Re: [Accord] presentation slot for draft-you-tsvwg-latency-loss-tradeoff

[Brian Trammell <[e-mail]>]

hi Gorry,

> On 22 Mar 2016, at 10:23, [e-mail] wrote:
> 
> 
> I'm not finding this ID very easy to get into... and I think there may be
> various reasons, but I suspect David may be digging in a good place to
> start.
> 
> This looks and feels like a -00 draft.

As indeed it is. :)

> From the current, I still need to
> understand why the code points would be needed and why new ones would
> help.

Quoting myself, below:

>> ... it really does seem to be that all existing
>> Diffserv PHBs have shown themselves to be irrelevant for the case we are
>> addressing here -- applications on the content side signaling loss/latency
>> tradeoffs to access networks across the open Internet.


>> The reason to ask for new codepoints is to make it *possible* to continue
>> blocking all existing PHBs while not bleaching PHBs that are not tied in
>> any network (whether by definition or by local configuration) to a request
>> for "better" treatment. A firewall rule could be configured to pass La and
>> Lo (along with other potential future explicit-tradeoff codepoints) while
>> still bleaching other PHBs to DF.


To elaborate:

Every marking other than DF is an explicit request for treatment which will disadvantage DF traffic. There is no mechanism for tradeoff in the DS architecture outside legal contracts between two directly connected networks, which limits its deployability in the general (interdomain, open-Internet) case.

Lo and La are defined such that there is an *explicit tradeoff*. The hope is that by putting these new codepoints in part of the code space that is always subject to these explicit tradeoffs -- marking Lo or La need not disadvantage DF -- makes it possible to bleach the existing DS architecture codepoints, which are used internally but cannot deploy in the general interdomain case due to mark inflation, while passing the interdomain-friendly tradeoff-based codepoints, which are defined such that there is no incentive for mark inflation.

This is, I admit, perhaps laughably optimistic. But it would allow the treatment to deploy on current hardware, as DSCP-aware boxes can already make forwarding and queueing decisions based on DSCP codepoints.

> I think previous experience in TSVWG suggests a lack of precise
> language can surely end with rathole discussions.
> 
> I can see value in a heads-up for people who know and care about Diffserv
> deployment to be encouraged to show-up at ACCORD.

Yes, please. :)

> However, I continue to be really confused about which drafts may be linked
> to the ACCORD BOF, and what this could lead to, that doesn't stop me
> thinking that the BoF could be a good place to start discussion.

Broadly speaking, there are two things I hope we accomplish with the ACCORD BoF:

(1) Have a discussion about how radio access networks (RAN) work, and how this architecture doesn't fit with common IETF perceptions about what happens at layer 2.

(2) Determine the minimum set of information that RAN schedulers actually need (i.e., in deployed networks, not in the 3GPP architecture) about application traffic to function correctly in the absence of traffic classification provided by DPI.

We believe the loss-latency tradeoff bit to be a potential (partial) solution to the problem in (2), given how few different classes of traffic seem generally to be deployed.

