[PATCH] [RFC] Babel: Implement route daming with fixed delay

[dxld at darkboxed.org]

Hi Juliusz,

On Mon, Mar 06, 2023 at 03:59:14PM +0100, Juliusz Chroboczek wrote:
> > In order to prevent RTT based routing from causing persistent traffic
> > oscillations we delay core rte announcement of each prefix by a
> > configurable but metric invariant amount of time.
> 
> Could you please explain how that works?  How does it interact with cost
> smoothing and route selection hysteresis?

I would approach the stability problem from an electronics/signal
processing/control theory background. In my view the current approach is a
roundabout way of implementing a simple filter to prevent (high-frequency)
oscillations from propagating from the input (measured link cost) to the
output (the datapath).

I have essentially two gripes with the current approach:

  1) we use not one, but two nonlinear (i.e. hard to understand) filters
     for what one well designed linear (i.e. easy to reason about) filter
     could do and

  2) the continous nature of the "smoothing" causes a blowup of the number
     of bird core announcements that could be avoided.

In the present patch I'm mainly concerned with 2) so I just threw out the
smoothing to keep things simple.

Frankly I haven't yet been able to reverse-engineer the weird exponential
smoothing filter behaviour enough to build a model, perhaps you (or
someone) could expand on the rationale for the smoothing filter
construction and the addition of hysteresis?

Thanks,
--Daniel