Re: [Udp35] TAPS BOF and moving the transport API forward

[Brian Trammell <[e-mail]>]

On 26 May 2014, at 12:17, [e-mail] wrote:

> Hi all,
> 
> I've been struggling to see the synergy between UDP 3.5 and TAPs - I'll
> try and work through the emails again, but it seems to me that I need
> help:  I don't yet see where you people think we have arrived.

So there are a couple of things we're trying to address here:

(1) TAPS forsees a selection/negotiation mechanism, but if we have a selection/negotiation mechanism that ends up telling us "sorry, on this path you get to use SOCK_STREAM because I (the kernel) only support that and SOCK_SEQPACKET for reliable transports, and this path isn't SCTP-transparent" then we've built a bunch of machinery to tell us that transport is broken, which we knew in the first place. And I'm afraid this will be the case more often than not. We need tools in the toolbox beyond those we already have, and "roll-it-yourself over UDP" is one of those tools.

(2) We need to look into ways we can rescue "roll-it-yourself over UDP" from creeping middlebox breakage. Happy/angry* eyeballs is one approach here, but there may be things that can be done in a UDP shim layer that can help too.

I think we need to make progress down the path of solving these problems in parallel with figuring out the right way to decompose the dimensions of transport.

(* "angry eyeballs" -- the logical conclusion of escalating aggression in happy eyeballs across multiple dimensions; in other words, we can only go so far down this path alone before we have to be very careful that we're not part of the problem.)

> To me, what I understand dimly of UDP 3.5 seems to be coming from that
> areas that inspired work on applications-layer building blocks to assemble
> protocols in user-land.

This is one reason "UDP" is in the name: if you want raw sockets in userland without privilege on the machine, UDP is what you get. We can try to fix the privilege problem here in the meantime, but probably not within the IETF.

> At one time (history) the IESG/IAB turn focus
> pushed towards kernel-transports and things changed track and the
> development went in a different direction. I think the tide reversed some
> time ago. In the IETF, RMT tried something different, and that's another
> story. Some rethinking may be useful here, what I am not sure yet -
> because I haven't really seen this presented.

Brainstorming in front of a whiteboard on this point is one of the main reasons we want time in a smaller room in Toronto.

> I do see some synergy between the thread with the ideas of building TAPS
> components to handle specific protocol mechanisms - SPROUT tried something
> else, and IMHO exposed the weakness of this, in that everything becomes
> tuneable and basic stability or ability to share capacity robustly is
> compromised or at least needs to match some overall guidelines (or the
> danger is we ignore the interactions needed for inclusive support of all
> types of path and all types of apps sharing the path).

One of the things we'd want to see come out of the UDP35 effort is a recommended restriction of the parameter space for congestion control when rolling your own: "here's how to compose a transport protocol without breaking the Internet or rediscovering CC by yourself".

> But this idea of
> modular design was not what I thought of as  the goal of TAPS.. (maybe my
> vision was different?)

This is... not so much a complimentary vision, as an activity to start addressing what we think the next problems will be once one goes down the TAPS path...

> In TAPS, I was quite excited at the potential to find a set of robust ways
> that the transport framework could figure out how to best use a network
> path - and then choose an appropriate transport.

...these being: (1) a transport framework that only has one transport to choose from is not extremely useful, and (2) a transport framework that requires new kernels and libraries everywhere is less quickly deployable than one relying only on application-layer changes.

> There has long been a
> need for a transport transition/selection framework.  Sure if the
> framework succeeds, it would really open itself to enabling the building
> block method in user-land (with all that involves).

... which is why I'd like those building blocks to be ready for the framework once it's done.

> The thing that is
> really bewildering me (I may be wearing different spectacles to you) is
> why you think the expertise lies in RAI and APPS to work out how to do
> this?

It's not RAI and APPS that we need to work out how to make it work, it's that we need people who aren't wearing transport spectacles to help figure out what's useful. These areas represent the communities who will choose to use (or to ignore) TAPS.

> I think it is IMPORTANT to get these concepts/boundaries clear. I probably
> don't understand the "derails" in London (maybe because I partly
> contributed to them), but most of what I heard within the BoF seemed to be
> helping define what should/should not be done. But maybe I missed some
> later discussion?

Much of what I labelled "derail" was the API discussion, because we spent too much time talking about the details of non-transport APIs (and about whether the IETF should be doing APIs at all) as opposed to the abstract interfaces necessary. But that might also be a function of my expectations...

> I think it is important that TAPS meets in some way at the IETF (losing
> this chance could stall things - and hawaii may turn out to be an odd ietf
> meeting for many outside the IETF).

(...I'll not be able to make Hawaii, for one...)

