1. Peer-to-Peer Systems Rodrigo Rodrigues Peter Druschel Max Planck Institute for Software Systems

2. Paper Overview  Survey paper  Long list of references  Presents a high view of various aspects of P2P systems  Describes distributed hash tables in more detail

3. In the beginning  1999  Napster music sharing system  Gave a bad reputation to P2P systems  Freenet anonymous data store  SETI@home volunteer-based distributed computational project

4. Now  BitTorrent  Skype P2P telephony system  Skinkers enterprise communication management system  P2PLive, CoolStreaming, BBC’s iPlayer

5. What I think  Previous list mentioned commercial products  Does it mean all major issues have been solved?

6. Email from Skype To our valued customers: As 2010 draws to a close, I would like to take a moment to thank each of you for your patience, understanding, and support during Skype’s recent outage. … Kind regards, Tony Bates CEO Skype

7. Defining properties of P2P systems  High degree of decentralization  Few or no dedicated central nodes  Multiple administrative domains  Low barriers to deployment  Organic growth  Resilience to faults and attacks  Abundance and diversity of resources

8. What I think  P2P systems are  Very cheap  Very easy to deploy  Highly scalable

9. Applications (I)  Sharing and distributing files:  Napster (quickly shutdown)  Gnutella, FastTrack aka Kazaa (all decentralized)  eDonkey, BitTorrent (faster)  Streaming media:  PPLive, Coolstreaming (academia)  BBC’s iPlayer, Skinkers Livestation (industry)

10. Applications (II)  Telephony:  Skype  Scientific Computing:  SETI@home  BOINC  Other:  Distributed storage systems (Freenet)  Content-delivery networks (CoralCDN, CoDeeN)

11. Typology (I)  Degree of centralization:  Partly decentralized:  BitTorrent tracker,  Skype billing subsystem  Fully decentralized:  More scalable  Resilient to failure, attacks and legal challenges  Can have supernodes

12. Typology (II)  Overlay maintenance  Overlay is graph G = (N, E) describing set of links E among members of set N of participating nodes  If there is a link in E between two nodes, they are aware of each other  Overlays can be structured or unstructured

13. Unstructured overlays  When a node joins, it acquires a set of "neighbors" by  Contacting the tracker (BitTorrent)  Contacting a system participant  Must have a mechanism advertising these nodes

14. Structured overlays (I)  Use key-based routing  Each node has a unique identifier  160-bit integer  Identifiers are uniformly distributed  Addressing is based on keys  Each key is mapped into exactly one of the current overlay nodes  Smallest integer "larger" than key value: make identifier space circular

15. Structured overlays (II)  Key-based routing implements primitive KBR(n o, k) that produces a path going from a node n o to the node holding key k  Big tradeoff is between  Keeping paths short  Minimizing state information kept by nodes

16. Typology (III)  Distributed state  In partly decentralized systems state is maintained by  The central node(s)  The peers assigned by it/them to each node  In decentralized systems, state is kept by  The content providers  Individual peers

17. Locating data  In unstructured systems, nodes wanting to access a specific object flood their neighbors, which flood their neighbors and so on  Structured systems use distributed hash tables  All data have keys  Stored at node responsible for key value and replicated at its successors

18. Typology (IV)  Distributed control:  In unstructured systems, it is typically done by epidemic techniques  Can also build a spanning tree among the nodes if membership is fairly stable  In structured systems, it is much easier to build spanning trees

19. Content distribution  Tree-based protocols  Main disadvantage is that leave nodes ado not contribute anything  Full binary tree of height n has 2 n+1 - 1 nodes and 2 n leaves  Swarm-based protocols  BitTorrent  All nodes can participate

20. Challenges (I)  Controlling membership:  Preventing Sybil attacks  One node pretending to be many  Can require proof of work or use trusted identities (FARSITE)

21. Challenges (II)  Protecting data:  Integrity and Authenticity:  Can use digital signatures  Data stored in DHTs can be self- certifying by making DHT keys function of data themselves  Can use voting (LOCKSS)  Availability and Durability  Replicate data and keep system alive

22. Challenges (III)  Incentives:  Fighting Free riding  Big problem  BitTorrent tit-for-tat  Not always feasible  Managing P2P Systems:  Lack of centralized control can make system hard to manage  Skype collapses

23. P2P and ISPs  P2P systems consume a lot of bandwidth  Current ISP billing models assume that customers send much less bits than they receive  Flat-rate pricing for residential customers  Bandwidth-based pricing for information providers  ISPs have no way to bill anyone for P2P traffic

24. Conclusions  P2P is a disruptive technology with great potential  Major strength is lack of centralized control  Also creates new challenges that can be  Technical  Commercial  Legal