[altq 1479] Re: RED implementation

[Kenjiro Cho <kjc@csl.sony.co.jp>]

Nguyen-Tuong Long Le wrote:
> I can follow and understand the code but my question is
> why does ALTQ/RED compute (prob/(2 - count*prob)) instead
> of (prob/(1 - count*prob)) as in the Floyd and Jacobson paper.
> 
> I'm not saying that this is wrong since one can always scale inv_pmax
> but I'm interested in why it was done this way.

It was recommended by Sally and implemented in ns.
I don't remember where it was described, though.

The idea was to keep the drop interval close to 1/Pb.
The original algorithm drops packets with the probability more than
Pb.
With the modified algorithm, the drop probability starts with Pb/2,
becomes Pb at (count = 1/Pb), and 1 at (count = 2/Pb).

-Kenjiro