[Ecn-sane] rtt-fairness question

[David P. Reed]

Hi Sebastian -
 
Actually, fq in fq_codel does achieve throughput-fairness on the bottleneck link, approximately, given TCP.
And I do agree that throughput fairness is about all you can define locally.
 
That is, no matter what the RTT (unloaded), dropping and ECN marking all flows equally at the bottleneck link will achieve approximate throughput sharing. The end-to-end windows of independent TCP will size themselves to the underlying RTT, as they are wont to do, and as desired if you want to get both good utilization and minimize queueing delay across all paths in the network as a whole. (a reasonable definition of a good operating point).
 
To do this, each router need not know at all what the RTT of the packets flowing through should be. The router strategy is RTT agnostic.
 
My concern was focused on trying to balance RTT among all flows by decisionmaking in a router watching the packets flowing by. That seems like a terrible idea, though I suppose any metric might have some supporters out there. [snide remark: look at all the "diffserv control points", some person actually is a fan of each one, though I doubt anyone knows what each one has an implementation technique that actually will achieve anything like what is stated in the RFC's describing them. It's why I think that diffserv couldn't have resulted from any process like "rough consensus and working code", but instead came from the usual committee-style "standards process" that has produced the millions of useless standards in the world standards organizations].
 
The nice thing about fq_codel and cake, to me, is that they come close to achieving a pragmatic throughput-fairness, eliminate queueing delay on a link - two important factors that allow good end-to-end protocols to be built on top of them. (A key property is reducing the likelihood of load-based starvation of new flows, etc. as long as those flows handle drops and marks by reducing sending rate compatibly with TCP flows). Of course, if implemented badly (like refusing to drop or mark some packets based on some theory like "lost packets are just evil") they may not work well.
 
Thinking about measuring the right things rather than the wrong things, to me, is crucial. Optimizing for 100% link utlilzation is an example of the wrong metric. It should be obvious why, but apparently it is the metric most financial executives in network operators want to see prioritized. With a second metric of "lost packets/sent packets" being optimized to 0. Imagine if roads were required to be 100% utilized by cars at all times... Yup, I've talked to folks at RBOCs in charge of financing (and BT executives, too) who actually respond to that analogy with cars, by saying "data is nothing like cars - you must be kidding" and then going back to saying that they want 100% utilization and 0 dropped packets. That's what accountants do to you.
 
 
 
On Wednesday, April 20, 2022 8:54am, "Sebastian Moeller" <moeller0 at gmx.de> said:



