Network Computingwww.networkcomputing.com

Of the distributed routing vendors that did participate, we tip our hats to Bay and Foundry. Bay's BCN surprised us with its ease of configuration, variety of interfaces and overall stability. In the Layer 3 switch comparison, Foundry's NetIron Switching Router displayed several features that Extreme's Summit2 LAN Switch lacked, including OSPF, PIM (Protocol Independent Multicast) and DVMRP (Distance Vector Multicast Routing Protocol) routing protocols, IPX routing support and Layer 3 policy-based VLANs (virtual LANs).

Our test results indicate that both the Foundry NetIron and the Extreme Summi t2 are extremely fast Layer 3 devices with a variety of features that make them quite suitable for enterprise deployment. However, we also uncovered some serious shortcomings in their software design--Bay's BCN was the only product that operated flawlessly on our real-world network. The Foundry and Extreme routers run code versions 2.0 and 1.09, respectively, while the BCN is up to version 12.

There is something to be said for the tried-and-true product, but these start-ups will get it right, and when they do, their products will be a force to be reckoned with.

Real-World Gore Network Computing's real-world environment offers an unusual opportunity to push routers and Layer 3 switches to their extremes. The backbone network of our University of Wisconsin lab carries more than 1,300 routes that provide connectivity to the university campus, the Internet and the vBNS (Very High-Speed Backbone Network Service). Unlike Layer 2 switches, routers must be constantly aware of the routing updates that p eriodically flood the network. These updates have the potential to seriously impair performance. The size of the routing table on your network may also have an adverse effect on routing performance.

To test these routers, we connected each of the products to our live production network and configured it to participate in that routed network. We also configured each interface on the router to be a distinct subnet. We used a combination of Pentium Pro workstations and Netcom Smartbits analyzer ports to generate traffic through the router.

More Than Packets Per Second Going into this testing we knew that measuring raw packets per second in a sterile environment wouldn't tell us much about how the router would function in a production network. Indeed, the most difficult part of routing traffic is not in moving the packets, but in deciding where they should go. This is especially true in an environment where route paths are changing.

To determine h ow well these routers deal with changing environments, we first set up a mesh of traffic between the router ports. We benchmarked the performance of each router while it was participating in our live network topology. Each packet the router forwarded could take one of approximately 10,000 different paths. We repeated this test with the router isolated from the network. In all cases, the performance of the forwarding engine was the same.

We then ran the same benchmark again, but this time we pulled the router's connection to the real world, forcing it to flush 1,365 routes while under load. Here, we saw significant differences in the products' behavior. When subjected to this "route flush," Bay's BCN never blinked a lid. Foundry's NetIron router also handled the change gracefully.

However, Extreme's Summit2 had a fit. The loss of routes caused the Extreme router to stop forwarding for nearly 12 seconds while it flushed the routes. We told Extreme engineers about the problem, and they provided us with a fix that prevented the routes from being flushed when the cable was pulled. The engineers also acknowledged that their route-flush algorithm was very untuned, and that a more permanent and effective fix would be integrated into a future software release.