Re: [OPS-DIR] Opsdir last call review of draft-ietf-tsvwg-sctp-ndata-12

[Benoit Claise <bclaise@xxxxxxxxx>]

On 8/31/2017 11:01 PM, Spencer Dawkins at IETF wrote:

Hi, Michael,

On Thu, Aug 31, 2017 at 2:56 PM, Michael Tuexen <tuexen@xxxxxxxxxxxxxx> wrote:

> On 31. Aug 2017, at 15:46, Spencer Dawkins at IETF <spencerdawkins.ietf@xxxxxxxxx> wrote:
>
> Hi, Michael,
>
> On Thu, Aug 31, 2017 at 8:25 AM, Michael Tuexen <tuexen@xxxxxxxxxxxxxx> wrote:
> > On 31. Aug 2017, at 14:50, Spencer Dawkins at IETF <spencerdawkins.ietf@xxxxxxxxx> wrote:
> >
> > Just interjecting ...
> >
> > On Thu, Aug 31, 2017 at 7:25 AM, Michael Tuexen <tuexen@xxxxxxxxxxxxxx> wrote:
> >
> > > On 31. Aug 2017, at 14:14, Benoit Claise <bclaise@xxxxxxxxx> wrote:
> > >
> > > Hi Michael,
> > >>> On 30. Aug 2017, at 18:51, Benoit Claise <bclaise@xxxxxxxxx> wrote:
> > >>>
> > >>> Michael,
> > >>>
> > >>> I stumbled on the same point as Stefan.
> > >> Hi Benoit,
> > >>
> > >> thanks for reviewing the document. See my comments in-line.
> > >>
> > >> Best regards
> > >> Michael
> > >>> First of, I had to review RFC4960 for TSN/SSN. I was not sure whether TSN was per stream or not
> > >>> It would have helped me to repeat this information in the intro.
> > >>> Ok, to be candid, I deduced this information from figure 1 and 2.
> > >>>>>>> And the rest are only musings not needing any action if the authors don't want
> > >>>>>>> to action on them:
> > >>>>>>>
> > >>>>>>> The introduction presents a good use case, quoting: "e.g., when the
> > >>>>>>> transmission of an urgent
> > >>>>>>>  message is blocked from transmission because the sender has started
> > >>>>>>>  the transmission of another, possibly large, message".
> > >>>>>>>
> > >>>>>>> The later queueing example in figure 1 has three SIDs being queued
> > >>>>>>> simultanseously. It is not clear which of those "has already started" and which
> > >>>>>> There is actually one message queued for stream 0, three messages for stream 1,
> > >>>>>> and one message for stream 2.
> > >>>>>>> are the important ones being delayed. For a better understanding of the reader,
> > >>>>>> The figure is about in which sequence they are put on the wire. So no
> > >>>>>> transmission has startet. This examples also does not deal with the case
> > >>>>>> of one stream having a higher priority than another. That would be an
> > >>>>>> example using a priority scheduler. This examples illustrates the round
> > >>>>>> robin scheduler.
> > >>>>>> The point here is that a single scheduler (the round robin scheduler, for
> > >>>>>> example) behaves differently when user message interleaving is used or not.
> > >>>>>> That is an important point: You can even use the schedulers with regular
> > >>>>>> DATA chunk, i.e. with user message interleaving being negotiated.
> > >>> This was my source of confusion.
> > >>> With figure 1 and figure 2, you want to show the effect of Round Robin Scheduler with and without User Message Interleaving, while we were expecting the figure 2 to be the solution from this spec. Hence Stefan and I were looking for this "urgent message" in figure 2.
> > >> I don't know what you mean with "solution from this spec" means. This spec introduces
> > >> stream schedulers and a protocol extension to allow interleaving of user messages.
> > >> Using this to implement an "urgent" concept is only one possibility: usage of a
> > >> strict priority scheduler (SCTP_SS_PRIO). So this "urgent" is misleading. It seems
> > >> to ring the wrong bells. That is why I suggested to remove it...
> > >>>>>>> it might be useful to revise the figure so that the head-of-line blocking
> > >>>>>>> happens for SID 1 because transmission of  0 and 2 has already started (and
> > >>>>>>> declare SID 1 to be the "important" message). The ASCII art would just have to
> > >>>>>>> indent SID 1 content a bit (placing 0 and 2 earlier in time), and the resulting
> > >>>>>>> serialisation would then put all the SID 1 messages at the very end, when 0 and
> > >>>>>>> 2 have been completely submitted (right?).
> > >>>>>> One could think about adding another example based on the priority scheduler
> > >>>>>> and giving stream 1 a higher priority. But for illustrating the point of
> > >>>>>> schedulers behaving differently depending on the negotiation of user message
> > >>>>>> interleaving this seems not necessary.
> > >>>>> That second example might indeed be useful.
> > >>>>>
> > >>>>> It's simply a bit strange that in the introduction you speak about an
> > >>>>> "urgent" message (something deserving a higher-priority sending) and
> > >>>>> that another transmission has already started - but then the example has
> > >>>>> none of those two.
> > >>> Exactly.
> > >>> Please add this extra example.
> > >>> This would also clarify from the introduction that when you write ...
> > >>>
> > >>>   The sending of such large messages over
> > >>>   SCTP as specified in [RFC4960] can result in a form of sender side
> > >>>   head of line blocking (e.g., when the transmission of an urgent
> > >>>   message is blocked from transmission because the sender has started
> > >>>   the transmission of another, possibly large, message).
> > >>>
> > >>> .. the notion of transmission urgency is not within a stream but across streams, and that can be achieved with a priority scheduler.
> > >> I would actually prefer to remove the word "urgent" from that sentence. This just rings the wrong bells.
> > > That's a missed opportunity in mind, but your call.
> > Hmm. Let me remove the word "urgent"...
> >
> > Yes, please! I wasn't wild about it, either.
> It's done:
> https://github.com/sctplab/sctp-idata/commit/cf7b331512cd85d6195eb88fa96360673293323f
> >
> > >>> The text might have to be rephrased to explain that this spec provides the ability to create stream scheduler with different relative stream treatments.
> > >> What about:
> > >>
> > >> OLD TEXT
> > >>    This document also defines several stream schedulers for general SCTP
> > >>    associations.  The stream schedulers may behave differently depending
