BGP Confederation versus Route Reflectors

[Philip D'Ath]

I have a client with data centres all around the world.  They have a public AS number from pretty much every registry.

Lets consider a smaller part of the picture, just Asia Pacific.  They use maybe 6 data centres in the APNIC area, all using the one public ASN from APNIC.  The data centres do not have a common carrier/connectivity between them.  A couple of sites are connected via VPN, 2 or 3 are connected with one carrier and the others with a different carrier again.  Most sites have more than one link to other sites for redundancy.  All links are either layer 2, or point to point layer 3.  Most sites have 3 or 4 BPG devices in them, all in the same AS.

Creating a full iBGP mesh within a singe data centre is ok.  But I have had to use a lot of route reflectors to build the connectivity between the different data centres.

I have never used a BGP Confederation before, but I am starting to think this will be the most manageable way.  Make each datacentre a a separate sub-AS, because I can easily create a full mesh within a data centre.  Then peering between the sub-AS's will become trivial, while still providing the appearance of a single AS to all the external peers.

Have you use BGP Confederation before?  Had a good/bad experience?

[Masoud Pourshabanian]

Hello,

I did not use BGP confederation before, but it is common. Route reflector is a good option, but you need make a tree structure. In my opinion, router reflector is easier to config, but BGP confederation is more flexible. You can devide your main AS to several sub ASs with different administration groups and you can enjoy using EBGP behaviour between confederation groups.

I used RR before, but my network was structured in a tree form originilly, so I did not have any proplem to make a logical tree structure.

If the network is sporatic, sometimes RR is confusing to design due to RR rules regarding RR client and NON-client and so on.

You will have the same result with both in terms of the nexthop and propagation of attributes.

Masoud