"The Ethernet Principle" ... there's worse news. Like all queued
systems (e.g.: T1 lines, Ethernet, or standing in line at McDonalds) the effective queuing time rises as the rate of demand approaches the
service rate. With small load, the utilization is near zero,and the delay is the time to process one transaction (e.g., get a burger).
Queued systems work great as long as the utilization is less than about 70%. But as utilization increases and approaches 100%, the effective delay gets longer and longer (theoretically, infinity). In reality, you get out of line and go to Taco Bell.

It's the long delay that drives people nuts. Everyone wants a high peak rate (for downloads and quick response time). But a queued system that's oversubscribed (i.e., beyond around 70%) has a long queuing delay. To make matters worse, most IP-based systems have an adaptive re-transmission time; each time they get tired of waiting in line, they *increase* the wait time for the next response (in order to prevent network congestion). In that case, performance is worse than ISDN or dial-up connections.

The cable companies never had to think about this stuff (queuing and over-subscription) when they were just blasting out video. The result:
instead of satisfied customers, they get grouchy and disillusioned ones with low blood sugar (... and we can't just say "ˇyo quiero Taco Bell!").

-Dave Retz
SYV.COM