[Bloat] Bloat goes away, but with ~25% speed loss?

[Dave Taht]

On Fri, Jun 5, 2015 at 10:27 AM, Adrian Kennard <a at k.gg> wrote:
> On 05/06/2015 18:23, Dave Taht wrote:
>>>> A&A struck me as an extremely clueful ISP (I think they have had ipv6
>>>> /48s for forever?)
>>>> and it has been my impression that folk like that were using things
>>>> like HFSC + SFQ already
>>>> in their "rate limiters", and had experimented also with fq_codel by now.
>
> Oh, I meant to say, the policers used on broadband links have a very
> very simple logic of small packets (<1000) are allowed more predicted
> lag, and so VoIP "just works" as does DNS, interactive key presses, TCP
> ACK packets and all sorts of time critical stuff.

Yes, I have thought about many improvements to the linux based policer
with "bobbie", including this one.

Still, prior to running out of cpu with inbound shaping + fq_codel on
the cpe side, we were doing so well that I basically figured that the
same stuff was also very applicable to the isp's side, particularly as
we approached higher rates.

Matt Mathis is always whinging about bad policer implementations
showing up in the google mlabs datasets... certainly linux's is
insanely primitive even compared to the rfcs.

So I am curious as to how well A&A's AS20712 (?) clients are doing on
the new dslreports.com/speedtest, which shows the "bloat" grade and
actual behavior over time?

And I think dslreports is now publishing summary statistics somewhere??

 shibbolet!

-- 
Dave Täht
What will it take to vastly improve wifi for everyone?
https://plus.google.com/u/0/explore/makewififast