Subscriber Edge & Carrier-Grade NAT

The cable plant is the physical half of residential broadband. This is the other half: the point where a household becomes a subscriber, gets an address, is told what it may consume, and is counted.

It is the same plane of the network that a broadband network gateway occupies in the telephone company world — under a different name, over a different access technology, and in this case on the CMTS itself rather than on a platform behind it.

Policy, and which direction matters

Downstream and upstream DOCSIS policy governed what each subscriber class was entitled to consume, on the customer-to-headend path in both directions.

Downstream is the easy half. Upstream is where a residential plant actually fails: the return path is shared, noisy and finite, so a policy that is generous in the wrong place degrades a whole segment rather than one house. Sizing the upstream is the design decision that residential broadband lives or dies on, and it is not the one most people look at.

Addressing, before the houses exist

Address management is the unglamorous half of a subscriber network and the half that fails visibly. A household does not experience a lease problem as a lease problem; they experience it as the internet being down.

On Cisco Network Registrar that meant DHCP and DNS across the estate, with separate scopes for the cable modems themselves and for the subscriber equipment sitting behind them — the subscriber side in private space, which is what the translation below then works on. The two are one design: the scope plan decides what the translation is handed.

At greenfield buildout scale the scopes have to be laid out before the houses exist rather than extended once they do, which is the same sequencing constraint the plant itself is under. Registry allocation sits above all of it, and the reason any of this is necessary at all is that routable space ran out before the households did.

Translation is a state problem

Carrier-grade translation gets discussed as though it were an addressing trick — many households behind one routable address, problem solved. In operation it is a state problem.

The translation table is finite, and when the translation runs on the same box as subscriber termination and address assignment, that table competes for resources with everything else the platform is doing. The per-subscriber session ceiling is the real control: it decides who breaks first and which applications break for them, and applications differ enormously in how many sessions they open. A ceiling that is invisible to a household reading mail is the reason a different household’s software cannot connect.

And once one public address serves many households, the address stops identifying a household. That is a design consequence rather than a footnote — it changes what an address means to everyone downstream who assumed otherwise, from abuse handling to geolocation to anything that treats an address as a person.

Accounting

Per-subscriber rate control and accounting has to reconcile with billing, which is a harder constraint than it sounds: the network’s own record of what a subscriber consumed becomes a number somebody is invoiced for, and a discrepancy is a customer dispute rather than a monitoring gap.

Hands-on from 2000 through the Coleman Technologies tenure and onward under Michael Young Technologies past 2005. The Cisco cable expert-track written examination and the cable, QoS and telephony specialist examinations of 2003 all sit on this work.