However these are with a very high address-sharing ratio (several thousands users per address). Using a sparser density (<= 64 users per address) is likely to show much less dramatic user impacts.
I think you have the numbers off, he started with 1000 users sharing the same IP, since you can only do 62k sessions or so
These numbers were not off. From page 19: "...we should assign at least 1000 [..] ports per customer to assure good performance of IPv4 applications" "At least 1000 ports per customers" is roughly the same than "less than 64 users per address" as I stated above. Btw, 90% of subscribers have less than 100 active connections at any time, if I read these tiny graphs correctly: http://www.wand.net.nz/~salcock/pam2009_final.pdf
and with a "normal" timeout on those sessions you ran into issues quickly.
Agreed for UDP, but most of these sessions are TCP, which arguably time out rather rapidly after a FIN and an extra hold time. Normal duration of a TCP session is usually under a few seconds. This study saw an average connection time of 8 seconds for DSL, but it's from 2004. http://www.google.com/#q=A+Comparative+Study+of+TCP/IP+Traffic+Behavior+in+B... /JF