Re: [v6ops] [SUSPECTED SPAM] Google Alert - IPv6

[Dave O'Reilly <[e-mail]>]

….and there it is. Sorry, that’s what I get for reading through the thread in order from earliest to most recent posts!

Apologies!

daveor

> On 28 Oct 2017, at 06:19, Fred Baker <[e-mail]> wrote:
> 
> 
> 
>> On Oct 28, 2017, at 7:15 AM, Ole Troan <[e-mail]> wrote:
>> 
>> Fred,
>> 
>>>> The bitter truth is that we can make IPv4 'scale' forever. ;-(
>>> 
>>> Actually, from what I have been told, we can't. We can make it work for a number of sources per translated address (Xing Li gave a talk on his experiments in that regard a few years back, although I would have to do some research to find it). As I recall, he started getting user complaints when he had 100-or-so active sessions on each external address. That's far from "forever".
>> 
>> Complaints about what?
> 
> What I'm thinking about specifically is slide 6 of https://www.ietf.org/proceedings/92/slides/slides-92-v6ops-3.pdf, and some of the other slides around it. I got the number wrong: he saw issues when he had more than 512 addressed machines sharing a single address using NAT address. I believe that the context was splitting the port number so that any given address had a certain number of ports that could be tracked to it rather than fussing with logging.
> 
>> You might be assuming the IPv4 Internet isn't changing to accommodate address sharing. And there are many ways of building NAT.
>> The mechanisms I described earlier do work. It's on my todo list to get hold a sufficiently large traffic sample to run the numbers of how what sharing ratio you could achieve. It is orders of magnitude more than 1:100 (if you intended s/sessions/users).
>> 
>> Each session to 8.8.8.8:53 would burn a port.
>> But a single port could be reused for millions of connections as long as DA+DP was different.
>> (And you could of course be nice to things like SIP and give those endpoint-independent behaviour).
>> 
>> That said the path, of extending A+P to the end host is also open to us. Similarly to what Brian C proposed at IETF 29.
>> No state in network. Each host gets a single port. I.e. we have expanded the address space to 48 bits.
>> Requires new transport protocol on top of IP+UDP of course. With it's own name space. But that's pretty much happening over in transport now anyway, right?
>> 
>> Ole
>