Subject: Re: Average queue length vs instantaneous queue length
Date: Wed, 26 Nov 1997 10:40:30 PST
From: Sally Floyd <floyd>

>If I choose to use instantaneous length to do "RED", do you think 
>that would do much harm?   Would that cause much oscillation.
>
>I'm thinking about this because people do EPD with instantaneous
>queue length info and get significant improvement.  So I guess
>RED would also improve things even without queue length averaging.

I would say that EPD is most appropriate to implement with the
instantaneous queue length, but that RED is **only** appropriate
to implement with the average queue length.  

EPD is invoked when it is judged that there is some danger of the
buffer overflowing, and therefore danger that later cell fragments of a
packet would be dropped.  This danger is a function of the
instantaneous queue length.

RED is specifically designed to allow for transient bursts in the
queue, but to begin to inform nodes when more persistent congestion
occurs.  The 1993 RED paper discusses some earlier research with queue
management mechanisms that use the instantaneous queue size (e.g.,
Early Random Drop).  I would assume that this would be an improvement
on Drop Tail, but it is not RED.  The use of the **average** queue size
is essential in RED to give a tolerance for transient congestion (so
that a packet is not dropped from a transient queue caused by a TCP
slow-start, for example), to eliminate global synchronization, and to
eliminate biases against bursty traffic.

- Sally