Re: [tcpm] [tsvwg] Cross-area alignment on "L4S and RACK"

["Scheffenegger, Richard" <[e-mail]>]

Hi,

Just to chime in here, for a (unfortunately common) real world example 
that I tend to run into more frequently recently:

Some implementations of (L2 or L3) ECMP appear to perform regular / 
load-triggered rebalancing events. I have observed these in particular 
with certain VSwitch implementations, as well as some modern, 
proprietary L2 "trill-like" switch topologies.

While there are adaptive experimental mechanisms in TCP to deal with 
relatively consistent reordering (measured in packets), the timescales 
involved in my examples above are such, that these adaptive variants 
really can not cope properly.

And the argument, to measure the reordering extent in units of time 
rather than packets seems sensible to me, as it bounds the maximum delay 
incurred before a retransmission happens.

In the above mentioned examples, doing rebalance events only when the 
relative buffer utilization allows for a bounded (in time) reorder 
extent should be relatively easy to work out. Compared to the status quo 
(where the potential worst case reordering extent is not considered), 
this seems to be an improvement.

Best regards,
   Richard


Am 09.11.2018 um 20:28 schrieb Scharf, Michael:
> First, I don’t know if the currently discussed parameters for the RACK 
> reordering window are indeed future proof. Maybe TCP stacks could use 
> different parameters in one or two years from now? So what is the 
> real-world benefit for a link designer **now** in relying on an 
> parameters that can change easily? And I would actually assume that the 
> exact parameter values in a (hypothetical) future standards track 
> specification of RACK would be a SHOULD. Different TCP stacks may pick 
> different values. How useful is this then to a link designer?
> 
> Second, I don’t understand why one cannot try to provide some other 
> guidance, say, in terms of number of triggered DUPACKs or the like. 
> Let’s assume we said a stack must be able to deal with a reorder degree 
> of 10? For which link technologies would this not be good enough?
> 
> Third, some real-world data would facilitate this discussion quite a 
> bit, IMHO. I have a hard time to follow this proposal without examples 
> for different RTTs, different number of parallel connections, different 
> mixtures of traffic, etc.
> 
> Michael
> 
> *From:*Bob Briscoe [mailto:[e-mail]]
> *Sent:* Friday, November 9, 2018 7:52 PM
> *To:* Scharf, Michael; [e-mail]
> *Cc:* [e-mail]
> *Subject:* Re: [tsvwg] Cross-area alignment on "L4S and RACK"
> 
> Michael,
> 
> The idea is that (as now in the RACK draft since discussion on the tcpm 
> list earlier this year under a subject something like vicious circle or 
> virtuous circle) there will be a limit on the initial RACK reordering 
> window expressed in fractional RTT units.
> 
> Link vendors know the min e2e RTT of the paths that their links are 
> usually deployed over. Therefore they can estimate the max reordering 
> they allow their links to introduce.
> 
> I don't expect this limit to be set in stone while both RACK and L4S are 
> experimental. But as both these experiments progress, I think the 
> industry will get a better idea of where the limit should lie.
> 
> It was the realization that, if TCP adapted its reordering window to the 
> measured reordering degree without bound, the L2 world would continually 
> increase the reordering degree of their links (the vicious circle).
> 
> If there is any aspect of this limit expressed in units of packets, the 
> limit will decrease in time as flow rates increase, and therefore not 
> serve to give the L2 world any clear guidance - no de facto limit agreed 
> between the L4 world and the L2 world.
> 
> If we keep the limit in units of RTT, hopefully then we have a virtuous 
> circle, not a vicious circle.
> 
> 
> 
> Bob
> 
> On 09/11/2018 18:28, Scharf, Michael wrote:
> 
>     I may miss something, but I read the outcome of the appendix as a
>     requirement “must be more robust to reordering”. As far as I
>     understand, that can be implemented in different ways.
> 
>     And I may be wrong, but those link technologies could be well
>     implemented in a router and then one needs to analyze how that whole
>     architecture fits into the scheduler, queue, policer, and line card
>     hardware design of a router. But that is something routing people
>     know better than me. My point is just to get the experts into the
>     loop early.
> 
>     Michael
> 
>     *From:*Bob Briscoe [mailto:[e-mail]]
>     *Sent:* Friday, November 9, 2018 7:04 PM
>     *To:* Scharf, Michael; [e-mail] <mailto:[e-mail]>
>     *Cc:* [e-mail] <mailto:[e-mail]>
>     *Subject:* Re: [tsvwg] Cross-area alignment on "L4S and RACK"
> 
>     Michael,
> 
>     Please address the scaling rationale in the appendix for why it does
>     matter how a sender realizes this internally.
> 
>     The list of access links types I mentioned are rarely connected
>     directly to routers.
> 
> 
> 
>     Bob
> 
>     On 09/11/2018 17:44, Scharf, Michael wrote:
> 
