[Bloat] DOCSIS + Codel paper published by cablelabs

[Dave Taht]

Recently Greg White and a few others at Cablelabs, along with Kathie
Nichols, did an exhaustive simulation of a cable modem with a default
amount of buffering, vs the new short queue option, vs codel, in a
wide variety of scenarios.

The paper has just been released officially.

http://www.cablelabs.com/downloads/pubs/PreliminaryStudyOfCoDelAQMinaDOCSISNetwork.pdf

If you dig deeply into it, you'll see how really small queues can hurt
performance (so don't do that), the effect of losing TCP syns in drop
tail (don't saturate your buffers!), the effects of overbuffering at
the driver level (hard to tell what reducing that will do in this
sim), and a few other things like that.

One problem with the web portion of the sim is that it doesn't
simulate enough streams, and more importantly, it doesn't emulate the
effects of DNS lookups. Still, it's interesting and more accurate than
many others tests.

I'm relatively happy with the voip test therein, although I structured
the rrul equivalent to be much more difficult than the simulated one.

And if you look at the (fairly few) places in the paper where codel
doesn't shine, and think about where fq_codel compensates for that,
you might go to bed with a glow on.

I look forward to the followup paper.

Thx cablelabs and kathie!

-- 
Dave Täht

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