[Cerowrt-devel] Biggest bufferbloat offenders in 2019?

[Dave Taht]

you might get more traction on this on the bloat, ripe, and nanog lists.

On Sun, Jun 16, 2019 at 9:27 AM Valdis Klētnieks
<valdis.kletnieks at vt.edu> wrote:
>
> So I'm writing up something, and The Great Google is failing me. A few things
> I can't seem to track down:
>
> Have Microsoft and/or Apple gotten any sort of AQM into their products?

fq_codel became the default on osx wifi devices over 2 years ago. I
know that the codebase is shared with IOS,
but to this day I've not got confirmation that it is on or not on IOS
on wifi or 3g.

I'm really happy to note that intel added support
for fq_codel for wifi in linux 5.1 - and neither I or toke noticed
until someone pointed it out. I'd love someone
with the right iwl devices on their laptop to try it (I'm pure ath9/10k here)

microsoft has not shown a pulse around here lately.

> Have all the carrier routers (Cisco, Juniper, etc) gotten their act together? If not,
> who's still lagging?

The only major cisco bufferbloat solution appeared for a single cisco
device a couple years back. Weirdly I have 6 folk from arista now on
my linkedin feed, and their product is as bloated as the come.

On my bad days, I tend to think that the chief impact of the
bufferbloat project was merely to hold carrier gear buffer sizes
constant (sigh, not even reduce them) as bandwidths went up, not our
fancy schmancy algorithms. I'm hoping high end customers are buying
more shallow buffered carrier products or using VOQ effectively.

on the lower end, like I said, bandwidths went up, buffer sizes didn't:
http://www.dslreports.com/speedtest/results/bufferbloat?up=1

> Who are the top offenders in the currently-shipping CPE,

CPE is a different term than "home router", and for all I know, CPE
still universally sucks. what sort of CPE? GPON fiber ONTs are not bad
at 60ms
worth of buffering at 100mbit. early 5g tested out at 1.6 sec.

Independently purchased home routers are looking pretty good though.

> and who's gotten
> decent anti-bufferbloating into currently shipping gear?

I would prefer we don't the term "anti-bufferbloat" as that is used by
the netduma folk, which while it does use fq_codel for part of it,
seem to have
got framing wrong and does other packet inspecting stuff I'm not into.
I do hope they keep improving their code, but I always worry about
anything
that needs external databases to work. Streamboost, for example, also
used fq_codel in some versions, but their database stuff I've heard
has
been basically discontinued.....

so I use "bufferbloat-fighting" or something like that instead.

The wifi stuff is spreading like wildfire, the sqm stuff more slowly.
all QCA derived products for the wifi side (e.g eero/google
wifi/openwrt), openwrt and related for default fq_codel support on
everything, all of linux tends to default to fq_codel now, with redhat
8 being the last join the party.

someone should maybe poke into what yocto uses as a default.

Somehow doing a comprehensive survey would be good.

> _______________________________________________
> Cerowrt-devel mailing list
> Cerowrt-devel at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cerowrt-devel



-- 

Dave Täht
CTO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-831-205-9740