[Battlemesh] blowing up the network tests?
dave.taht at gmail.com
Wed Jul 3 21:28:57 CEST 2019
On Wed, Jul 3, 2019 at 11:58 AM Filipe Borges Teixeira
<filipebteixeira at gmail.com> wrote:
> Sounds interesting; I'd suggest that we discuss its feasibility (or any other test/testbed plan) during the first days of WBM.
One fun thing we did at battlemesh v8 that is worth re-viewing is:
https://www.youtube.com/watch?v=Rb-UnHDw02o - especially the demo at
25:20 - which lays out the best explanation of
the wifi architectural rework we needed for per-station queuing I'd
ever made, which was subsequently completed a year later.
Regrettably we didn't manage to accomplish all that talk laid out -
minstrel-blues is still not "there" to my knowledge, we didn't crush
retries or multicast as much as we should, and so on, but we did knock
it out of the park on the fq_codel per-sta queue rework which
was mainlined, backported to openwrt, and written up in "ending the anomaly".
Google's separate implementation (for google chromebooks and google
wifi) has a great set of of RvR vs latency charts here:
skip to the end and work back. A complaint that I had then was that
more p2mp testing needed to happen....
Anyway, now that we have vastly shorter and fairer queues on wifi,
there are new problems to see.
1) QoE for web traffic on a sparse client
2) QoE for movies... netflix has some good code here...
3) ECN is enabled universally. Given the current carnage in the IETF
over L4S vs SCE and ECN responses in general, I'm increasingly down on
it. A test, flooding a link with, say 50 TCP flows with ECN enabled on
the client and server might do bad things to routing protocols that
lack ECN capability, and you'll see meshy routes moving around more
due to congestion. In stepping down to low bandwidths (20mbit) under
heavy load, leaf babel nodes would consistently drop off the network
under a load like that with a 1000 routes....
> Besides SAn, guifipedro and me, who else can bring some hardware?
I wish I could come.
> Filipe Teixeira
> Dave Taht <dave.taht at gmail.com> escreveu no dia domingo, 30/06/2019 à(s) 23:45:
>> I would love to see how much better we do on these tests now that all
>> the fq_codel stuff has landed.
>> (not coming this year. too broke. :( )
>> Dave Täht
>> CTO, TekLibre, LLC
>> Tel: 1-831-205-9740
>> Battlemesh mailing list
>> Battlemesh at ml.ninux.org
> Battlemesh mailing list
> Battlemesh at ml.ninux.org
CTO, TekLibre, LLC
More information about the Battlemesh