[Ecn-sane] is FQ actually widely deployed?

[Dave Taht]

On Mon, Jul 22, 2019 at 11:21 PM Mikael Abrahamsson <swmike at swm.pp.se> wrote:
>
> On Tue, 23 Jul 2019, Sebastian Moeller wrote:
>
> >       That, and most other features the linux network stack offers (it
> > is easier ot be fast if you do less ;) ) I believe users need to
> > actively enable HW offloading, so shpould be in control whether they
> > want to trade in a ship-load of features for alloing their router to
> > punch well above its weight...
>
> On most HGW SOCs today the choice between HW offload on/off is 200
> megabit/s unidirectional with large packets, or very close to wirespeed.
>
> Anyhow, my statement is that of the users today on the Internet, their
> congestion point will not have FQ enabled, and this for 99% of users. I
> don't see any trend that this is on the verge of drastically increasing
> either.
>
> The trend I am seeing is that delay control is being deployed by means of
> RED, buffer size caps (basically implementing adaptive buffers to only
> provide 10ms buffering until tail drop), PIE/CODEL or similar.

I don't see a trend on "RED" - I see existing implementations.
Certainly most hardware offloads have fairly short buffer sizes. I
have been hoping for a pie implementation in non-cable-hw for years
now.

The title of this mail could just as well been, "have any hw offload
makers been paying the
slightest bit of attention to fixing bufferbloat with *any* aqm or fq
tech?" - with an answer of no.

On 10gige+ ethernet cards, 64 hw queues is the norm. That's fq.
There's some support for aqm
on some. I think the work on timerwheel paced stuff has high potential.

In the switch chip market we see big buffered switches and small
buffered ones, targetted at different markets. There's work in p4
that's promising on the switch front.

We've hit it out of the park on wifi, with atf and fq_codel. Very
visible "trend" there.

What happens in the offloaded word on small routers like the
edgerouter series is to do "qos" at a lower
than line rate, offloads are disabled and it's done in software.



The tit
> I'd be happy to be proven wrong.
>
> --
> Mikael Abrahamsson    email: swmike at swm.pp.se



-- 

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