>         I’ll not comment on link technology implementation, scheduler
>         and queue design assumptions in Appendix A.1.7 of
>         draft-ietf-tsvwg-ecn-l4s-id-05. People e.g. in RTG area may be
>         more familiar with the implications of internal designs e.g.
>         inside a router. IMHO the best way to get routing experts into
>         the loop before further discussing what is currently listed in
>         Appendix A.1.7.
> 
>         Thus, I limit my response to the actual wording in the main part
>         of draft-ietf-tsvwg-ecn-l4s-id-05:
> 
>              o  A scalable congestion control MUST detect loss by
>         counting in
> 
>                units of time, which is scalable, and MUST NOT count in
>         units of
> 
>                packets (as in the 3 DupACK rule of traditional TCP),
>         which is not
> 
>                scalable (see Appendix A.1.7
>         <https://tools.ietf.org/html/draft-ietf-tsvwg-ecn-l4s-id-05#appendix-A.1.7>for
>         rationale).
> 
>         As individual contributor, I disagree with this requirement. I
>         would probably be OK with a wording that a TCP sender must be
>         able to tolerate reordering beyond 3 Duplicate ACKs. I believe
>         it does not matter how a sender realizes this internally.
> 
>         And if that requirement is not just removed, I believe the
>         wording “traditional TCP” has to be replaced by “standard TCP”.
> 
>         BTW, also, in the rest of the document there has to be a clear
>         distinction between experiments and standards. History tells us
>         that some experiments can evolve into standards track
>         specifications, while other experiments may fail. And there is
>         no way of knowing that in advance.
> 
>         Michael
> 
>         *From:*Bob Briscoe [mailto:[e-mail]]
>         *Sent:* Friday, November 9, 2018 11:08 AM
>         *To:* Scharf, Michael; [e-mail] <mailto:[e-mail]>
>         *Cc:* [e-mail] <mailto:[e-mail]>
>         *Subject:* Re: [tsvwg] Cross-area alignment on "L4S and RACK"
> 
>         Michael,
> 
>         On 09/11/2018 06:21, Scharf, Michael wrote:
> 
>             I don’t see the fundamental difference between a link
>             technology that guarantees in-order delivery within a flow
>             (say by flow hashing, see ECMP) vs. some link scheme that
>             guarantees in-order delivery, say, only for some fraction of
>             the traffic characterized by some other means (say, other
>             header bits).
> 
>         [The "other means" and "other bits" wording is pretty ambiguous
>         here. I reckon I can guess what you mean but then I don't know
>         why you're saying this, 'cos I don't think I've disagreed with
>         you on this. So perhaps you could be more specific and we might
>         be able to see why you've brought this up.]
> 
>         Also I wasn't really talking about all "link technology that
>         guarantees in-order delivery within a flow". {Note 1} I was
>         talking about link technology where all flows are blocked while
>         waiting to get the aggregate back in order that it was in when
>         it arrived. I specifically said:
> 
> 
> 
>         Note that there can be multiple flows over one link. So
>         'in-order delivery' for a link does not mean in order of TCP
>         sequence number. It means 'delivered in the order that packets
>         arrived at the link' (the link ingress adds its own sequence
>         numbers, and the link egress buffers them until it can send out
>         in the same order, or until a time-out or max no. of link
>         retransmissions).
> 
>         For example:
> 
>           * A packet resequencing buffer after allowing for link-layer
>             retransmissions is common for radio link technologies such
>             as LTE (PDCP layer) and 802.11 WiFi.
>           * A packet resequencing buffer after merging multiple bonded
>             link channels is common for other access link technologies
>             such as downstream DOCSIS bonded channels. 
> 
> 
> 
> 
> 
>         Clearly, reordering is a problem that needs to be discussed in
>         the IETF as a whole. And we could discuss whether a revision of
>         RFC 3366 is needed.
> 
>         No. There is no proposal to alter general guidance for all links
>         like RFC 3366. No-one is saying this.
> 
>         There is not even a proposal to require or even recommend that
>         L4S-specific links do anything.
> 
>         The specific proposal is to require L4S /sources/ to use a
>         RACK-like scheme, as a condition for setting the ECT(1)
>         codepoint. See the last bullet in S.4.3 of
>         draft-ietf-tsvwg-ecn-l4s-id-05
>         <https://tools.ietf.org/html/draft-ietf-tsvwg-ecn-l4s-id-05#section-4.3>.
> 
> 
>         This says nothing about what L4S links MUST, SHOULD or MAY do.
>         It just gives an L4S link an opportunity not to have to do
>         resequencing.
> 
> 
> 
> 
>         But I am not convinced that this topic has to be tied into an
>         experiment such as L4S.
