BGP Filtering

Date: Tue, 15 Jan 2008 15:16:04 -0500
From: "William Herrin" <>
Subject: Re: BGP Filtering

> I think I understand what you want, and you don't want it. If you
> receive a route for, say,,, and
>, you want to drop the /17s and just care about the /16. But
> a change in topology does not generally result in a complete update of the
> BGP table. Route changes result in route adds and draws, not a flood event.
> So if you forgot about the /17s and just kept the /16, and the /16 was
> subsequently withdrawn, your router would not magically remember that it had
> /17s to route to as well.


That's half-true.

The "routing table" is comprised of two components: the Routing
Information Base (RIB) and the Forwarding Information Base (FIB). The
RIB sits in slow, cheap memory and contains routes and metrics for
every route as announced by every neighbor. The FIB sits in fast,
expensive memory and contains the currently "best" route for each
destination. The FIB is built by choosing the best routes from the
RIB. Packet-forwarding decisions are made by consulting the FIB.

Opportunistically filtering routes from the RIB would have exactly the
problem you point out: routing updates are incremental. The knowledge
that the /16 has been withdrawn may not accompany the knowledge that
the /17s are available.

Opportunistically filtering more-specific routes from the FIB,
however, could be very valuable at the edge of the DFZ. If Cisco
supported such filtering, those Sup2's could have another few years of
life left in them. With 512m ram in a two-transit provider scenario a
Sup2 could handle upwards of 1M routes in the RIB. Unfortunately, they
can only handle 244k routes in the FIB.

Ben, coming back to your question: I don't think there is a way to
make the software filter the routes inserted into the FIB. I don't see
a reason why it couldn't be programmed to do that. But the fine folks
at Cisco didn't see fit to write that software. Its a pity 'cause it
would be very useful.

This leads to one 'obviuous' approach to the matter. a smart BGP 'proxy'
that accepts 'full' data from all the peers, manages the RIB and outputs
=only= the current 'most desired' subset of routes to the target router.
In _that_ router the RIB and FIB are going to be equivalent,

This gets the 'hog' part of things off into 'commodity' hardware, and
where you can afford to burn CPU cycles implementing 'smarts' to compensate
for other people's 'dumbs'.

At a quick glance it looks like this would not be terribly difficult to
build, using Quagga/Zebra as a base platform.


That is where I got to last night with my cogitations before I feel