> My understanding on how and why the
> group meets would depends on understanding more of this thread.

The intention of the meeting on Saturday is more or less to continue the discussion we're having right now, in person, in front of a whiteboard; the format is "by-invitation brainstorming session" as opposed to an IAB workshop, for example, mainly because the whiteboard's there, and larger workshops are less productive per unit time.

Beyond that, we're planning on introductory talks about the outcome of that meeting in dispatch, appsarea, and tsvarea. It sounds like there's an open request before the TSV ADs to have a WG-forming BoF for TAPS with a smaller scope, as well, and some of the discussions on Saturday could feed into that BoF as well.

Cheers,

Brian

>> hi Michael, all,
>> 
>> On 26 May 2014, at 10:03, Michael Welzl <[e-mail]> wrote:
>> 
>>> ... so, now:
>>> 
>>> 1) we agree that a few things need to be done ASAP no matter what
>>> (define the "supply side")
>>> 2) we may have to refine our views a little bit about what needs to be
>>> done immediately
>>> 
>>> - and this should not be done as an update to the TAPS charter, leading
>>> to a WG-forming TAPS BoF in Toronto, because... ?
>> 
>> ...since the responsible ADs are the ones who make the decision here, if
>> their opinion differs from mine, what they say goes; but...
>> 
>> ...because IMO (1) it's not clear that having a TAPS WG in the TSV area
>> would give us the best combination of efficiency and wider participation
>> from the communities (represented by RAI and APP) that would use it, (2)
>> I'm afraid we'd have the same derails we had in London which would take
>> cycles away from work we already have a pretty clear boundary drawn
>> around, (3) there are _clearly_ things outside the scope of TAPS that are
>> nonetheless related (essentially anything under the "signaling" heading,
>> which itself seems split into two things) s.t. we need more understanding
>> about those relationships, and (4) the current scope of agreement is a
>> single document (albeit a somewhat large and quite important one), which
>> would appear to me to draw heavily on the present draft-monpetit-* (unless
>> the current group had something else in mind), for which WG formation
>> seems kind of a heavyweight process.
>> 
>> Indeed, for a TAPS WG we'd need to find two non-author co-chairs, who
>> would have to work pretty hard to avoid a derail on the API issue at a
>> Toronto BoF, and keep the charter that comes out of the BoF focused enough
>> that all the new contributions you'd get would further the current work on
>> this document. My reticence here comes from a suspicion that this will be
>> hard to do.
>> 
>> I don't see how "not creating a working group" equates with "delaying the
>> work"; indeed, you've already got a mailing list, a core team of
>> contributors, documents in pretty good shape, etc.; i.e. you've already
>> pretty much built a working group. Now, if there is some (political)
>> reason that the work needs to be seen as being done within a WG in the
>> IETF, let's talk about that. But otherwise it seems to me we could publish
>> this one document on the IAB stream as a product of the
>> (soon-to-be-renamed TCP/)IP Evolution program in the IAB with somewhat
>> less overhead, and use this (more architectural) document after
>> publication as background for engineering work done in (a) future WG(s).
>> 
>>> Don't get me wrong, I understand the need for the meeting on Saturday in
>>> Toronto, indeed there are more things that we should try to agree on,
>>> but since we already agree (to some level of detail) what TAPS should be
>>> doing, why do we delay it?
>> 
>> I'm not at all adamant about this; if there's wider agreement now that the
>> "dimensions-of-transport" work is probably best done in a TSV area working
>> group, we could also parallelize this: put together a smaller charter as a
>> starting point for a Toronto WG-forming BoF to which the outcome of
>> Saturday's discussions could be input.
>> 
>> (But on that point I defer to the responsible ADs. :) )
>> 
>> Cheers,
>> 
>> Brian
>> 
>>> If the idea of having a WG seems problematic (slowing things down for
>>> reasons I don't understand), is that still the case with the scope on
>>> "supply side"? I don't quite get that.
>>> 
>>> Cheers,
>>> Michael
>>> 
>>> 
>>> 
>>> On 23. mai 2014, at 10:34, Michael Welzl wrote:
>>> 
>>>> 
>>>> On 22. mai 2014, at 18:38, Brian Trammell wrote:
>>>> 
>>>>> 
>>>>> On 22 May 2014, at 16:13, Michael Welzl <[e-mail]> wrote:
>>>>> 
>>>>>> 
