Re: [Bloat] less than best effort: TCP - flexis - A New Approach To Incipient Congestion Detection and Control

[Qian Li <biz.tinalee@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 5052 bytes --]

Hi Dave,

I was just told that I am allowed to distribute the code freely. I will
upload it to GitHub and will send you the link as soon as I am done with it.

As for the AQM test, I set the QDisc to FQ-CoDel, CoDel, and CAKE, but none
worked on CORE. In contrast, RED and PIE worked as expected. As far as I
know, the major difference between these two groups of AQMs is the time
when the packets are dropped. But I am not 100% sure it was the cause. I
may somehow test FlexiS or FlexiR (FlexiS adapted to the receiver side) on
a testbed with various AQMs. But it will be toward the end of the
adaptation I guess :)

Best regards,
Qian

On Thu, Apr 7, 2022 at 8:52 AM Dave Taht <dave.taht@gmail.com> wrote:

> On Tue, Apr 5, 2022 at 8:44 AM Qian Li <biz.tinalee@gmail.com> wrote:
> >
> > Dear Dave,
> >
> > Thank you for your interest in my work.
> >
> > I have read another paper authored by D. Rossi at el. presenting the
> priority inversion problem of LEDBAT when it is used together with AQM. And
> it has become one of the factors that motivated me to devise a new LBE CC
> that can preserve low priority even when AQM is used.
>
> We'd given up hope circa 2014 as of the publication of the paper I
> cited, and moved on.
>
> >However, I could not test FlexiS with CoDel on the CORE emulator probably
> because CoDel drops packets at the dequeue time.
>
> I don't really understand that statement.
>
> > More tests should be done to verify that FlexiS does preserve low
> priority in the presence of various AQM algorithms.
>
> Yes. until fairly recently I had had a testbed setup that allowed
> testing of various tcps and aqm systems, but its been in storage since
> covid.
>
> > I am now adapting FlexiS to the receiver side. The main motivation to do
> so is that there might be HTTP/TCP proxies between the sender and the
> receiver. A receiver side LBE CC and make the connection between the proxy
> and the receiver LBE. In this work, I am going to tackle some open issues
> with FlexiS. For example, I am going to test if trend analysis can be done
> based on one way delay so that the throughput is less affected by ack path
> congestion. And I am going to evaluate various techniques to reduce rate
> below 2 mss per RTT. This may include what you have suggested -- use small
> packets and sub-packet window. I am also interested in using pacing to slow
> down sending rate and maybe more alternative solutions.
>
> Cool!
>
> >
> > I don't have a git tree for the source code mainly because I don't know
> if I am allowed to publish the code as open source. If you are interested
> in the source code, I can ask the University of Oslo if I am allowed to
> distribute it freely?
>
> I would hope they would allow publication. The world is full of half
> baked projects that if only someones new also stepped in, were
> completed. An example of this is BBR which originally was about half
> what it is today, until source was released among the right people.
>
> >
> > Best regards,
> > Qian
> >
> >
> >
> > On Sun, Apr 3, 2022 at 6:38 AM Dave Taht <dave.taht@gmail.com> wrote:
> >>
> >> Dear Qian:
> >>
> >> Pretty promising paper. I liked that it tackled congestion on the ack
> >> path, among other things.
> >>
> >>
> https://www.techrxiv.org/articles/preprint/TCP_FlexiS_A_New_Approach_To_Incipient_Congestion_Detection_and_Control/19077161/1/files/33905018.pdf
> >>
> >> I like also that you tackled, inter-rtt fairness, and, ledbat's
> >> latecomer advantage problem, and in fig 9, the basic problem with
> >> delay based LBE vs AQMs (in that ledbat degrades to reno)... [1]
> >>
> >> Towards your conclusion...
> >>
> >> I have always disagreed with the "don't reduce segment size" crowd,
> >> btw. If you have a rate where you need to go below 2mss, it doesn't
> >> hurt the network to reduce the size of the packet, and you can keep
> >> the signal strength up by reducing that size and continuing to sample
> >> rtt, to respond quickly.
> >>
> >> Even if you are only passing a single byte of data, by lowering this
> >> below everyone else's 2mss noise floor, you still eventually win, and
> >> also you occupy space in packet fifos, reducing overall latency, as
> >> bytes=time. IMHO.
> >>
> >> elsewhere, sub-packet windows are being experimented in bbrv2, I'm
> >> told, but not in LBE.
> >>
> >> I'm also a big believer in packet pacing, and I think this is the
> >> first paper I've seen that attempted LBE with it. Thx!
> >>
> >> Got a git tree?
> >>
> >> [1] do wish you'd had cited
> >> https://perso.telecom-paristech.fr/drossi/paper/rossi14comnet-b.pdf
> >>
> >> --
> >> 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
>
>
>
> --
> 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
>

[-- Attachment #2: Type: text/html, Size: 6547 bytes --]