> > >>    on whether user message interleaving has been negotiated for the
> > >>    association or not.
> > >>
> > >> NEW TEXT
> > >>    This document also defines several stream schedulers for general SCTP
> > >>    associations allowing different relative stream treatments.
> > >>    The stream schedulers may behave differently depending
> > >>    on whether user message interleaving has been negotiated for the
> > >>    association or not.
> > >>
> > >> Would that address this issue?
> > > ok.
> > Fixed.
> > >>
> > >>> And also that urgent message would be placed in this high priority stream scheduler (reliable, I guess, but that's orthogonal) stream.
> > >> I really would like to get rid of this. I don't think giving the strict
> > >> priority scheduler as specific role here is a good idea.
> > >>
> > >> The RR scheduler was chosen as the simplest scheduler to illustrate
> > >> that scheduler can behave differently when using interleaving or not.
> > >>>> What about removing the word "urgent" from that sentence.
> > >>>>> In fact I wonder why you speak about urgent things in the introduction
> > >>>>> at all. Neither the scheduler nor the interleaving are designed to help
> > >>>>> urgent things to be sent earlier.
> > >>>>>
> > >>>>> All they do is make the transmission fairer so that *all* chunks get a
> > >>>>> more even serialised distribution on the wire.
> > >>>> That is not true in general. If you use the priority scheduler, one
> > >>>> stream gets preferred treatment. It also makes sure that the time
> > >>>> of a message spent in the stream queue is shorter. So these messages
> > >>>> get sent earlier.
> > >>>> In the case of the round robin scheduler, you are right that this is
> > >>>> about getting fairer treatment. If you take the WFQ scheduler, the
> > >>>> streams don't get the same treatment, but it depends on the weight
> > >>>> given to the stream.
> > >>>>> So, rather than saying that the spec helps urgent things to be sent
> > >>>>> earlier, it should maybe rather state that the spec helps every message
> > >>>>> getting an equal, and equally distributed, share of the medium.
> > >>>> It is not meant that the spec does this. It was one example, what
> > >>>> a particular stream scheduler does. But there are multiple schedulers
> > >>>> defined in the document.
> > >>>>
> > >>>> The RR scheduler example is not intended to illustrate what you can do
> > >>>> with all the schedulers, but that using message interleaving or not
> > >>>> has an impact on the scheduler. That is meant by stating:
> > >>>>
> > >>>> This document also defines several stream schedulers for general SCTP associations.
> > >>>> The stream schedulers may behave differently depending on whether user message
> > >>>> interleaving has been negotiated for the association or not.
> > >>>>
> > >>>> This is followed by illustrating this using the RR scheduler.
> > >>>>
> > >>>> Best regards
> > >>> Also, I don't see the value of the last sentence in this paragraph, especially with RFC2119 MAY.
> > >>>
> > >>>   This section defines several stream schedulers.  The stream
> > >>>   schedulers may behave differently depending on whether user message
> > >>>   interleaving has been negotiated for the association or not.  An
> > >>>   implementation MAY implement any subset of them.
> > >> During the review the question was brought up which scheduler have to be implemented
> > >> by an implementation. This sentence was add to address it.
> > >> If you want, I can remove it.
> > > The point is that the MAY is that sentence doesn't mean anything.
> > > Maybe you mean:
> > >
> > > An implementation MUST implement at least one stream scheduler.
> > That would require an implementation to implement at least one of
> > the ones being listed. You are right in the sense that an implementation
> > MUST implement at least one stream scheduler, but the original text
> > allows that one not to be specified in this document.
> >
> > Are you saying that an implementation MUST implement one of list defined
> > by this document?
> >
> > I'm pretty sure (and this came up in my AD review) the problem is that the original specification DID define a stream scheduler, that's wasn't called a stream scheduler, but SCTP implementations did schedule streams. And that didn't matter until there was more than one way to schedule streams (as in this document).
> The original specification did NOT specify a stream scheduler and left this completely implementation
> specific.
> Right now Linux is using (and only supporting a single scheduler) what we call in this document
> SCTP_SS_FCFS. FreeBSD's default scheduler is what we call in this document SCTP_SS_RR and
> Solaris uses one of these, forgot which one. Among these kernel implementations, only FreeBSD
> supports multiple streams schedulers and its selection. Possibly other closed source SCTP
> implementations might support other schedulers. Don't know.
>
> Thanks for getting that right for me ...
>
> Spencer
>
>  > So you're working out better wording that says something like "in addition to the way the original specification scheduled streams, we're defining more of them". I think.
> I think we wanted to say something like:
>
> Here is a list of schedulers. Your implementation might support some of them, but this is
> left open. So you can't rely on it.
>
> The list of schedulers we specified includes the ones being used as default by FreeBSD, Linux and
> Solaris and was extended by the ones supported by FreeBSD and the one explicitly required by
> WebRTC, which is SCTP_SS_WFQ.
>
> What about this change:
>
> OLD TEXT:
>    An implementation MAY implement any subset of them.
>
> NEW TEXT:
>    An implementation MAY implement any subset of them. If the implementation
>    is used for WebRTC Datachannels as specified in [I-D.ietf-rtcweb-data-channel]
>    it MUST implement the Weighted Fair Queueing Scheduler defined in Section 3.6.
>
> That works for me, but I'll let Benoit reply.

