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

[Dave Taht]

On Fri, Jun 5, 2015 at 10:59 AM, Adrian Kennard <a at k.gg> wrote:
> On 05/06/2015 18:57, Kevin Darbyshire-Bryant wrote:
>> It was the uplink side and the recent adoption of Zyxel kit which
>> made me wonder out loud to AA-Andrew earlier today regarding A&A
>> bufferbloat experiences/testing on that side of things with the new
>> modems.  You're an ISP that would have some clue in that regard and
>> hence hopefully a bit of clout with the OEM to do things right (see
>> baby jumbo frames support)  It's me being curious again...sorry!
>
> We can try... Working hard to fix some showstoppers like MTU on bridging
> and the like, first.

I found dslreports.com's summary stats page:
http://www.dslreports.com/speedtest/results/isp/AS20712

not enough samples, but pretty good results. Comparisons were
interesting also. I love a competitive marketplace! (And am admiring
all the tools for continuous link monitoring AA does)

On the downlink, I am relatively uninformed until today. A lot of ISPs
have mentioned HFSC+SFQ without much details.

There are several things, conflated together, that are hurting dsl
performance on the uplink nowadays.

1) A lot of DSL modem firmware used to some form of fq buried deep in
the driver. FT used to mandate SQF, for example.

2) a lot of modems would exert ethernet hardware flow control at very
minimal packet depths. (hardware flow control is so correct for a
cable, fiber, or dsl "modem" - but as manufacturers started embedding
switches into the modem, this feature has been getting lost)

3) connecting routers used to have a default packet depth of 100 (or
less!) and thus responded to flow control more sanely than the current
linux default of 1000. I would be surprised if firebrick had a packet
depth that high.

as for 2 and 3, I know a LOT of people that passionately hold onto
their old dsl modems because they are "better" than anything newer
they've tried. I know one guy that treasures his circa-1998 one...

however, as fq_codel and pie (any latency sensitive aqm) work GREAT
with hardware flow control, this would work better than any fixed
packet limit. A metric ton of people have reported results like ipfire
did in this case:

http://planet.ipfire.org/post/ipfire-2-13-tech-preview-fighting-bufferbloat

(And I keep hoping the DCB people in DCs are taking notice.)

4) In order to get higher reliability (for things like multicast udp
tv), a lot of DSL providers use "interleaving", which incurs quite a
bit of added latency on the link.

5? anyone?

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