CGNAT - Seeking Real World Experience

[Adam <adamkuj () gmail com>]

I'm crunching the numbers on the cost effectiveness of implementing CGN vs
IPv4 auctions. The determining factor is how many ephemeral ports are
reserved for each customer. This is for a residential broadband environment.

Is anybody doing deterministic NAT/PAT (i.e. each customer gets X ports -
no more, no less)? My thinking is that X=8192 would cover even the power
users. However, with only 8 customers per public IPv4 address, CGN is not
worth the trouble. With X=8192, our money and time would better be spent
acquiring additional IPv4 space. Are people successfully using a smaller
deterministic port allocation? What's your X?

If I can't make the numbers work for deterministic NAT, I might be able to
live with dynamic port assignments. Specifically, I'm referring to what
vendor J calls "Port Block Allocation". I was thinking 1024 ports per
block, with up to 8 blocks per customer (and a bunch of log correlation to
determine who was using which ip:port tuple at a given datetime). I *can*
make the math work out in favor of CGN if the average customer uses <= 3072
ports (3 blocks). But is that going to be enough? I'd love to hear other
people's experiences.

Thanks!
-Adam