[j-nsp] Juniper best path selection process

[Joe Metzger]

Does anybody have a reference to the real best path
selection process actually used by an M7i running 9.3
that documents all of the tie-breaking criteria, and
when it might decide to just blow off recomputing a new
best route if the important attributes didn't change?

I was poking around today on a router that is excepting
2 identical iBGP feeds of routes to black hole. Approximately 70%
of the routes from 1 server were marked as best and 30%
from the other.

The published selection algorithm docs only include deterministic
criteria, and so I would have expected deterministic results and
that the routes from the black-hole server with the lowest router
ID would be preferred 100% of the time.

http://www.juniper.net/techpubs/en_US/junos/information-products/topic-collections/nog-baseline/topic-20357.html

The underlying system feeding routes to the 2 black hole
servers guarantees that both servers are sending identical
lists of prefixes, with identical local-pref, as-path,
origin code, med value, source protocol, etc. The import
policy explicitly sets the next-hop to discard for all
routes from both servers so IGP distance shouldn't be a factor.

The only differences between the routes announced from each
black hole server visible in show route extensive are router
ID, Task, and the time that the route was announced. It appears
that it is showing the oldest route as the 'best' one.

This isn't causing me any trouble, but it looks wrong, and making
routing decisions based on age has bit me in the past. :-(



Joe Metzger
metzger at es.net