Sunday, September 14, 2008

The forgotten historical effect of switches

I had a funny little chain of thoughts the other day. I was thinking about how my office's Ghost software (which we use to do disk imaging across the network) uses an old-style packet driver for the network boot disk. Most of my younger counterparts have never seen anything like it, but these drivers used to be quite common. I can remember configuring them for the various DOS-era networking clients waay back in the 1990's. Thinking about this also reminded me about how much effort used to go into designing networks so as to minimize collisions.

I used to work for the University of Venice (Italy!) and their entire campus was setup as ONE big collision domain -- every building was tied directly together and everyone had trouble with lost connections... imagine :-) At the end of my stay there, I suggested to the IT Center that they install a switch at the central hub. At the time, this was considered an extravagance. I explained to them how I'd been going around to all their Novell servers, upgrading the LAN drivers to make them more resilient to ethernet traffic. I think they subsequently used my suggestion ("Our American consultant said...") to justify the expense and I'm sure it was well worth it. Now, of course, switches are The Standard and it's hard to find an old-style dumb hub, and consequently there's no cost premium for it. (P.S. I'm assuming my reader already knows the difference between a hub and a switch?)

Well, the final thought I had was about the problems we used to deal with on hubs. First of all, there was a lot of research into how many workstations you could group together before you started getting more collisions than proper traffic. At the time, 17 workstations was considered the ideal maximum, above that number the percentage of collisions (and rebroadcast attempts) started to climb rapidly. One vendor that seemed prescient about this was 3Com. At first, they were pilloried for the design of their NICs because the hw didn't do any collision detection. Instead they relied on their sw driver to detect it. As long as there were relatively few collisions, this design was MUCH MUCH faster. But on a traditional hub-driven LAN the 3Com cards ended-up CAUSING more collisions since they didn't 'back off' as quickly.

Nowadays, every NIC vendor uses the 3Com-type model since switches have eliminated collisions. So the whole 17-workstation-max, avoid-bad-LAN-client-3Com-NICs advice is just so much ancient history.

No comments: