E.1. Some Historical Perspective

It is useful to briefly review the history of the topic, beginning with what is sometimes called the "classic model" of how a router makes routing decisions. This model predates IP. In this model, a router speaks some single routing protocol such as RIP. The protocol completely determines the contents of the router's Forwarding Information Base (FIB). The route lookup algorithm is trivial: the router looks in the FIB for a route whose destination attribute exactly matches the network prefix portion of the destination address in the packet. If one is found, it is used; if none is found, the destination is unreachable. Because the routing protocol keeps at most one route to each destination, the problem of what to do when there are multiple routes that match the same destination cannot arise.

Over the years, this classic model has been augmented in small ways. With the deployment of default routes, subnets, and host routes, it became possible to have more than one routing table entry which in some sense matched the destination. This was easily resolved by a consensus that there was a hierarchy of routes: host routes should be preferred over subnet routes, subnet routes over net routes, and net routes over default routes.

With the deployment of technologies supporting variable length subnet masks (variable length network prefixes), the general approach remained the same although its description became a little more complicated; network prefixes were introduced as a conscious simplification and regularization of the architecture. We now say that each route to a network prefix route has a prefix length associated with it. This prefix length indicates the number of bits in the prefix. This may also be represented using the classical subnet mask. A route cannot be used to route a packet unless each significant bit in the route's network prefix matches the corresponding bit in the packet's destination address. Routes with more bits set in their masks are preferred over routes that have fewer bits set in their masks. This is simply a generalization of the hierarchy of routes described above, and will be referred to for the rest of this memo as choosing a route by preferring longest match.

Another way the classic model has been augmented is through a small amount of relaxation of the notion that a routing protocol has complete control over the contents of the routing table. First, static routes were introduced. For the first time, it was possible to simultaneously have two routes (one dynamic and one static) to the same destination. When this happened, a router had to have a policy (in some cases configurable, and in other cases chosen by the author of the router's software) which determined whether the static route or the dynamic route was preferred. However, this policy was only used as a tie-breaker when longest match didn't uniquely determine which route to use. Thus, for example, a static default route would never be preferred over a dynamic net route even if the policy preferred static routes over dynamic routes.

The classic model had to be further augmented when inter-domain routing protocols were invented. Traditional routing protocols came to be called "interior gateway protocols" (IGPs), and at each Internet site there was a strange new beast called an "exterior gateway", a router that spoke EGP to several "BBN Core Gateways" (the routers that made up the Internet backbone at the time) at the same time as it spoke its IGP to the other routers at its site. Both protocols wanted to determine the contents of the router's routing table. Theoretically, this could result in a router having three routes (EGP, IGP, and static) to the same destination. Because of the Internet topology at the time, it was resolved with little debate that routers would be best served by a policy of preferring IGP routes over EGP routes. However, the sanctity of longest match remained unquestioned: a default route learned from the IGP would never be preferred over a net route from learned EGP.

Although the Internet topology, and consequently routing in the Internet, have evolved considerably since then, this slightly augmented version of the classic model has survived intact to this day in the Internet (except that BGP has replaced EGP). Conceptually (and often in implementation) each router has a routing table and one or more routing protocol processes. Each of these processes can add any entry that it pleases, and can delete or modify any entry that it has created. When routing a packet, the router picks the best route using longest match, augmented with a policy mechanism to break ties. Although this augmented classic model has served us well, it has a number of shortcomings:

• It ignores (although it could be augmented to consider) path characteristics such as quality of service and MTU.

• It doesn't support routing protocols (such as OSPF and Integrated IS-IS) that require route lookup algorithms different than pure longest match.

• There has not been a firm consensus on what the tie-breaking mechanism ought to be. Tie-breaking mechanisms have often been found to be difficult if not impossible to configure in such a way that the router will always pick what the network manger considers to be the "correct" route.