RED: distributions for intervals between marked packets.
To: Pankaj Gupta
Subject: Re: Question about 1993 RED paper
Date: Thu, 05 Feb 1998 15:40:16 PST
From: Sally Floyd
Pankaj -
>1. What is the explanation for having a uniformly distributed random variable
as opposed to
>a geometric variable (well, the simulations suggest that, but I am
>missing an intuitive understanding).
A geometric distribution is medium-tailed, and a uniform distribution
is light-tailed. And a light-tailed distribution is better, because
you are less likely to have these long periods with no drops at all.
But there is nothing special about the uniform distribution other
than its light tails, I think. Any number of other light-tailed
distributions might do just as well.
>2. Is there any real evidence that suggests that an implementation of
> RED with uniformly distributed intermarking time X will perform better
> than one with geometric distribution?
Lots of simulations. But nothing published.
>3. What about if we change the intermarking time distribution to be periodic
> deterministically. E.g. if we keep a counter that counts up to C, and
> drop every Cth packet? C can be a function of avg-minth.
> How will this or how is this expected to perform?
I think that anything periodic could lend itself to deterministic
patterns of bias in certain scenarios. But roughly periodic with a
little dollop of randomness should do just fine.
- Sally
This is the only reference that I have off-hand about tailweight
of distributions...
T. Hettmansperger, and M. Keenan,
``Tailweight, Statistical Inference, and Families of Distributions -
A Brief Survey'', in
{\em Statistical Distributions in Scientific Work\/}, V.1,
G.~P.~Patil et al. (eds), Kluwer Boston, pp. 161-172, 1980.