> 
>         It is an opportunity that currently only applies to L4S.
> 
>         L4S is merely an example of a case where we can mandate that
>         there will be no non-RACK traffic in a link, which creates an
>         opportunity for those who are implementing the L4S DualQ Coupled
>         AQM to ignore the RFC3366 resequencing advice for the L4S queue
>         (but we are not changing the advice, because L4S is only an
>         experiment). The prize is to remove the head-of-line blocking
>         problem for their users.
> 
>         This is not L4S-specific, altho there are no other examples at
>         the moment where this opportunity is possible (so in that sense
>         it is /currently/ L4S-specific).
> 
> 
>         The message has to be handled really carefully, because the
>         opportunities for misunderstanding are enormous (as evidenced by
>         this thread, and we haven't even started talking to
>         non-transport people yet).
> 
>         So Gorry and David B want me to draft a short I-D that they will
>         word-smith and author as tsvwg chairs to make the situation clear.
> 
>         I have said I don't want this discussion to be officially raised
>         outside the transport area until we are sure that a variant of
>         RACK is even feasible without any packet counting at all (i.e.
>         without the initial 3-DupACK rule in current RACK). So I'll
>         prioritize working that out for the next week or so, rather than
>         drafting anything.
> 
> 
> 
>         Bob
> 
>         {Note 1}: I hadn't considered that this would affect
>         technologies like ECMP that maintain order without a
>         resequencing buffer and without introducing HoL blocking. But
>         actually, the consequence there would be that L4S traffic could
>         be sprayed by ECMP without per-flow hashing.
> 
> 
> 
> 
>         Michael
> 
>         *From:*Bob Briscoe [mailto:[e-mail]]
>         *Sent:* Friday, November 9, 2018 6:42 AM
>         *To:* Scharf, Michael; [e-mail] <mailto:[e-mail]>
>         *Cc:* [e-mail] <mailto:[e-mail]>
>         *Subject:* Re: [tsvwg] Cross-area alignment on "L4S and RACK"
> 
>         Michael,
> 
>         That is the point I thought you were making, and my point /is/
>         about that.
> 
>         If an application needs all the packets (i.e. a reliable
>         protocol), it's not useful to reduce the latency of some packets
>         but not all. That's where the in-order requirement comes from.
> 
>         However, if a stream-based transport is used (i.e. TCP) the
>         /application/ still gets in-order delivery from TCP. It just
>         doesn't have to require links to provide in order delivery as well.
> 
>         Note that there can be multiple flows over one link. So
>         'in-order delivery' for a link does not mean in order of TCP
>         sequence number. It means 'delivered in the order that packets
>         arrived at the link' (the link ingress adds its own sequence
>         numbers, and the link egress buffers them until it can send out
>         in the same order, or until a time-out or max no. of link
>         retransmissions).
> 
>         So, if the link loses a packet from flow A (the link isn't aware
>         of microflows, I'm just saying the packet happens to be from
>         flow A), link-layer in-order delivery will hold back all packets
>         sent after that packet (from all flows, not just flow A) until
>         the link repairs the gap or gives up.
> 
>         So hop-by-hop in-order delivery gives every flow the same delay
>         as the worst delayed flow.
> 
>         Of course, someone building a network for DETNET would avoid
>         link technologies like WiFi or LTE that frequently lose and
>         retransmit packets. But wherever there is variable delay in a
>         link, guaranteed in-order delivery per-hop means every flow is
>         guaranteed to always get the worst delay from every link, which
>         would have been experienced by only one flow without per-hop
>         in-order delivery.
> 
>         And many of the applications of DETNET involve dumb industrial
>         machinery that just expects everything to arrive in order and to
>         a clocked schedule. But it's always as effective to deploy a
>         resequencing buffer as the penultimate hop, screwed onto the
>         receiving machine if necessary, rather than require per-hop
>         resequencing in the network.
> 
>         ________________
>         So, why do Internet links often ensure in-order delivery even
>         though TCP puts everything in order for the application?
> 
>         Short answer: the 3 DupACK rule (which is to do with
>         loss-recovery, a different issue from the discussion above).
> 
>         Long answer: As flow rates have scaled up, the typical time
>         between 3 packet arrivals has become so small that TCP's
>         3-DupACK rule makes links have to provide delivery with only a
>         very small re-ordering degree - so small that it's easier for a
