On Mon, Jun 18, 2018 at 12:45 PM, <***@csail.mit.edu> wrote:
> Thanks, Dave, for the comments. The paper does have results comparing to Cubic-over-sfqCodel in a couple of sections. For example, "In an experiment to test RTT-fairness, Copa, Cubic, Cubic over CoDel and Newreno obtained Jain fairness indices of 0.76, 0.12, 0.57 and 0.37 respectively (higher the better).” Figure 6 in the paper illustrates.
It is not clear from the paper it was actually sfqcodel under test at
any point. The codel results look like codel results.
My largely empirical analysis is that delay based tcp's work great
against fq_codel derived aqms, even competing against cubic or bbr. In
the case of copa
I would expect the loss based portion of the algo to rarely engage,
and the relative smoothness-es of copa for rate changes to work quite
well against the smoother rate sharing flow-queuing generally
Not clear if you were using tcp pacing, either.
Got patches? Test scripts?
> Most of the other results were obtained over real-world paths, or via trace-driven emulation where the traces were obtained over real paths. In those cases we are not sure how to compare with AQM.
Well, I would certainly love it if one day y'all tore apart
against whatever you are dreaming up. :)
>> On Jun 18, 2018, at 3:09 PM, Dave Taht <***@gmail.com> wrote:
>> Looks promising, like all papers do...
>> Keep hoping folk will try their CC-scheme de-jure against one of the existing
>> aqm designs....
>> Dave Täht
>> CEO, TekLibre, LLC
>> Tel: 1-669-226-2619
CEO, TekLibre, LLC