BGP Path Selection
amb at gxn.net
Mon Aug 30 17:53:36 UTC 1999
> I keep hearing bitching about how broken the BGP path selection
> process as 1) defined in the protocol spec and 2) as implemented
> by vendors, is.
Well this is a yet-another-knob-request. I have for a while now thought
it would be useful to have an attribute which is kept local to the AS,
and is if you like an "exit-penalty". This could be added to your IGP
distance at the decision stage process. It's difficult to combine
internal IGP metrics (especially if you use next-hop-self style peerings
internally so lose the distance on the interface itself) *and* add
make intelligent use of different quality interconnects (especially
if these are interfaces on the same router).
Consider the case where you are prepared to send traffic from SOME parts
of the country "around the long way" to avoid a particularly crappy
interconnect, point, but not from ALL parts of the country.
On your point (2) I reckon Cisco implements the standard pretty
well (i.e. where it diverges it's for reasonable reasons), or at
least it does when you've switched on "bgp deterministic-med" which
gives you a good indication of how MED decisions are taken without
this knob switched on.
GX Networks (formerly Xara Networks)
More information about the NANOG