> Hi David,
> 
> 
> > On Apr 19, 2022, at 22:40, David P. Reed <dpreed at deepplum.com> wrote:
> >
> > Sebastian - all your thoughts here seem reasonable.
> >
> > I would point out only two things:
> >
> > 1) 100 ms. is a magic number for human perception. It's basically the order
> of magnitude of humans' ability to respond to unpredictable events outside the
> human.
> 
> Yes, with this I fully agree, "order of magnitude", the actual numerical value of
> 100 is for convenience and has no real significance IMHO. Which I should have
> phrased better. Side-note such experiments typically require the subject to
> create a measurable response, which will take additional time to the initial
> event detection, but that still fits within the 100ms order of magnitude much
> better than a hypothetical 10ms. (for visual events at 10ms the frontal lobe will
> not even have the information available that something changed, vision is
> amazingly slow*)
> 
> > That's why it is magic. Now humans can actually perceive intervals much, much
> shorter (depending on how we pay attention), but usually it is by comparing two
> events' time ordering. We can even synchronize to external, predictable events
> with finer resolution (as in Jazz improv or just good chamber music playing). A
> century of careful scientific research supports this, niot just one experiment.
> 
> Quite a number of experiments however are misinterpreted (or rather interpreted
> without the required nuance) on the internet (yes, I know shocking ;) that the
> internet can be factually imprecise).
> 
> 
> > Which is why one should take it seriously as a useful target. (the fact that
> one can achieve it across the planet with digital signalling networks makes it a
> desirable goal for anything interactive between a human and any entity, be it
> computer or human). If one can do better, of course, that's great. I like that
> from my home computer I can get lots of places in under 8 msec (15 msec RTT).
> >
> > 2) given that a particular heavily utilized link might be shared for paths
> where the light-speed-in-fiber round trip for active flows varies by an order of
> magnitude, why does one try to make fair RTT (as opposed to all other possible
> metrics on each flow) among flows.
> 
> I think the measure that is equalized here is throughput per flow, it is just
> that if done competently this will also alleviate the inherent disadvantage that
> longer RTT flows have compared to shorter RTT flows. But then again, other
> measures are possible as well assuming the bottleneck can get at these easily.
> 
> > It doesn't make any sense to me why. Going back to human interaction times,
> it makes sense to me that you might want to be unfair so that most flows get
> faster than 200 ms. RTT, for example, penalizing those who are really close to
> each other anyway.
> > If the RTT is already low because congestion has been controlled, you can't
> make it lower. Basically, the ideal queue state is < 1 packet in the bottleneck
> outbound queues, no matter what the RTT through that queue is.
> 
> Well, why RTT-fairness? My answer is similar as for why I like FQ, because
> equitable sharing is the one strategy that without information about the flows
> relative importance avoids the pitfall of starving important flows that just
> happen to have a long RTT or a less aggressive controller... So IMHO RTT fairness
> does not need to be absolute but simply good enough to keep all flows at making
> decent forward progress. The very moment someone comes in knowing more about the
> different flows' importance, more optimal capacity sharing becomes possible (like
> in Vint's example)... in a sense neither FQ nor the "accidental" RTT-fairness it
> offers are likely optimal but they are IMHO considerably less likely to be
> pessimal than any uninformed inequitable sharing.
> 
> 
> Regards
> Sebastian
> 
> 
> *) Given that vision is essentially our long-range sense** that internal latency
> typically is not an issue, since events/objects will often be far enough away that
> detection can afford that extra time
> 
> **) In space and time, just look at the stars ;)
> 
> 
> >
> >
> >
> > On Thursday, April 14, 2022 5:25pm, "Sebastian Moeller"
> <moeller0 at gmx.de> said:
> >
> > > Just indulge me here for a few crazy ideas ;)
> > >
> > > > On Apr 14, 2022, at 18:54, David P. Reed
> <dpreed at deepplum.com> wrote:
> > > >
> > > > Am I to assume, then, that routers need not pay any attention to
> RTT to
> > > achieve RTT-fairness?
> > >
> > > Part of RTT-bias seems caused by the simple fact that tight control
> loops work
> > > better than sloppy ones ;)
> > >
> > > There seem to be three ways to try to remedy that to some degree:
> > > 1) the daft one:
> > > define a reference RTT (larger than typically encountered) and have all
> TCPs
> > > respond as if encountering that delay -> until the path RTT exceeds
> that
> > > reference TCP things should be reasonably fair
> > >
> > > 2) the flows communicate with the bottleneck honestly:
> > > if flows would communicate their RTT to the bottleneck the bottleneck
> could
> > > partition its resources such that signaling (mark/drop) and puffer size
> is
> > > bespoke per-flow. In theory that can work, but relies on either the RTT
> > > information being non-gameably linked to the protocol's operation* or
> everybody
> > > being fully veridical and honest
> > > *) think a protocol that will only work if the best estimate of the RTT
> is
> > > communicated between the two sides continuously
> > >
> > > 3) the router being verbose:
> > > If routers communicate the fill-state of their queue (global or per-flow
