This was a double-wide session presented by Clem Leung, from IBM Networking Systems in Research Triangle Park. He talked very quickly, using big words and a Chinese accent (he's a native of Hong Kong) so I probably only understood half of what he had to say. I still came away with a severe case of information overload.

He started with a quick review of the OSI model, ranging from the physical layer up to the application layer. He pointed out that when you implement new function, it is cheaper (i.e. costs less aggregate resource) to implement it on the lowest layer you can.

Connection-oriented and Connectionless protocols were reviewed. Did you know that Appletalk confirms connections once every ten seconds? This is a performance nightmare for larger Appletalk nets. Novell confirms every 90 seconds, a little smaller problem, but a problem nonetheless.

Bridges are simple-minded devices that link together two LANs, making them appear to be one big LAN. They operate on OSI's "data link" layer. They don't support the concept of "class of service", because there aren't any classes! A 2 megabyte file download will completely hose over any interactive traffic you have going through the bridge. The best, but only partial solution is to throw bandwidth at the network (T1-T3).

Time-outs can be a problem on bridged networks when you have connection oriented protocols at work. If someone is sending the proverbial 2 megabyte spreadsheet on a bridged network, an interactive connection oriented session might time out.

Clem characterized the IBM 3745 as a combination router and gateway. Can a 3745 be replaced by a plain old router? Maybe it can, if you have intelligent stations on the net. The NCP still performs boundary function such as polling for dumb (3270) terminals. Routers don't do this. (Clem said "Who wants to do this anyway? The mainframe is dying, or so they say.")

Routers are complex beasties which require careful planning and somewhat more money than simple bridges. Clem warned that routers are not plug-n-play devices; implementing them is not as simple as an office LAN.

If you implement a router on the Internet, be aware that routers use IP addresses. IP addresses are 32 bits long, and there is only room in a class "B" IP address for 65,536 physical networks. This limitation won't be fixed for a couple of years.

Routers live on the OSI "network" layer and implement the "data in a cloud" idea. LANs connect to routers, and the routers all talk to each other using point-to-point router protocols. How a message gets from one LAN to another is entirely up to the routers; they make the routing decisions on the fly. Some routers do this simply -- they just count the number of router hops, and send data the shortest distance. Other routers (Cisco among them) are much smarter and collect detailed transmission metrics for more intelligent, dynamic routing.

Clem is "really a fan" of frame relay for PPP, mostly because it provides bandwidth on demand.

Some routers can "spoof" unproductive polling.