Benoit?

Since I'm not sure which timezone Benoit is in (except that I know it's Yang Central Time), I'll respond by saying that I did talk with Benoit on the call, after sending him this privately:

This is tripping over the same thing I was tripping over - SCTP standards haven't had a stream scheduler, but SCTP implementations scheduled streams. So Michael was thinking that you were asking about forcing existing implementations to implement one of the new ones. 

His response was to propose text basically saying "in addition to what you're doing now, you MAY do one or more of the following". 

I'm sorry I garbled my point so badly.

Benoit said on the call (if I understood him correctly) that my explanation (filtered through your correction) worked for him. The point I was trying to make earlier was that SCTP implementations scheduled streams somehow, and that your document wasn't an effort to tell them to stop doing what they were doing, but that implementers now had more standardized options than they did before.

I think your proposed text says that well. 

At this point, I think you're ready to submit a revised ID to reflect the comments you've received. I'll review your changes, and that will give Benoit a chance to interrupt if he needs to -

no interrupt needed. I'll trust the venerable Spencer.

Regards, Benoit

he's quite responsive, especially when he doesn't spend the evening on a telechat.

Thanks, and it's always a pleasure.

Spencer

 

>
> Spencer
>
> Best regards
> Michael
> >
> > ?
> >
> > Spencer
>
>