[Bloat] fixing bufferbloat in 2017

[Toke Høiland-Jørgensen]

Dave Taht <dave.taht at gmail.com> writes:

> On Wed, Nov 23, 2016 at 10:18 AM, Sebastian Moeller <moeller0 at gmx.de> wrote:
>>
>>> On Nov 23, 2016, at 19:09, Mikael Abrahamsson <swmike at swm.pp.se> wrote:
>>>
>>> On Wed, 23 Nov 2016, David Lang wrote:
>>>
>>>> Deploy what we already know to work on the real edge devices and things get vastly simpler.
>>
>>         One problem will be that the actual edge devices are often ISP
>>         supplied and hence extremely cost sensitive; combined with increasing
>>         bandwidth in many ISP offerings, having CPEs that can perform ingress
>>         shaping at “modern” rates looks challenging if the same CPEs also need
>>         to be very cheap. Egress shaping is a different kettle of fish though
>>         and for most asymmetric plans the CPE either should have enough punch
>>         or might be amendable to a BQL-like solution that could e actually
>>         relatively computationally cheap. But for that to happen we would need
>>         to convince ISPs and/or CPE chipset manufacturer (or better those
>>         engineers that create the drivers for the SDKs).
>
> One benefit of the fq_codel on wifi work is that for homes that are
> primarily wifi, you no longer need inbound rate shaping to work, you
> can do it on the wifi naturally.
>
> A long term plan might be to try to develop code that could be used at
> the ISP, a transparent bridge that would take over customer rate
> shaping and queue management. Hardware "good enough" to do this,
> ranging from high end xeons to the next generation NXP products, to
> mellonox's bluefield thing, is arriving, and I anticipate being able
> to effectively fq and shape 1000s of customers with gear that costs
> less than 10k - maybe not with linux, but with vpp and/or dpdk.

I happen to know of at least one company that offers this kind of middle
box and also ships CoDel and BLUE (I think it was). Those are DPI boxes,
though (*shudder*) -- and I think the price tag is quite a bit more than
$10k... :/

-Toke