RED: Optimum functions for computing the drop probability?
Date: Mon, 20 Oct 97 12:57:38 PDT
To: "Benjamin Nieh"
From: Sally Floyd
Benjamin 
>Read your paper "Random Early Detection Gateways for Congestion
>Avoidance". I have a question about the parameter max_p.
>
>Beside hardware efficiency (keep max_p flexible to implement
>probability update as shifts and additions), is there any other
>reason why max_p is not set to 100%?
The optimal function (between average queue size and packet dropping
probability) would probably be to have something more sigmoid on the
upper end, or to have the dropping probability somewhat quickly approach
100% as the average queue size exceeds maxthresh. But I think that is
also a second order concern. I think that the important thing is that
while the average queue size exceeds minthresh only moderately, the
packet dropping probability vary over the 0%10% range, say. And that
the packet dropping move from the 10% to the 100% value considerably
more quickly as the average queue size reaches or exceeds maxthresh.
But I don't know what the optimal function would be, and I do know that
the simple function that is there now works fairly well.
(In the earlier simulations I had max_p set to 2%, and that was
not good  based on obsolete 1993 ideas about what the typical
levels of congestion might be. Now I set max_p to 10%, and that
seems much better.) I wouldn't recommend setting max_p to 100%,
and keeping the same linear function  I think in that case,
the packet dropping probability would increase too quickly as
the average queue size exceeds minthresh.
 Sally
My current idea of a good simpletoimplement function would be a
packetdropping probability linear between 0% and 10% as ave queue
varies from minthresh to maxthresh, and linear between 10% and 100% as
ave queue varies from maxthresh to 2*maxthresh, or maxthresh + XXX,
say. But I haven't looked at it.
dropping probability






_____________________________________________ average queue
size
 
minthresh maxthresh