>>>>>> On 22. mai 2014, at 14:05, Brian Trammell wrote:
>>>>>> 
>>>>>>> hi Michael,
>>>>>>> 
>>>>>>> (copying the udp35 list)
>>>>>>> 
>>>>>>> On 22 May 2014, at 12:56, Michael Welzl <[e-mail]> wrote:
>>>>>>> 
>>>>>>>> Hi,
>>>>>>>> 
>>>>>>>> Here's a suggestion. I'm cc'ing the mailing list because in case
>>>>>>>> you disagree this isn't helpful - after all I'm missing context of
>>>>>>>> what happened at your retreat. But I do have a suspicion that we're
>>>>>>>> all on the same page here, so I'll try a simple suggestion:
>>>>>>>> 
>>>>>>>> When I first approached the IETF, my take was to go bottom-up:
>>>>>>>> define the services of existing IETF transports, create a
>>>>>>>> protocol-independent socket API from that. Not because I think that
>>>>>>>> this is what the long-term solution should be, but because I think
>>>>>>>> this is needed as a starting point.
>>>>>>> 
>>>>>>> I see the starting point differently, though also bottom-up:
>>>>>>> 
>>>>>>> (1) define the dimensions of transport services...
>>>>>>> (1a) provided by existing IETF transports, and/or
>>>>>>> (1b) required by existing applications (e.g., there's a gap in
>>>>>>> RTCWEB here plugged by SCTP/DTLS)
>>>>>>> (2) determine other combinations of these transport service
>>>>>>> dimensions that make sense (e.g. Minion occupies another spot in
>>>>>>> this space)
>>>>>>> (3) make recommendations for combining these dimensions in
>>>>>>> user-space transport protocols (in a way that leads to interoperable
>>>>>>> implementations of these; this is the "mix-ins" idea); this requires
>>>>>>> us to
>>>>>> 
>>>>>> "mix-ins" => do you have a pointer? Sorry if you presented it at an
>>>>>> IETF and I forgot about that.
>>>>> 
>>>>> Nope, not yet; the idea is still not quite fully-formed. Essentially,
>>>>> instead of providing an implementation of an API that has various
>>>>> knobs for various features (the SCTP approach), we provide a set of
>>>>> standards/implementations that each implement a small set of choices
>>>>> for dimensions as a coherent whole, which are all designed to play
>>>>> nicely with each other. This is sort of a "userspace transport
>>>>> construction kit" as opposed to a layer which selects an existing
>>>>> transport.
>>>> 
>>>> That sounds very similar to this:  http://nena.intend-net.org/
>>>> 
>>>> 
>>>>>>> (3a) define a standard method for implementing user-space transport
>>>>>>> protocols that is deployable on the present Internet, preferably in
>>>>>>> a way that does not explicitly escalate the middlebox arms race.
>>>>>>> (4) (possibly) make recommendations for the interface that
>>>>>>> implementations of (3) should provide to applications.
>>>>>> 
>>>>>> I'm not sure I buy the need and/or feasibility of 3a, but I'm also
>>>>>> not sure we need to agree about this, at this point in time. Probably
>>>>>> it's enough if we agree on the starting points  :-)
>>>>> 
>>>>> Yep. (3a) is also something which is can be done almost completely
>>>>> independently of the rest, so will probably be a separate thread
>>>>> within the "udp35" work.
>>>> 
>>>> ACK, let's stop discussing it for now to keep this discussion focused.
>>>> 
>>>> 
>>>>>>>> At the Vancouver bar BOF, I was surprised that folks wanted more:
>>>>>>>> support of higher-level APIs, not just being bound to current
>>>>>>>> RFC-defined transports, etc. This is good, but it does create a
>>>>>>>> larger space of work, and is potentially a much longer-term story,
>>>>>>>> which is why I didn't think we could go there in the first place.
>>>>>>>> 
>>>>>>>> => would it now perhaps be the right time to split this into:
>>>>>>> 
>>>>>>> It seems to me that splitting the effort in to "things we can/should
>>>>>>> do now" and "things we are not sure about" is a very good idea. I
>>>>>>> have some detail-level disagreements with where to draw the line, I
>>>>>>> think.
>>>>>>> 
>>>>>>>> 1) an IETF TAPS WG that does a *socket* API extension only, based
>>>>>>>> on an analysis of given transports, and a recommended way of
>>>>>>>> implementing this (happy eyeballing ++ )
>>>>>>> 
>>>>>>> It seems like (1) and (2) above mostly overlap between the way we've
>>>>>>> been looking at this in udp35 and the path TAPS has taken; this
>>>>>>> intersection could be the scope for a WG-forming BoF or a direct WG
>>>>>>> formation, after a bit more discussion to nail down the differences.
>>>>>> 
>>>>>> I agree (although my proposal for a starting point was even narrower;
>>>>>> I'm fine with broader, whatever works...)
>>>>> 
>>>>> I think (2) may simply be a set of recommendations in the document
