At the last uknof meeting I had two questioners ask me for data on ADSL vs
fq_codel and various rate shapers at the tail end of my talk:

http://www.youtube.com/watch?v=quAaZKBHvs8

Until recently the cable industry was paying the bills, not the adsl
industry, and although I encouraged the questioners to do some of their own
testing on their networks, I alluded to some puzzlement over the results
we'd got with some preliminary testing with the ADSL devices we were able
to open.

We spent some more time poking into ADSL stuff recently and it turned out
the Linux "tc" rate shaping mechanism for linklayer atm and related framing
things like PPPoe has been broken in Linux for quite some time. Various
fixes have been discussed, but there is no one working the problem
directly, presently.

The email thread discussing the problem(s) is here:

https://lists.bufferbloat.net/pipermail/bloat/2013-May/001540.html

And the bug turned out to be even deeper and older than we thought. This
possibly explains a lot of the anomalies we've seen so far (but by no means
all! waay too many binary blobs in dsl!) with rate shaping with hfsc/htb +
fq_codel, and perhaps, also, some of the ones you've seen using your
existing shapers.

Whenever and however a fix or set of fixes arrives we'll go back to looking
into it.

-- 
Dave Täht

Fixing bufferbloat with cerowrt:
http://www.teklibre.com/cerowrt/subscribe.html

Reply via email to