[p2p-hackers] Statistcs about % of natted peers in P2P systems?

[David Barrett]

Aha, tracked down the original post, on the BEHAVE list.

http://www.nabble.com/RE%3A-Re%3A-removing-application-milestone-td11622258.html#a11628671

Here's the data itself:

> Type        Count       Connectivity Open Time
> Fullcone    150 (9%)    92%          1056ms
> Fullstatic  397 (25%)   94%          867ms
> Fw          68  (4%)    2%           374ms
> Rstrcone    598 (37%)   89%          706ms
> Rstrstatic  82 ( 5%)    91%          769ms
> symmetric 293 (18%)   94%          906ms
> 
> This is read as "598 tested clients had a restricted cone NAT (making up 37%
> of all clients tested), with an average peer connectivity success rate of
> 89%".  ("FullStatic" and "Rstrstatic" means a non-NAT'd client with endpoint
> independent/dependent filtering, respectively; "Fw" means a client that
> failed our STUN operation and thus might just have UDP blocked altogether;
> "Open Time" is the time it took to establish the connection and begin
> sending actual connection data.)  
> 
>            fullcone fullstatic fw      rstrcone rstrstatic symmetric
> fullcone   98%      96%        0%      96%      94%        98%    
> fullstatic 97%      97%        5%      97%      90%        98%    
> fw         0%       5%         0%      0%       0%         0%    
> rstrcone   96%      97%        0%      16%      66%        27%    
> rstrstatic 94%      92%        0%      66%      100%    
> symmetric  98%      98%        0%      12%      100%    
> 
> Here are the pairwise peer connectivities.  As you can see, this is a pretty
> small test (only maybe 1000 total clients), an thus not all pairwise
> scenarios got tested (ie, no rstrstatic:symmetric or symmetric:symmetric
> tests were attempted).
> 
> Thus I wouldn't put too much faith in those numbers (especially anything
> claiming 100%) until I have numbers on a broader scale.  For some reason I'm
> seeing peer connectivity problems at larger scales (average peer
> connectivity across the board is 90% in the above case, but drops to 40%
> when tested on 10x the clients), and I'm still ironing them out (there
> appears to be a problem in my test infrastructure).
> 
> Basically, I'd love to see some results on ICE in the above format, as
> that'll let me do an apples-to-apples comparison of it versus mine.
> 
> -david
>