> does not
> > > matter all that much) flows in theory can do a better job at not putting
> way too
> > > much data in flight remedying the cost of drops/marks that affects high
> RTT flows
> > > more than the shorter ones. (The router has little incentive to lie
> here, if it
> > > wanted to punish a flow it would be easier to simply drop its packets
> and be done
> > > with).
> > >
> > >
> > > IMHO 3, while theoretically the least effective of the three is the only
> one that
> > > has a reasonable chance of being employed... or rather is already
> deployed in the
> > > form of ECN (with mild effects).
> > >
> > > > How does a server or client (at the endpoint) adjust RTT so that it
> is fair?
> > >
> > > See 1) above, but who in their right mind would actually implement
> something like
> > > that (TCP Prague did that, but IMHO never in earnest but just to
> "address" the
> > > L4S bullet point RTT-bias reduction).
> > >
> > > > Now RTT, technically, is just the sum of the instantaneous queue
> lengths in
> > > bytes along the path and the reverse path, plus a fixed wire-level
> delay. And
> > > routers along any path do not have correlated queue sizes.
> > > >
> > > > It seems to me that RTT adjustment requires collective real-time
> cooperation
> > > among all-or-most future users of that path. The path is partially
> shared by many
> > > servers and many users, none of whom directly speak to each other.
> > > >
> > > > And routers have very limited memory compared to their
> throughput-RTdelay
> > > product. So calculating the RTT using spin bits and UIDs for packets
> seems a bit
> > > much to expect all routers to do.
> > >
> > > If posed like this, I guess the better question is, what can/should
> routers be
> > > expected to do here: either equitably share their queues or share queue
> > > inequitably such that throughput is equitable. From a pure router point
> of the
> > > view the first seems "fairest", but as fq_codel and cake show, within
> reason
> > > equitable capacity sharing is possible (so not perfectly and not for
> every
> > > possible RTT spread).
> > >
> > > >
> > > > So, what process measures the cross-interactions among all the
> users of all
> > > the paths, and what control-loop (presumably stable and TCP-compatible)
> actually
> > > converges to RTT fairness IRL.
> > >
> > > Theoretically nothing, in reality on a home link FQ+competent AQM goes a
> long way
> > > in that direction.
> > >
> > >
> > > >
> > > > Today, the basis of congestion control in the Internet is that each
> router is
> > > a controller of all endpoint flows that share a link, and each router is
> free to
> > > do whatever it takes to reduce its queue length to near zero as an
> average on all
> > > timescales larger than about 1/10 of a second (a magic number that is
> directly
> > > derived from measured human brain time resolution).
> > >
> > > The typical applies, be suspicious of too round numbers.... 100ms is in
> no way
> > > magic and also not "correct" it is however a decent description of
> reaction times
> > > in a number of perceptul tasks that can be mis-interpreted as showing
> things like
> > > the brain runs at 10Hz or similar...
> > >
> > >
> > > >
> > > > So, for any two machines separated by less than 1/10 of a
> light-second in
> > > distance, the total queueing delay has to stabilize in about 1/10 of a
> second.
> > > (I'm using a light-second in a fiber medium, not free-space, as the
> speed of light
> > > in fiber is a lot slower than the speed of light on microwaves, as Wall
> Street has
> > > recently started recoginizing and investing in).
> > > >
> > > > I don't see how RTT-fairness can be achieved by some set of bits in
> the IP
> > > header. You can't shorten RTT below about 2/10 of a second in that
> desired system
> > > state. You can only "lengthen" RTT by delaying packets in source or
> endpoint
> > > buffers, because it's unreasonable to manage all the routers.
> > > >
> > > > And the endpoints that share a path can't talk to each other and
> reach a
> > > decision in on the order of 2/10 of a second.
> > > >
> > > > So at the very highest level, what is RTT-fairness's objective
> function
> > > optimizing, and how can it work?
> > > >
> > > > Can it be done without any change to routers?
> > >
> > > Well the goal here seems to undo the RTT-dependence of throughput so a
> router can
> > > equalize per flow throughput and thereby (from its own vantage point)
> enforce RTT
> > > independence, within the amount of memory available. And that already
> works today
> > > for all identifiable flows, but apparently at a computational cost that
> larger
> > > routers do not want to pay. But you knew all that
> > >
> > >
> > > >
> > > >
> > > >
> > > >
> > > > On Tuesday, April 12, 2022 3:07pm, "Michael Welzl"
> <michawe at ifi.uio.no>
> > > said:
> > > >
> > > >
> > > >
> > > > On Apr 12, 2022, at 8:52 PM, Sebastian Moeller
> <moeller0 at gmx.de>
> > > wrote:
> > > > Question: is QUIC actually using the spin bit as an essential part
> of the
> > > protocol?
> > > > The spec says it’s optional:
> > > https://www.rfc-editor.org/rfc/rfc9000.html#name-latency-spin-bit
> > > > Otherwise endpoints might just game this if faking their RTT at a
> router
> > > yields an advantage...
> > > > This was certainly discussed in the QUIC WG. Probably perceived as
> an unclear
> > > incentive, but I didn’t really follow this.
> > > > Cheers,
> > > > Michael
> > > >
> > > > This is why pping's use of tcp timestamps is elegant, little
> incentive for
> > > the endpoints to fudge....
> > > >
> > > > Regards
> > > > Sebastian
> > > >
