Chapter 2: Network and Switch Architecture

Chapter 2 is the technical core of the 1997 report. It follows a single modem call from the living room to the internet, names the eight elements it crosses, and asks a narrow question: which of them can actually run out of room?

Reading the network before judging the claim

The Bell operating companies went to the FCC in 1996 with a claim about equipment: dial-up sessions were loading the telephone network in ways it was never built to carry. Lee L. Selwyn and Joseph W. Laszlo of Economics and Technology, Inc., writing for the Internet Access Coalition in January 1997, answer that the question cannot be settled by how long people stay online. It is settled by naming the hardware being asked to do more than it can.

Two terms carry the chapter. Holding time is how long a connection stays open once set up. Blocking is the chance that a new call attempt finds every path taken, measured as a grade of service: P.01 means one attempt in a hundred is turned away. A four hour session produces an enormous holding time on the caller's own line. Whether it produces blocking depends on whether that traffic passes through something other people are waiting for.

One modem call, eight numbered elements

Figure 1 of the report drew the local network as a chain of eight numbered elements, and the numbering is what makes the analysis possible. Element 1 is the subscriber line, the wires from the customer premises to the telephone company building serving that address. Element 2 is that building's switch, known as the serving central office, the end office, or the Class 5 switch: it supplies dial tone and decides where the call goes next. Element 3 is the interoffice trunk, the shared circuits carrying calls between central offices. Element 4 is the tandem switch, an intermediate point used when two end offices lack a direct trunk group or the direct route is full.

Element 5 is the end office at the far end, serving the internet provider. Element 6 is the lines between that office and the provider, normally a hunt group: a block of lines published under one telephone number, with the terminating switch handing each call to the first line that is not busy. Elements 7 and 8 are the provider's own modems, servers and software, plus the data network beyond.

Nothing along that path is special to the internet. A provider on ordinary analog lines is just another business customer, and a call to it is routed like any other local call. What differs is the shape of the traffic: calls from across a metropolitan area converge on the office serving the provider, so whatever concentration exists shows up there.

"Any congestion at these points is clearly separate from the local exchange network."

Selwyn and Laszlo on elements 7 and 8, the provider's own equipment and the data network beyond it

Only three points can affect anyone else

Several of those elements cannot inconvenience anyone but the parties to the call. Element 1 belongs to the subscriber alone: a line held for six hours is one its owner has chosen not to use otherwise, and no other customer is queued behind it. Element 6 is the provider's own supply of lines, so congestion there means it bought too few, and the effect lands on people trying to reach it. In a network running Common Channel Signalling System 7, that busy signal comes from the originating switch before routing, so the failed attempt consumes no interoffice capacity.

That leaves the public network proper: elements 2, 3, 4 and 5. Of these the report identifies three where blocking could reach other customers. If enough subscribers on one end office call a provider at once, paths through that switch, element 2, can fill, and other customers on it lose the ability to place or receive calls. Interoffice trunks, element 3, normally offer several routes between two offices, directly and by way of the tandem, so exhaustion should not occur, though it stays conceivable where trunks are underprovisioned or traffic jumps without warning. And the terminating switch, element 5, can fill with calls for the provider.

The loads attributed to provider lines were not extraordinary. The Bell studies reported peak hour figures of roughly 26 to 32 CCS per line, or hundred call seconds per hour, against a ceiling of 36. A 900 station private branch exchange averaging 3 CCS per station needs 97 trunks to hold a P.01 grade of service, and each trunk then carries about 28 CCS. Such loads had been routine for decades.

Where dial tone actually runs out

The place where dial tone genuinely runs out has a name. Subscriber lines do not each get a path through the switching matrix, because an average line sits idle for most of the busy hour, the sixty minutes in which a switch carries its heaviest load. Instead they terminate on a Line Concentration Module, which lets a large group of lines share a smaller group of paths. That is line concentration. In the Nortel DMS-100 used in the report's diagrams, one module can terminate as many as 640 lines against as few as 180 paths. The 181st subscriber to lift a handset gets no dial tone, and callers to that line get a fast busy, the reorder tone.

The scale is worth holding on to. One switch handles on the order of 20,000 lines and needs up to 32 of these modules; the largest urban offices may hold 156. The 5ESS, the AT&T design later sold by Lucent Technologies, does the same job with a Line Unit, handling up to 512 lines across 64 ports. Congestion, where it occurs, occurs inside one module, on one switch, in one neighborhood. Carriers had ordinary answers: mix lines whose peaks do not coincide, or deload the module so half rather than a quarter of its lines can be off hook at once.

The rest of the plant was built with more headroom. Fiber optic interoffice cable is laid with strands left dark until electronics are added, so capacity rises without new construction. Switch processors carry far more capability than typical line and trunk configurations use. Elements provisioned one for one are non-blocking by design.

"The economic rationale for this price differential is difficult to understand."

Selwyn and Laszlo on tariffs that priced 24 analog lines below an equivalent digital trunk group

A bypass nobody was priced to buy

There was a clean answer for a customer whose lines really are busy most of the hour, and it sat inside the switch already. A high volume subscriber can bypass the concentration module and terminate on a dedicated trunk port, functionally the same port the module's lines share. Contention for dial tone then disappears. Tariffs said otherwise. Ordering 24 analog business lines in a hunt group obliges the carrier to run 24 loops, 24 line cards and 24 line ports; the same 24 channels over a T-1 need two loops, no line cards and one trunk port. Yet in New York, California, Maryland and Oregon, the hunt group priced 22 to 65 percent below the trunk group.

The closing arithmetic is worth keeping. A switch costing $250,000 a month and carrying 150 million minutes costs about $0.0017 per minute. Raise the total to 165 million minutes without touching the peak and it falls to $0.0015. Even in the worst case, with the busy hour shifting to 10 pm and peak demand climbing from 500,000 to 550,000 minutes, a five percent cost increase against the larger volume still lands at $0.0016.

What 127 offices can and cannot show

Which brings the chapter to the size of the evidence. There were 23,686 central office switches in the United States. The Bell studies examined 127 offices, roughly 2.4 percent of the approximately 5,200 offices in the territories concerned. Selection matters more than the count. These were not offices drawn at random: the studies measured only offices serving internet and online service providers, where such traffic actually terminates. Everywhere else, data was a negligible share of what the switch carried. Presenting the most affected 127 offices as the condition of the network describes the sample, not the network.

And inside even those offices, only one of the eight elements ever produced an example. Not the originating switch, not the interoffice trunks, not the tandem: only element 5, the switch terminating calls to the provider, and specifically the concentration modules within it. That is a provisioning question with well understood remedies, made worse by a tariff that discouraged the best of them. Not a meltdown.