BGP TTL

[Geoff Huston]

>At 3/21/2002 07:16 AM, Andrew Partan wrote:
>>However even a setting of (say) 10 would limit the count-to-infinity
>>problem - see, e.g.: the current 6bone & their *very* long AS paths.
>>[As the 6bone today is mostly full of mutual-transit peerings, you
>>can get some very long count-to-infinity paths.]


Following on from Jo's quick snapshot, you can see a set of AS path length 
histograms for a whole set of route-views peers at 
http://www.potaroo.net/bgp They all show some form of strong clustering of 
AS's at a distance of between 2 and 4 AS's.

But more to the point, I now have had time to wonder why a BGP AS TTL is 
necessary for the count to infinity problem. I can't understand what the 
problem is, frankly, and maybe an explanation would help educate a poor 
dumb operator such a myself.

Let me explain my ignorance:

Now that I'm away from noisy crowded meeting rooms, I recall that Craig and 
Abha's work on BGP instability indicated that, under certain timing 
conditions, BGP would explore a whole set of alternate AS paths for a 
prefix at a constant AS Path length, as I recall, which in turn created a 
lengthy time for a remote listener to see the BGP advertisement to withdraw 
completely. What was not there was count to infinity. Of course what is 
happening here is that because BGP maintains an AS path vector, count to 
infinity is not a count through a loop, but its a potential count through 
ever-longer bogus AS paths until the original withdraw manages to catch up 
with the wave of bogus announcements.

As I understand it (and being wrong is quite a probability!) this could 
conceivably be caused by a progressive condition where a withdraw proceeds 
very slowly through BGP along one set of paths, while the same withdraw 
proceeds quickly through another set. The set of AS's which see the faster 
withdraw may see an update from an AS where the withdraw has not yet 
propagated and learn that as an alternate path, and propagate that. In 
practice I'm not sure that this behavior is susceptible to damping by an AS 
Path TTL.

So I don't understand - if AS path TTLS are a Good Idea, what is the errant 
BGP behaviour that makes it a Good Idea?

Thanks,

     Geoff