Scalability of full mesh, Layer 3 designs is a major issue when building
very large networks. Chapter 3, “Redundancy,”
discussed problems with Layer 2, full mesh designs. Layer 3, full mesh designs
have many of the same problems.
At some point, the number of possible neighbors and paths becomes overwhelming,
and you need to reduce the amount of work that needs to be done by the core
One possible solution to the Layer 3 meshing problem is the Next Hop
Resolution Protocol (NHRP). NHRP is technically a routing protocol rather
than a new Layer 2/3 switching mechanism. Figure
9-1 provides an example network for discussion.
Now, as you know from Chapter 3, the full mesh design in Figure
9-1 results in 15=>6(6-1)/2 paths through the network. Suppose that
you want to reduce the paths through the network by making it a hub and spoke
design. It could look like the one illustrated in Figure
The real difficulty with this design (other
than the single point of failure) is the amount of traffic that must pass
through the hub router. If all of these links are 2.4 Gbps, the hub router
needs to switch traffic at 12 Gbps or faster. There must be some way to spread
this work out a bit.
If you are using a lower layer media that supports switched virtual circuits
(SVCs), such as ATM (or Frame Relay SVCs), you should be able to
take advantage of them to make direct connections between the spoke routers
The problem with this is routing. How does the router know that a given
destination is reachable through some other means than the hub router? How
does it know which SVC to use (what number to dial, so to speak) to reach
This is where NHRP comes in. In NHRP, a number of routers are configured
as route servers. Each router advertises its reachable destinations to this
route server along with a SVC to use to reach them.
When a router wants to reach a given destination, it queries the route
server to find out if there is a direct path through the cloud. If there is,
it will bring up a SVC to the next hop and pass the traffic along.
This effectively provides the advantages of a full mesh topology while
also providing the scalability of a partial mesh topology.