AW: [e2e] Queue size of routers

[David P. Reed]

At 11:46 AM 1/20/2003 +0100, Sebastian Zimmermann wrote:
>Doesn't RED allow us to define a "target" buffer size that is larger
>than 0?

If you just think about Little's theorem, the impact of RED is not to 
stabilize the system around a "target buffer size".

Little's theorem says that when arrival rate grows, queue size grows 
non-linearly, getting to infinity at the point where arrival rate = queue 
size.  This leads to a phase transition that propagates towards the 
congestion sources, filling all the intervening buffers back from the point 
of congestion.

What RED does is apply gradually increasing pressure to the arrival rate, 
via drops, so that the congestion sources get the message more quickly, 
while still allowing the system to absorb burstiness in its buffers.

ECN works even faster, because the simulated drops propagate faster.   (in 
fact, thinking about it, I wonder if flagging packets with ECN from the 
front of their queue (rather than the tail) would provide a faster 
congestion notification that would be worthwhile trying...).