[v6ops] Re: Traffic control protocols (PCP and UPnP IGD)

[Brian Candler <[e-mail]>]

On 26/07/2024 18:06, Daryll Swer wrote:
> > Assumed IPv6 mostly network deployment, 464XLAT is the appropriate 
> choice.
>
> 464xlat is *stateful*, MAP-T is *stateless*, I heavily prefer 
> stateless approaches, why avoid chances to minimise computing overhead 
> in general…

Correct me if I'm wrong, but my understanding is that MAP-T is stateless 
inside the provider's network only. It still requires a stateful CPE for 
every customer, to map internal traffic from multiple customer devices 
to the provider-assigned IP and port range (NAPT44). And what the 
customer then gets is dual-stack on their side of the CPE.

If you want to do MAP-T but with "IPv6 mostly" on the customer's own 
network - i.e. when a device declines to pick up an IPv4 address - then 
you'd still need this stateful part to take place somewhere. As you say, 
traditionally this would be a stonking big stateful PLAT in the 
provider's network.

You can't delegate this to the client devices. It would mean all client 
devices sharing the same IP address, and having to apportion the port 
range between themselves.

However, I don't see in principle why you couldn't put an intermediate 
node in the CPE doing the stateful mapping to the assigned IPv4 
address/port. Effectively it would be a PLAT back-to-back with MAP-T CE. 
The CPE would then have to advertise itself as the PREF64 to the 
customer side network.

I've never seen anybody attempt this, but I'd certainly be interested to 
see it if it exists.