BitTorrent

Incentives Build Robustness in BitTorrent
Bram Cohen

Peer Distribution

[Content]

• Tracker return the interaction peer list to a peer. Random graphs algorithm is used to get good robustness, i.e. avoid the power law graph's unbalance.


• Peers info each other the pieces it had obtained. Then Piece Selection.

[Questions]

• How random is the Random Graphs algorithm? Does it take the proximity or anything else into consideration?


• How many peers in the Tracker's return list?


• Is the peer list no change after ONE Tracker-Peer interaction? Can Peer ask another list? Which situation Peer shall reask?

Piece Selection

[Content]

• The basic principle is: Each one do for the optimization of the total system, e.g. Rarest First, Random First Piece. So, BT is intrincically a Cooperation system:


• Peer only download one piece at one time, but it shall request sub-pieces of this piece from multiple peers.

[Questions]

• Because the cooperation is limited into the peer lists returned from the tracker, so the list from tracker is Key? Anyone analyze and try to optimize the tracker algorithm?

Choking Algorithm

[Content]

• Unchoke for 4 peers with quickest download rate. Rate is measured at 10s.


• 1 Optimistic unchoke for possible quicker download peer. Rotated at 30s. In the 30s, the business can be reached in both directions.


• It seems it is based on an assumption that up and down link between peers are symmetric or near-symmetric, so the quickest download of one peers also means quickest download for the other peer, so peer shall select each other. The deal is reached. The precondition is:


• 
  
  
  
  • Peers have different pieces, so they need each other. So Piece Selection is important. Rarest first is key here.
  
  
  
  


• 4 thread make TCP saturate. 4 is also the net-ants default threads, which seems not to be occassional.

[Questions]

• If the pair is not equal? e.g.


• 
  
  
  
  • link is not symmetic, i.e. uplink rate is lower than downlink rate.
  
  
  • one peer has lower bandwidth than the other,
  
  
  
  


• It is predicable that the low-capability peer shall be unchoked for high-capability peers. It is understandable, and there are methods in BT, e.g.


• 
  
  
  
  • The Random First Piece make the low-capablity peer has a rarest piece.
  
  
  • The Seed shall prefer peers with noone else upload, so at least, Seed shall not abandon the low-capablility peers.
  
  
  
  


• But is there any analysis for this issue?

[Futher Reading]

• bitconjurer.org

Technorati : BitTorrent, P2P