Re: Pseudorandom Flow Labels

[Thomas Narten <[e-mail]>]

> Discrepancy is a measure of the deviation of a point set from a
>  uniform distribution.

Actually, the core issue is that we do not need the Flow Label to be
uniformly distributed.

In an ideal world, that would be a nice property to have. And all
things equal, better to have that propert than not.

But for ECMP and other types of things being discussed, such a
property is simply *not* required to get perfectly fine results.

What is *required* is that the hash function (or whatever function
that is used) on the router maps the tuples in a *uniform* way across
the range of possible outputs.

If you have 10 links, and all your Flow Labels are clustered around
low ten values, but in an approximately uniform way, a simple modulo
hash will get you the kind of distribution you need.

The range of values of the flow label itself does not need to be
uniformly distributed.

That is why incrementing a counter for each new flow provides adequate
properties. I remain unconvinced that we need to require that they
exhibit pseudo randomness, or even be uniform.

If that is a requirement, I'd like to see the justification, not just
hand waiving.

Thomas