[Codel] ingress rate limiting

[Dave Taht <dave.taht@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1260 bytes --]

I know we have a tendency to focus on just one side of the queue. In
asymmetric networks, both sides get important.

Attached are some plots I took yesterday (from the rrul test prototype
which measures ping rtt in combination with 8 up/down tcp streams) of
normal comcast behavior, the prototype of what I'm calling "cake" on
both ingress and egress, and same prototype with ingress shaping off.

these were taken on a "normal" comcast line of 20Mbit down, 2Mbit up.

2_20mbit_comcast/irene_speedboost_unshaped.ps # "normal" behavior
2_20mbit_comcast/irene_cake_ingress_egress.ps # simple_qos.sh shaper in cerowrt
2_20mbit_comcast/irene_cake_no_ingress.ps # ingress limiting turned
off, egress left on

I also spent some time looking at the behavior of dropbox and netflix.
(no graphs attached)

Both mark their traffic CS1. I knew dropbox did, didn't know netflix did!

How dropbox responds to congestion is interesting. It's pretty good at
not doing bad things to the network. Is there a paper on how it works
somewhere?

netflix does the classic video trick of loading up a giant burst every
few seconds, then going quiescent.


-- 
Dave Täht

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

[-- Attachment #2: irene_cake_ingress_egress.ps --]
[-- Type: application/postscript, Size: 116890 bytes --]

[-- Attachment #3: irene_cake_no_ingress.ps --]
[-- Type: application/postscript, Size: 112299 bytes --]

[-- Attachment #4: irene_speedboost_unshaped.ps --]
[-- Type: application/postscript, Size: 118446 bytes --]