Re: [tsvwg] Related to "Non-L4S traffic abusing the L-queue" discussion during the interim

[Bob Briscoe <[e-mail]>]

Jonathan, Rod, Pete,

Chia-Yu has now run the "Fast-lane" or "Throughput Bonus" experiment 
brought up during L4S WG last call - as closely as possible to the 
rather high level description in your objection.

Summary:

  * No evidence for the 'fast-lane' claim, by reproducing the experiment
  * The result can, however, be reproduced very closely by suppressing
    congestion control as well
      o But objection stated "bonus is easily exploited ... _without_
        disabling congestion control"
  * Once congestion control is disabled, the same result can be
    reproduced in either queue.
      o This proves the bonus in the objection was due to
        straightforward unresponsiveness, not the DualQ
      o So, the claimed 'fast lane' is no faster than the other lane
      o The experiment in the objection was likely faulty, and somehow
        suppressed congestion control
  * Dual Queue Coupled AQM meets its stated goal of not allowing
    unresponsive flows to cause more harm to existing traffic than in a
    single queue
      o unresponsive thru'put: same in either queue
      o unresponsive delay: lower in L than C queue, but not at the
        expense of anyone else's delay

The full Disproof of the experiment in the objection, with detailed 
experiment set-up and results are given here, which also includes links 
to your original experiment report:
https://github.com/L4STeam/l4steam.github.io/blob/master/disproof-fast-lane/202205-ThruputBonusDisproof.pdf

This is now the fifth request (starting 17 Feb 22) asking Jonathan to 
either defend or retract the experiment in the objection, which seems to 
have been faulty.


Bob


On 22/04/2022 22:03, Bob Briscoe wrote:
> Jonathan,
>
> This concerns the experiment in your report last August'21, where you 
> claimed that there was an L4S 'throughput bonus'. I asked you to 
> either do an A/B experiment or explain your experiment setup, so that 
> we can check why we cannot reproduce any bonus with either:
> a) the experiments that Koen did with an unresponsive L4S flow.
> b) or the many A/B experiments done with two responsive flows, one 
> paced ECT(1) vs. one CUBIC ECT(0).
>
> This is the fourth request (starting 17 Feb 22). Please refer to the 
> thread with my other three reminders here:
> https://mailarchive.ietf.org/arch/msg/tsvwg/Bthg3orBBE09Q45HAREXZgoLgcM/
>
> in case your mailer is ignoring or junking that thread, I'm sending 
> this email on another thread about this experiment that was running in 
> parallel.
>
>
> Bob
>

-- 
________________________________________________________________
Bob Briscoehttp://bobbriscoe.net/