Re: [v6ops] Happy Eyeballs and preferring IPv6 [was RE: CGN vs Native IPv6 latency]

["Dan Wing" <[e-mail]>]

...
> And port exhaustion should not be expected in the real world of
> stateful translation that may use port overloading

It can help, but if all the subscribers visit the same popular 
IPv4 address, port overloading can fail to provide service -- its
law of large numbers collapses.  There is a proposal to improve one 
of the problems (TCP server going into TIMEWAIT, see (*) and
RFC6191) but the number of active connections to a popular server 
remains a problem if the server is suddenly more popular than 
anticipated by the port allocation.  Stateless MAP ("A+P")
techniques can suffer significantly from attempts to use port
overloading, because their user base is smaller (often one 
house), and a stateful CGN suffers fewer problems (because
it can utilize more port numbers if there is sudden popularity
to a single IPv4 server).  The use of SPDY, which reduces the 
number of TCP connections to servers, will help all of these problems.

(*) http://www.ietf.org/proceedings/83/slides/slides-83-tcpm-2.ppt
    http://tools.ietf.org/html/draft-naito-nat-resource-optimizing-extension

Port overloading also interferes with, or breaks, peer to peer 
applications such as SIP, Skype, Bittorrent, Octoshape, and the 
upcoming WEBRTC.  I know those applications are often not well loved 
for a variety of reasons.  That is documented in REQ-1 of RFC5382, which 
says port overloading is bad.  I realize business reasons cause a 
lot of networks to be deployed violating that requirement.

-d