Khirman & Son

With 40-70% of total Internet traffic, P2P is unquestionably the main factor in ISPs' technological and business considerations. Would "Managed Peer Selection" be a solution?

The high throughput of P2P file-sharing applications is often achieved by establishing multiple simultaneous connections with other peers sharing the same content. Popular content, such as pirated movies or TV shows, is usually shared by hundreds or even thousands of peers distributed across the globe. When activated, the P2P client establishes connection with 50-100 other peers, usually selected randomly from a large set of appropriate peers.

Due to the random peer selection strategy, most of the peer connections have to go out/in of a given ISP network, significantly increasing the transit cost paid by the ISP. For example, in South Korea and Japan, as much as 80% of broadband traffic leaving the country does so over under-sea cables, which results in high transit charges. A large portion of this traffic belongs to P2P and can be eliminated by interconnection of local peers.

Managed Peer Selection is an actively discussed technology to optimize P2P overlay networks in accordance with ISP network topology and business rules. This discussion was initiated by the P4P working group and is supported by IETF: an Application-Layer Traffic Optimization (ALTO) working group is about to be established. These efforts are supported by some P2P application vendors (BitTorrent, Pando Networks) as well as major ISPs (Comcast, Verizon, Telefonica).

To estimate a feasibility of ALTO/P4P aproaches for real life torrents , I collected <ip,port> information for peers from one of the most popular "PirateBay" torrents ( almost 20k peers) and mapped their IPs to corresponded ASs (Autonomous Systems - group of IPs that belong to the same ISP. ISPs usualy have one or more ASs).

For investigation I chose a torrent that represents bootlegged copy of "Hellboy 2" movie released July 11 , 2008 for screening in US and Canada, but not yet legally available in other countries.

Data collection had been performed August 16, 2008. Peers information had been collected by periodic "announce" polling of the ThePrateBay tracker.

"IP to geo" and "IP to AS" mapping had been performed using WHOIS service from whois.cymru.com .

To make a long story short - for such kind of swarm (20k peers):

• almost 1/2 of peers belong to ASs with 100 or more peers,
• almost 2/3 of peers belong to ASs with 50 or more peers.
• less then 15% of peers belong to ASs with 10 or less "neighbors".

For illustration, the following graph represents cumulative distribution of the peers as function of the AS rank.

(another interesting graph is here)

Speculation:

Usually, 10 "good neighboring peers " sufficient to provide BitTorent client with maximum download bandwidth. If P2P application manage to give priority to connections inside of their own AS and "localise" traffic inside of ISP (AS) segment, localisation will eliminate 70-80% of BitTorrent traffic on the global backbone.

Certainly, it is a BIG speculation - I intentionally chose an unusually big BitTorrent swarm to this investigation. It will be interesting to evaluate AS distribution for small and medium swarms and extrapolate to all known torrents.

Please find more details in my work notes. You also welcome to download raw data in "per-AS" aggregation. Any comments, addition, corrections, etc are VERY welcome!.