Cheers,

Brian


>> Hi, David,
>> 
>>> On 21 Mar 2016, at 04:07, Black, David <[e-mail]> wrote:
>>> 
>>> I'd think that this draft is also planned for presentation in the ACCORD
>>> BOF - how is the discussion in TSVWG expected to differ from ACCORD?
>> 
>> In ACCORD, we're planning very short "here's what Lo and La mean, we have
>> a draft that defines Lo/La-over-DSCP, we also have a use case (in
>> draft-kuehlewind-spud-use-cases) for Lo/La-over-SPUD". The presentation in
>> TSVWG would get into the details of why we think we need new DSCP
>> codepoints for this.
>> 
>> If we assume everyone who might be interested in this will be in ACCORD
>> anyway, we can compress our discussion of "why do we need new codepoints"
>> into that single presentation.
>> 
>>> I'm concerned about the draft's brief dismissal of each existing
>>> Diffserv PHB as irrelevant with very little explanation.
>> 
>> As am I. But I personally consider that a problem with interdomain
>> deployability of existing Diffserv PHBs: it's not really what they were
>> designed for. I'd be happy to be shown counterexamples of n>2-domain
>> deployments at scale involving networks not party to a contract involving
>> DSCP-implemented SLAs, but it really does seem to be that all existing
>> Diffserv PHBs have shown themselves to be irrelevant for the case we are
>> addressing here -- applications on the content side signaling loss/latency
>> tradeoffs to access networks across the open Internet.
>> 
>> The reason to ask for new codepoints is to make it *possible* to continue
>> blocking all existing PHBs while not bleaching PHBs that are not tied in
>> any network (whether by definition or by local configuration) to a request
>> for "better" treatment. A firewall rule could be configured to pass La and
>> Lo (along with other potential future explicit-tradeoff codepoints) while
>> still bleaching other PHBs to DF.
>> 
>> We could potentially do a better job of clarifying this argument in a
>> future draft revision.
>> 
>>> Here are some combinations of existing PHBs that have the potential to
>>> deliver something like the desired behavior:
>>> 
>>> - DF (Default) & EF
>>> - DF (Default) & AFx1 (have to pick an x value)
>>> - AFx1 & EF (have to pick an x value)
>>> - AFx1 & AFy1 (have to pick different x & y values)
>>> 
>>> In all cases, some specific queue configuration is involved to obtain
>>> the desired results, particularly for the queues that correspond to AFx1
>>> & AFy1.  I've included EF because its queue behavior matches latency
>>> preference well - a queue for EF is supposed to deliver low latency in
>>> part by dropping traffic that it can't forward promptly.
>> 
>> Either of these against DF are "request for better treatment": we need
>> some way to distinguish Lo from DF treatment. Though they'd basically be
>> implemented the same way *today*, one can envision future networks where
>> ubiquitous multipath connectivity using different radio technologies means
>> that Lo really *is* a request to go over the lower-loss path.
>> 
>> So we can disregard La = AFx1 / Lo = DF as well as La = EF / Lo = DF.
>> 
>>> OTOH, EF's admission control requirement feels like a poor fit for
>>> what's wanted here.
>> 
>> Yep, so I think we can disregard La = EF / Lo = AFx1
>> 
>> AFx1/AFy1 is intriguing, though rereading 2597 it appears that the
>> "assured" in assured forwarding doesn't quite fit (2597 section 2 para 3):
>> 
>>   A DS node MUST allocate a configurable, minimum amount of forwarding
>>   resources (buffer space and bandwidth) to each implemented AF class.
>>   Each class SHOULD be serviced in a manner to achieve the configured
>>   service rate (bandwidth) over both small and large time scales.
>> 
>> For now, a DS node allocates buffer to Lo and La (where the La minimum
>> might be as small as a single packet), and will practically have to
>> allocate a bandwidth floor for each. But, as above, the actual
>> implementation in the future might rather involve completely different
>> forwarding paths. While these would meet the letter of this paragraph, as
>> there would be an implicit bandwidth floor on each next hop, conceptually
>> the bandwidth requirement is orthogonal to the Lo/La tradeoff, and as such
>> I don't think really fits what I read as the "spirit" of AF.
>> 
>> In addition, this use might collide with some providers' internal use for
>> these codepoints. Of course, this is a potential problem with new
>> codepoints in Pool 3, as well, since these were given over for potential
>> local use in the IANA registry. And it doesn't solve the problem with
>> declaring codepoints potentially unbleachable at all.
>> 
>> Thanks, cheers,
>> 
>> Brian
>> 
>>> I'm not trying to say that this is a solved problem, but I would like to
>>> see more attention paid to possible reuse and/or adaptation of what's
>>> been done before.
>> 
>>> Thanks, --David ++Sent from Android smartphone ...
>>> -------- Original message --------
>>> From: Gorry Fairhurst <[e-mail]>
>>> Date: 3/17/2016 11:02 AM (GMT-05:00)
>>> To: Brian Trammell <[e-mail]>, [e-mail]
>>> Cc: [e-mail]
>>> Subject: Re: presentation slot for draft-you-tsvwg-latency-loss-tradeoff
>>> 
>>> On 17/03/2016 14:49, Brian Trammell wrote:
>>>> Greetings,
>>>> 
>>>> I'd like to request a slot (if one hasn't already been requested) to
>>> quickly present draft-you-tsvwg-latency-loss-tradeoff at the most
>>> appropriate tsvwg session in Buenos Aires.
>>>> 
>>>> Thanks, cheers,
>>>> 
>>>> Brian (for the authors of the loss-latency "one bit" draft)
>>>> 
>>> I'll note the request. I did glance through the draft, and I'm not sure
>>> quite how this fits in to the DS architecture, but I'd need to look
>>> further and consult more to gte a better picture.
>>> 
>>> I need to speak to my co-chair about planning the Agenda, and we'll
>>> discuss - and get back next week.
>>> 
>>> Gorry
>> 
>> 
>