Network Computingwww.networkcomputing.com

Interestingly, the mechanism for communicating between 3Com local-area implementations and a Cascade WAN is through Ipsilon's Ipsilon Flow Management Protocol (IFMP). Networking makes strange bedfellows.

Wide area networking, at least at the level that applies to most of Cascade's customers, is about scaling. Running an environment with a few thousand routes is one thing; setting up a network with 50,000 sites (each with many nodes) is another. Setting up and tearing down virtual circuits (VCs) on an Internet scale brings even the most robust hardware to its knees.

On a typical carrier platform, a Cisco router sits next to a Cascade switch. The Cisco device routes data traffic and the Cascade switch handles the rest.

Cascade's goal is to take the Layer 3 router out of the picture and use the Cascade switch to move data with Layer 2 grace and expedience. Cascade deploys VCs over frame relay or ATM to build its Layer 2 fabric.

There are essentially two reasons why scaling is such an important issue: Not only is it difficult to set up and tear down connections quickly, but a high number of routes in a network can be problematic because it could require the use of a huge number of VCs. This is known as the N-squared problem.

How It Works If there are N sites in a network, each site needs to know how to reach every other site. Thus, the ne twork will have to know N-squared paths (actually Nx[N-1]). Routers don't have to know about the individual paths to every endpoint in the network, but, in a switched network, they should know about paths as close to the edge of network as possible. This leads to switches dealing with as many as N-squared routes.

To prevent a geometric growth in the need for VCs, Cascade uses a route-aggregation system called Multipoint to Point (MTP). MTP builds a tree data struct ure within each switch, eliminating the need for VCs to each end from every other endpoint. Special VCs are set up as trunks for all sites on a switch, so rather than knowing about N-squared points at the edge of a network, a Cascade network has to know only about roughly N switches. Thus, the growth of VCs is linear rather than geometric.

This addresses the growth issue, but it doesn't address the potential need to create and tear down many virtual connections very rapidly. Here, Cascade takes a brute-force approach by building permanent MTP VCs from each switch to every other switch.

To handle IP routing, Cascade will run IP routing software on each switch and will maintain route-to-VC translation tables on every interface card. Cascade believes it can handle as many as 250,000 routes with this approach. That's roughly five times the size of the existing Internet, according to Cascade.