Comparison of WAN IGP protocols

Most companies with large WANs use either RSTP+flooding/learning or OSPFv2 for two purposes: to route around network link failures, and to automatically find the correct path to a destination address on a large network with many hops. Including RSTP in the comparison is a realistic acknowledgement of the way L2 switching is abused these days. Including OSI in the comparison should reveal some habitually irritating aspects of switched IP networks that are mere accidents of history, and others that are more fundamental.

I'll provide background about how Ethernet switching works, what is an IGP, and what the now mostly-abandoned supposed-future OSI world feels like.

IGP's are not just for fault-resilience. They also function as a network management protocol: like SNMP, it's impractical to deal with a very large network without using an IGP.

The talk will not discuss EGP's, exterior gateway protocols like BGP. All these IGP's are used within one administrative domain. They are never used between one ISP and another, nor between a customer and his ISP.

I say ``loosely about'' because I'll be introducing three protocols, only one of which is an IGP in common use on today's Internet:

* RSTP + flooding/learning. L2 switches use a combination of RSTP and the flooding/learning system to route Ethernet frames. They can route traffic around a failed link, and they can direct traffic toward a wireless user that's roaming among access points.

* OSPF. This is the standard routing protocol for IPv4. There are other important/popular ones like 'Integrated IS-IS' and 'EIGRP', but OSPF is the best example and probably the most popular. ISP's and big companies use this protocol to route around the failure of WAN links within their own networks, and to manage their large networks.

* IS-IS. This is _the_ routing protocol for the ISO/OSI CLNP (ConnectionLess Network Protocol). OSI networking was designed many years ago through expensive necktie conferences in mountain resort towns, and proposed as the replacement for IP on the future Internet. It included CLNP which was analagous to IP. No one fell for it. At least, no one uses it at the edge, although it's still used ubiquitously, I'm not sure how exactly, on most Sonet/SDH rings. More importantly, almost everything complicated and difficult at the core of the modern Internet right now is a simplified/adapted version of some earlier OSI Master Vision which you can find referenced in the bibliography of the relevant IETF RFC. The difficult algorithmic and conceptual pieces in BGP, OSPF, LDAP, and many other things, are borrowed from OSI standards.

I want to teach people about these three protocols for two reasons. First, what's similar and what's different about the three types of routing? How, and with what limitations, do small and large networks route around failures? Second, which limitations are abstract, essential problems of network routing, and which are quirks of a particular implementation that has become overwhelmingly dominant, like IP or L2 switching.

I think most people don't really understand how L2 switches work---they just think ``they're like hubs, only better.'' Switches unfortunately fall a bit short of that ideal lego-networking model. Now that such a bastard hack has attained such prominence in large networks, I think we should have a second look at switches with the aim of borrowing ideas from them, like how most of the modern Internet's new ideas are borrowed from OSI.

Also, we've become accustomed to IPv4 networks where there is a strict rule: if you move to a different part of the L3 network, you have to change your end system's IP address to match the new subnet. OSI CLNP does not work that way. I think this will surprise and interest many people, as it did me.

I think the talk will broaden people's minds by introducing three standards central to our Internet heritage of which most people have vaguely heard, but have no idea how they work or even exactly what they are. Optimistically it'll help them think about new possibilities for how to design large networks and protocols, and will help them do it in a more historically grounded way than the usual creativity-heavy bikeshed moment of ``I've just invented this great idea for a new kind of peer-to-peer filesharing network! see, it's shaped like a tetrahedron, and when one node `drops out', then there's the `discovery phase,' to replace it, and...''