Re: [babel] Alvaro Retana's Discuss on draft-ietf-babel-rfc6126bis-12: (with DISCUSS and COMMENT)

["Vigoureux, Martin (Nokia - FR/Paris-Saclay)" <[e-mail]>]

Juliusz,

thanks again.
Based on my understanding of the discussions with Alvaro, the pending 
points were:
(A) he was expecting a RECOMMENDED route selection algo and this could 
for example be the most common algo in babel implems. I'm not sure if 
A.4 strictly addresses that.

(B) the “Plen MUST be 0” issue

(C) is resolved

-m

Le 2019-10-18 à 18:22, Juliusz Chroboczek a écrit :
> Dear Alvaro,
> 
> Unless I'm mistaken, I have not received a reply to my mail of 21 August.
> Thus, I have no idea whether you found my explanation of the issues
> related to route oscillation convincing.
> 
> I have just added the following appendix to the document, and referred to
> it in the (normative) section about route selection.  Please let me know
> if it satisfies you, and if it does not, please explain what more you
> would like me to do.
> 
> A.4.  Route selection
> 
>     Route selection (Section 3.6) is the process by which a node selects
>     a single route among the routes it has available towards a given
>     destination.  With Babel, any route selection procedure that only
>     ever chooses feasible routes with a finite metric will yield a set of
>     loop-free routes; however, not all route selection policies will
>     yield good reasults.
> 
>     The obvious route selection procedure is to pick, for every
>     destination, the feasible route with minimal metric.  When all
>     metrics are stable, this strategy leads to convergence to a tree of
>     shortest paths (assuming the metric is left-distributive).  There are
>     two reasons, however, why this strategy leads to instability in the
>     presence of continously varying metrics such as ETX (Appendix A.2.2).
>     First, if two parallel routes oscillate around a common value, then
>     the smallest metric strategy will keep switching between one and the
>     other.  Second, when a route is selected, congestion along it
>     increases, which may increase packet loss, which in turn causes its
>     metric to increase; this is a feedback loop, of the kind that is
>     prone to cause persistent oscillations.
> 
>     A simple strategy to avoid this kind of instabilities is to only
>     switch routes when the target route's metric is smaller by some
>     constant K than the currently selected metric.  However, this
>     approach is difficult to tune well: if K is too small, then it fails
>     to avoid oscillations, and if it is too large, then it leads to sub-
>     optimal routing.  A better strategy is to apply hysteresis
>     (sensitivity to a route's history) to the route selection algorithm.
>     The following hysteresis algorithm appears to yield good results with
>     a variety of metrics.
> 
>     For every route R, in addition to the route's metric m(R) maintain a
>     smoothed version of m(R) written ms(R) (we suggest computing ms(R) as
>     an exponential average of m(R) with a time constant proportional to
>     the Hello interval).  If no route to a given destination is selected,
>     then select the route with the smallest metric, ignoring the smoothed
>     metric.  If a route R is selected, then switch to a route R' only
>     when both m(R') <m(R) and ms(R') <ms(R).
> 
>     Intuitively, the smoothed metric is a long-term estimate of the
>     quality of a route.  The algorithm above works by only switching
>     routes when both the instananeous and the long-term estimate of the
>     route's quality indicate that switching is profitable.
> 
>