Time in Queue, bufferbloat, and... our accidentally interplanetary network

[Eric Dumazet]

Le lundi 05 décembre 2011 à 10:05 +0100, Dave Taht a écrit :

> And Eric dumazet also produced a preliminary patch a few weeks back
> that tied timestamping to before the head of a queue, but that tried to use a
> reserved field in the skb that appears from points A to Z is not guaranteed
> to be preserved.

It is guaranteed to be preserved, as it is part of skb->cb[], and
skb->cb[] content is private to each layer.

Here, Qdisc layer.

Adding a time limit is possible, all we need is a proper design and
implementation :)

Here is my suggestion :

Design a new tfifo/tred qdisc, with following properties :

Adaptative RED, (ECN enabled + head drop), but instead of using
bytes/packet qlen average, use time_in_queue average.

A single mandatory parameter to specify the tmin value
tmax default value would be tmin*3  (so that target avg is 2*tmin)
tlimit default value would be tmax*8

Adaptative RED dynamically adjusts maxp between 1% and 50%
[Using a timer, every 500ms, and AIMD]

tc qdisc add dev eth0 root tfifo tmin 1ms [tmax val] [tlimit val]


I volunteer to work on this new tfifo experimental qdisc :)

This can be used in replacement of pfifo/bfifo in a complex qdisc setup.
As it has RED included, it also can replace RED.

Please note that once skb leaves Qdisc and is delivered to device, any
time limit feature must be handled in the driver itself (if it can queue
packet for a long time period)