>         link to just deliver everything in the order it received it; not
>         because TCP doesn't put packets back in order, but just so as
>         not to trigger TCP into generating spurious retransmissions.
> 
> 
> 
>         Bob
> 
>         On 08/11/2018 16:35, Scharf, Michael wrote:
> 
>             Inline [ms]
> 
>             *From:*Bob Briscoe [mailto:[e-mail]]
>             *Sent:* Thursday, November 8, 2018 12:56 PM
>             *To:* Scharf, Michael; [e-mail] <mailto:[e-mail]>
>             *Cc:* [e-mail] <mailto:[e-mail]>
>             *Subject:* Re: [tsvwg] Cross-area alignment on "L4S and RACK"
> 
>             Michael,
> 
>             This is merely a symptom of a difference in opinion on where
>             the resequencing function should be located.
> 
>               * L4S takes the end-to-end approach saying that it is
>                 sufficient to do resequencing in one place (the
>                 receiving host) if it is needed. Then any resequencing
>                 delay only affects that flow.
>               * The DETNET approach is saying the resequencing must be
>                 done hop-by-hop. This means there appears to be no
>                 resequencing needed on the receiver (except if there's
>                 re-ordering on the final hop). 
> 
>               * In order to guarantee no resequencing in the network
>                 (DETNET approach), only the the worst case latency can
>                 be guaranteed, and every packet has to have that same
>                 worst-case latency.
>               * When you leave resequencing to the receiver (end-to-end
>                 approach), most of the packets arrive earlier than they
>                 would with DETNET, but out of order ones might not. Then
>                 the application chooses whether it wants to wait.
> 
>             [ms] My point is different: There seem to be other working
>             groups in the IETF that do talk about “low latency” as well
>             but apparently they also do need in-order delivery of
>             packets, too. So the assumption that e.g. a link layer
>             technology can simply disable in-order delivery for all “low
>             latency” applications seems not correct. The reality may be
>             a bit more complex.
> 
>             The 3 DupACK rule in TCP led the Internet not to follow the
>             end-to-end principle on re-sequencing. Now we're removing
>             the 3 DupACK rule, we can take advantage of the e2e principle.
> 
>             [ms] For TCP, three DupACKS are a SHOULD in RFC 5681 and RFC
>             5681 is standards track. So the bar for “removing” that rule
>             for TCP is not low…
> 
>             [ms] Also, the end-to-end principle comes with tradeoffs.
>             For instance, relying on the end-to-end principle for
>             recovery of bit errors is not efficient. For in-order
>             delivery there are tradeoffs as well, see e.g. Section 4 in
>             draft-ietf-tcpm-rack-04. There may be no free lunch.
> 
>             Of course, an application might choose to use TCP and L4S,
>             rather than UDP and L4S (e.g QUIC). Then there could still
>             be HoL blocking in the receiving TCP stack. But at least
>             there's no HoL blocking in the network, and the app can
>             choose between stream (TCP) or datagram (UDP).
> 
>             [ms] If LRO/GRO in the receiving TCP stack gets messed up, I
>             fail to see the benefit (see draft-ietf-tcpm-rack-04).
> 
>             Michael
> 
> 
>             Bob
> 
>             On 06/11/2018 19:20, Scharf, Michael wrote:
> 
>                 A comment on the TCPM presentation "L4S and RACK":
> 
>                   
> 
>                 As far as I understand, the DETNET WG in RTG area has quite some uses cases for ultra-low latency transport - in particular also with bounded jitter. And some of these applications (e.g. using UDP) apparently may not be able to tolerate *any* out-of-order packet delivery.
> 
>                   
> 
>                 So perhaps some cross-area alignment on ulta-low-latency application requirements would be useful?
> 
>                   
> 
>                 Michael
> 
>                 (who recently had to review draft-ietf-detnet-architecture)
> 
>                   
> 
>                   
> 
> 
> 
> 
> 
> 
> 
>             -- 
> 
>             ________________________________________________________________
> 
>             Bob Briscoehttp://bobbriscoe.net/
> 
> 
> 
> 
> 
> 
>         -- 
> 
>         ________________________________________________________________
> 
>         Bob Briscoehttp://bobbriscoe.net/
> 
> 
> 
> 
> 
>         -- 
> 
>         ________________________________________________________________
> 
>         Bob Briscoehttp://bobbriscoe.net/
> 
> 
> 
> 
>     -- 
> 
>     ________________________________________________________________
> 
>     Bob Briscoehttp://bobbriscoe.net/
> 
> 
> 
> -- 
> 
> ________________________________________________________________
> 
> Bob Briscoehttp://bobbriscoe.net/
>