> > > >
> > > > On 12 April 2022 18:00:15 CEST, Michael Welzl
> <michawe at ifi.uio.no>
> > > wrote:
> > > > Hi,
> > > > Who or what are you objecting against? At least nothing that I
> described
> > > does what you suggest.
> > > > BTW, just as a side point, for QUIC, routers can know the RTT today
> - using
> > > the spin bit, which was designed for that specific purpose.
> > > > Cheers,
> > > > Michael
> > > >
> > > >
> > > > On Apr 12, 2022, at 5:51 PM, David P. Reed
> <dpreed at deepplum.com>
> > > wrote:
> > > > I strongly object to congestion control *in the network* attempting
> to
> > > measure RTT (which is an end-to-end comparative metric). Unless the
> current RTT is
> > > passed in each packet a router cannot enforce fairness. Period.
> > > >
> > > > Today, by packet drops and fair marking, information is passed to
> the sending
> > > nodes (eventually) about congestion. But the router can't know RTT
> today.
> > > >
> > > > The result of *requiring* RTT fairness would be to put the random
> bottleneck
> > > router (chosen because it is the slowest forwarder on a contended path)
> become the
> > > endpoint controller.
> > > >
> > > > That's the opposite of an "end-to-end resource sharing protocol".
> > > >
> > > > Now, I'm not saying it is impossible - what I'm saying it is asking
> all
> > > endpoints to register with an "Internet-wide" RTT real-time tracking and
> control
> > > service.
> > > >
> > > > This would be the technical equivalent of an ITU central control
> point.
> > > >
> > > > So, either someone will invent something I cannot imagine (a
> distributed,
> > > rapid-convergence algortithm that rellects to *every potential user* of
> a shared
> > > router along the current path the RTT's of ALL other users (and
> potential users).
> > > >
> > > > IMHO, the wish for RTT fairness is like saying that the entire
> solar system's
> > > gravitational pull should be equalized so that all planets and asteroids
> have fair
> > > access to 1G gravity.
> > > >
> > > >
> > > > On Friday, April 8, 2022 2:03pm, "Michael Welzl"
> <michawe at ifi.uio.no>
> > > said:
> > > >
> > > > Hi,
> > > > FWIW, we have done some analysis of fairness and convergence of
> DCTCP in:
> > > > Peyman Teymoori, David Hayes, Michael Welzl, Stein Gjessing:
> "Estimating an
> > > Additive Path Cost with Explicit Congestion Notification", IEEE
> Transactions on
> > > Control of Network Systems, 8(2), pp. 859-871, June 2021. DOI
> > > 10.1109/TCNS.2021.3053179
> > > > Technical report (longer version):
> > > >
> > >
> https://folk.universitetetioslo.no/michawe/research/publications/NUM-ECN_report_2019.pdf
> > > > and there’s also some in this paper, which first introduced
> our LGC
> > > mechanism:
> > > > https://ieeexplore.ieee.org/document/7796757
> > > > See the technical report on page 9, section D: a simple trick can
> improve
> > > DCTCP’s fairness (if that’s really the mechanism to stay
> with…
> > > I’m getting quite happy with the results we get with our LGC
> scheme :-)
> > > )
> > > >
> > > > Cheers,
> > > > Michael
> > > >
> > > > On Apr 8, 2022, at 6:33 PM, Dave Taht <dave.taht at gmail.com>
> wrote:
> > > > I have managed to drop most of my state regarding the state of
> various
> > > > dctcp-like solutions. At one level it's good to have not been
> keeping
> > > > up, washing my brain clean, as it were. For some reason or another
> I
> > > > went back to the original paper last week, and have been pounding
> > > > through this one again:
> > > >
> > > > Analysis of DCTCP: Stability, Convergence, and Fairness
> > > >
> > > > "Instead, we propose subtracting α/2 from the window size for
> each
> > > marked ACK,
> > > > resulting in the following simple window update equation:
> > > >
> > > > One result of which I was most proud recently was of demonstrating
> > > > perfect rtt fairness in a range of 20ms to 260ms with fq_codel
> > > > https://forum.mikrotik.com/viewtopic.php?t=179307 )- and I'm
> pretty
> > > > interested in 2-260ms, but haven't got around to it.
> > > >
> > > > Now, one early result from the sce vs l4s testing I recall was
> severe
> > > > latecomer convergence problems - something like 40s to come into
> flow
> > > > balance - but I can't remember what presentation, paper, or rtt
> that
> > > > was from. ?
> > > >
> > > > Another one has been various claims towards some level of rtt
> > > > unfairness being ok, but not the actual ratio, nor (going up to
> the
> > > > paper's proposal above) whether that method had been tried.
> > > >
> > > > My opinion has long been that any form of marking should look more
> > > > closely at the observed RTT than any fixed rate reduction method,
> and
> > > > compensate the paced rate to suit. But that's presently just
> reduced
> > > > to an opinion, not having kept up with progress on prague,
> dctcp-sce,
> > > > or bbrv2. As one example of ignorance, are 2 packets still paced
> back
> > > > to back? DRR++ + early marking seems to lead to one packet being
> > > > consistently unmarked and the other marked.
> > > >
> > > > --
> > > > I tried to build a better future, a few times:
> > > > https://wayforward.archive.org/?site=https%3A%2F%2Fwww.icei.org
> > > >
> > > > Dave Täht CEO, TekLibre, LLC
> > > > _______________________________________________
> > > > Ecn-sane mailing list
> > > > Ecn-sane at lists.bufferbloat.net
> > > > https://lists.bufferbloat.net/listinfo/ecn-sane
> > > >
> > > > --
> > > > Sent from my Android device with K-9 Mail. Please excuse my
> brevity.
> > > >
> > >
> > >
> 
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/ecn-sane/attachments/20220420/ee63f4a3/attachment-0001.html>