>>>>> describing (1); we need to acknowledge that the dimensions of
>>>>> transport services are not orthogonal.
>>>> 
>>>> I agree - I wasn't so concerned about (2) though, I actually meant that
>>>> we might not even have to do (1b) yet. Perhaps, as long as we stick
>>>> with *existing* applications, as you wrote, that could be fine - but
>>>> how to draw the line? We're talking about using a future API, so we ARE
>>>> talking about a change to the app. Then again, for an intermediate
>>>> first step, I think the more futuristic fancy use cases set us on a
>>>> track that leaves TSV territory and may be a much longer term activity,
>>>> a broader (but also more interesting!) story (consider the breadth of
>>>> http://tools.ietf.org/html/draft-montpetit-transport-use-cases-01 and
>>>> http://tools.ietf.org/html/draft-deng-taps-datacenter-01 for example).
>>>> 
>>>> 
>>>>>>> I'm concerned (as in the previous message) that doing this in a TSV
>>>>>>> area WG will keep us from getting good cross-area participation, so
>>>>>>> another possibility is forming an IAB program around this activity.
>>>>>>> 
>>>>>>> In any case (1) should definitely continue moving forward regardless
>>>>>>> of venue.
>>>>>> 
>>>>>> ... and (1a) doesn't have the cross-area participation problem, it
>>>>>> can be done in TSV.
>>>>> 
>>>>> I'm still not 100% convinced but this is not a point we need to settle
>>>>> right now. :)
>>>>> 
>>>>>>> I'm not convinced building (4) makes a lot of sense without (3a).
>>>>>>> But discussing that further is what this list is for. :)
>>>>>> 
>>>>>> I agree. Can you elaborate on (3): why, what? I'm not sure I'm with
>>>>>> you here...
>>>>> 
>>>>> (see above, hope that's useful)
>>>>> 
>>>>>> 
>>>>>>>> 2) an IRTF activity of some sort that is a home for:
>>>>>>>> - how to map higher-level APIs onto a lower socket API
>>>>>>>> - how to define more general services ("give me low latency instead
>>>>>>>> of high bandwidth")
>>>>>>> 
>>>>>>> I'm not sure what's different between these two points and "socket
>>>>>>> API extensions only" in the TAPS WG scope above?
>>>>>> 
>>>>>> Sorry for not making it clear enough: with "socket API extensions
>>>>>> only" I meant something that's low-level, i.e. no mapping onto
>>>>>> higher-level APIs, no general services, but exposing the things that
>>>>>> transport currently do, at the abstraction level at which they do it,
>>>>>> without exposing the protocol. Exactly the exercise we did in:
>>>>>> http://heim.ifi.uio.no/~michawe/research/publications/futurenet-icc2011.pdf
>>>>>> http://heim.ifi.uio.no/~michawe/teaching/dipls/stefan_joerer.pdf
>>>>>> 
>>>>>> => I think this is the stuff we really need to do no matter what,
>>>>>> it's your (1a). It's not enough to solve the problem we want to
>>>>>> solve, but we won't get anywhere without having done that.
>>>>> 
>>>>> Got it. Yep. Agreed.
>>>>> 
>>>>>>>> - what services could/should be defined, based on use cases, which
>>>>>>>> aren't yet available
>>>>>>> 
>>>>>>> This third point covers my point 1b/2, and I think is also necessary
>>>>>>> to do earlier rather than later...
>>>>>> 
>>>>>> I'm fine with doing these things early - but pretty much anything
>>>>>> beyond your (1a) (plus signaling, as you suggest - see the second
>>>>>> email, coming..) will make things more complicated community-wise as
>>>>>> this then goes across areas. (1a) is a much simpler case.
>>>>> 
>>>>> The only question remaining in my mind about this work item, then, is
>>>>> whether or not it makes sense to separate 1a (a "supply side" survey
>>>>> of existing services) from 1b (which would be an expansion beyond
>>>>> what's already been done...).
>>>> 
>>>> I like the "supply side" phrasing! Indeed that's what 1a is. I think
>>>> the separation from 1b makes sense to get started, unless we can define
>>>> rules that place clear limits on 1b.
>>>> 
>>>> (one such rule that limits applications to what they are today could
>>>> be: "only consider what's already being offered by existing
>>>> higher-level APIs" => and, lo and behold, you're back at the TAPS
>>>> charter  ;-)   ).
>>>> 
>>>> Cheers,
>>>> Michael
>>>> 
>>>> _______________________________________________
>>>> Udp35 mailing list
>>>> [e-mail]
>>>> https://www.ietf.org/mailman/listinfo/udp35
>>> 
>>> _______________________________________________
>>> Udp35 mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/udp35
>> 
>> 
> 
> _______________________________________________
> Udp35 mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/udp35