I fail understand, however, why ATM over SONET is desirable when there is such a loss to overhead, especially when viable alternatives may exist to get more bang-for-the-buck.
Perhaps someone could enlighten me on this particular datapoint?
Telecom and datacom have historically been different worlds. Telecom kept coming up with faster ways to do standard-interface long haul, and datacom kept jumping into each new telecom technology. Eventually the telecom people and the datacom people started talking to each other and they decided to come up with an ubertechnology that would solve all problems in both fields, once and for all. Only a small number of the datacom people in the room knew anything about IP. IP people have historically been "tin cups and string" users; as a rule, we believe in context free switching (no virtual circuits), and we believe in end-to-end {congestion control, error checking, connection management, and reassembly}. IP's view is that you can build arbitrary complexity on top of a good fabric, whereas an overspecified fabric makes down-simulation very expensive and often unusable. Traditional datacom and telecom people believe the opposite of all these things. Witness X.25 as the canonical example of traditional principles in action. Both the X.25/ISO/ATM/telecom crowd, and the IP/datagram/end-to-end crowd, believe that their ideology is best since it enables all other things to be layered on top. What ATM has is an excellent set of buzzwords and a lot of vendor support. What IP has is an actual market with real end-to-end products and users. With iPhone and cu-seeme and mbone starting to come into wide use, we are running into a situation where IP needs more bandwidth than it can get from most of the pre-ATM telecom solutions. The IP looks at ATM and says "this thing wastes a lot of bandwidth*delay on things I'm already doing end-to-end, and it does them in a way that my end-to-end implementation considers pathological, so let's look for another solution." Hmmm, ATM seems to be running on top of SONET. SONET is faster than modulating a bent coathanger, let's run IP over SONET. A worldwide fabric of IP-over-SONET and IP-over-bent-coathangers and IP-over- tin-cups-and-string is _going_ to occur, and soon. A worldwide fabric of end-to-end ATM may or may not occur, depending on the PR capabilities of the folks who think it's the right way to go. One of these nets will come up sooner, work better, and be cheaper. The other one will go the way of X.25. Where's Padlipski when we need him?