CCs that did not work in the L4S queue
>>> without further adaptation.
>> The built in dropping law is for the old traffic ('pre-evolved' if you
>> like) that doesn't respond to anything else but loss. That's what
>> drop-based AQMs were for - to fool Reno etc into thinking they had hit
>> the buffers, so to speak.
> Yes, I understand that. But what happens if the classic traffic
> once vanishes and we only have low-delay congestion control and
> want to re-use the "second" queue for other purposes, e.g., using
> it for flows that do not use loss-based CC?
[BB] Let me repeat back the question, 'cos it seems rather odd, so 
perhaps I've misunderstood.

You want me to assume that the classic behaviour of repeatedly seeking 
out more capacity until loss occurs doesn't exist on the Internet any 
more. There's only low-delay CC, by which I think you mean CC based on 
L4S ECN (but I'm not sure that's your meaning).

Then, it sounds like everything would be working well. So why do you 
want to re-use the "second" queue for other purposes? Nothing is wasted 
if you don't use it. It has no capacity associated with it. it's just 
some lines of code sitting there in case packets appear from a classic CC.


>>>    > The key thing here is the wording of the Prague requirements
>>>> <https://tools.ietf.org/html/draft-ietf-tsvwg-ecn-l4s-id-08#section-4>.
>>>> We have a session in the 'Prague' side-meeting tomorrow to review them
>>>> (and I encourage this on this list too).
>>>>
>>>> Later down the line, if the L4S experiment is successful, there will be
>>>> an opportunity to review these requirements if a standards track doc
>>>> replaces the experimental (it is easier to relax CC requirements than
>>>> tighten them). So, for the expt track, the requirements are designed to
>>>> protect competing flows from harm in a tighter way than you might find
>>>> in RFC5033 or similar.
>>>>
>>>>
>>>> Nonetheless, even the 1/p isn't tightly spec'd, quoting:
>>>>
>>>>      The inverse proportionality requirement above is worded
>>>>      as a 'SHOULD' rather than a 'MUST' to allow reasonable flexibility
>>>>      when defining these specifications.
>>> Yes, but something has to be implemented (in hardware) and therefore it
>>> is fixed for some time and it should be consistent along a path, so I
>>> don't see a viable path for evolvement of this coupling law...
>> Why does the coupling law need to evolve? It's a relationship between
>> something invariant (scalable) and the past (which is over, by
>> definition, so it's not going to change now).
> See above, probably we want to put something into queue which is
> using a non-loss-based congestion control and/or we need to change
> the (1/p,1/p^2) marking.
[BB] The queue doesn't induce loss, the CC does. The queue isn't 1/p or 
1/p^2 or whatever, the different CCs are. So, if a CC is using the C 
queue without inducing any loss (e.g. a delay-based CC might be able to 
keep the queue under the AQM's target delay), then the coupling won't 
couple any marking across to the L queue.

But I can't imagine why a delay-based CC that can keep the queue below 
the target of the AQM would want to classify itself into a queue 
designed for CCs that need to induce a decent queue (in order to 
maintain full utilization), given such queue-inducing traffic might 
start up at any time.

>> If you want to evolve, you select what's best for you - probably the
>> nice clean L4S ECN queue. I still don't get what sort of evolution you
>> can't do? Evolvability isn't about a researcher being able to do what
>> they'd like.
> I also don't understand it in this way, but investigating
> invariables and degrees of freedom prudently may enable or facilitate
> deployment of new stuff. If that new stuff cannot be deployed
> it never gets a chance of being weeded out later either.
[BB] If you have something specific in mind, please say it. This is all 
getting rather abstract.

This does make me want to question your notion of evolvability. Usually, 
when we make sure the Internet supports evolvability, we mean 
evolvability of new applications. We don't mean the ability for 
researchers to come up with different ways to solve problems that are 
already solved. There seems to be a hint of a complaint in your emails 
that L4S doesn't leave room for researchers to solve the same problem 
differently. That sort of evolvability is rather a luxury isn't it? If 
one approach solves an enduring problem with the Internet, it would be 
rather churlish to say "No, we can't use this solution, because it might 
preclude some ideas that might lead to other solutions to the same 
problem in the future, but we're not sure."

Having said that, I hope I (and Koen) have shown that L4S still leaves 
scope for complementary delay-based CCs, not to mention scope for 
different AQMs and for different FQ-based solutions.


Regards



Bob



>
> Roland
>

-- 
________________________________________________________________
Bob Briscoehttp://bobbriscoe.net/