Introduction on Peer to Peer systems

Transcription

1 Introduction on Peer to Peer systems Georges Da Costa 1/55

2 Goal of this Lecture What can P2P do, not only as a buzzword What it can't do Shows some examples & algorithms A Survey and Comparison of Peer-to-Peer Overlay Network Schemes, by Eng Keong Lua and al. in IEEE Communications survey and tutorial March 2004 Harnessing the Power of Disruptive Technologies published by O'Reilly, 2001 dacosta@irit.fr 2/55

3 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 3/55

4 Plan 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 4/55

5 Universal What have in common Net Meeting, Skype, Ekiga Irc, Msn, Icq, Jabber Kazza, Freenet, Napster, Gnutella Ebay, Flickr, Facebook 5/55

6 Denition Philosophical one Participants gathering their resources in order to achieve a common goal dacosta@irit.fr 6/55

7 Why? Available resources Large Hard Drives Powerful CPUs Correct connexion to Internet Users want More freedom No link to commercial companies No infrastructure cost 7/55

8 A new (?) solution : Peer to Peer systems Denition Participant gathering their resources in order to achieve a common goal Computers are running the same code There is no global view of the system View is limited to neighboors Everyone has the same rights and duties dacosta@irit.fr 8/55

9 Peer-to-Peer: New name, old concept An architecture already there Internet connects most of existing computers Most computers are not fully used Idle time > 75% on personal computers Storage systems are mostly empty Already used between servers Usenet DNS IP Routing 9/55

10 Comparison with Client/Server In client/server each node is either a Client or a Server. Usually there are a few Servers and lots of Clients. Client/Server systems suer from single point of failure. Client/Server are mostly static, at least the Servers. Peer to Peer systems are dynamics. Client/Server systems need human administrators Client/Server does not scale dacosta@irit.fr 10/55

11 Comparison with Client/Server II Node Client Client Node Client Node Node Server Client Node Client Client 11/55

12 Comparison with Client/Server II When a new participant joins a service, the service increase the resource consumption Client/Server : increases the server power/connectivity Peer to Peer : uses the resources given by the participant dacosta@irit.fr 12/55

13 Not so easy Wanted Scalability (1K,100K,1M nodes) Dynamicity Security (user, task) Transparent For the user (CPU,memory,disk) For the network Heterogeneity Self-organization Participation (66% of Free riders) Go through NAT/Firewall 13/55

14 Self-organization Participants High volatility & voluntary No central administration Resource discovery Heterogeneity Hardware Users (15% of users have 94% of les) Distribution of the resources Trust 14/55

15 What's not new Partial solutions Scalability : Farm of web servers Dynamism : Cell phones Fault tolerance : Redundant servers dacosta@irit.fr 15/55

16 Current Peer to Peer systems Available applications File sharing Distributed storage Content delivery Distributed computing Telephony/Chat Games 16/55

17 Current Peer to Peer systems (cont) Widely used 2004 : According to British Web analysis rm CacheLogic, BitTorrent accounts for an astounding 35 percent of all the trac on the Internet more than all other peer-to-peer programs combined and dwarfs mainstream trac like Web pages Start-ups Skype (ok, no more a small start-up) BitTorrent UbiStorage dacosta@irit.fr 17/55

18 Two worlds Internet Users Problem of security Large scale No control Motivation needed Private Area (Corp., Univ.) Other mean of security Medium to large scale Total control 18/55

19 Plan 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 19/55

20 Index Method Client Client Server Client Client Client Static connexion Sending the files list Sending a request for a file File transfert Users send the list of their les to a server To nd a le, you send a request to the server It answers with the list of clients owning the le You directly contact the owners for the transfer dacosta@irit.fr 20/55

21 Index Method II Systems Napster, Mojonation, Yaga, Filetopia, Problems Scaling Price HotSpot Attack Single point of failure 21/55

22 Useful when... Small number of client Need a total control of transfers (video game industry) Performance is more important than cost dacosta@irit.fr 22/55

23 BitTorrent Same approach as Napster, but : Downloads are done in parallel One server per le Server manages all the details of transfers Server enforces the rule The more you share, the more you get Dierences Specialized for large les Distributed due to the One server per le rule dacosta@irit.fr 23/55

24 Privacy No privacy Napster : The server knows all transfers BitTorrent : For each le, a server knows all transferts dacosta@irit.fr 24/55

25 Flooding Client Client Client Client Client Client Client You send your request to your neighbors They forward it to their neighbors, and so on until reaching the Time To Live depth Users with les corresponding to the request answer 25/55

26 Flooding II Systems Gnutella, Direct Connect Characteristics Distributed structure No single point of failure Denial of service dicult (but possible) Not scalable Resource consumption (network) Not complete answers 26/55

27 Privacy Average to good privacy Onion routing (good privacy) No global view of the system Usually easy to obtain the shared list of a node Dicult to have a global impact dacosta@irit.fr 27/55

28 Super Peers SuperPeer SuperPeer SuperPeer SuperPeer SuperPeer SuperPeer Peer Peer Peer Super Peers act as local servers Some reliable nodes act as super peers Super peers are connected with a gnutella protocol Each super peer acts as a local server for several peers dacosta@irit.fr 28/55

29 Super Peers II Systems Gnutella2, Kazaa Characteristics Less distributed structure Scalable Some nodes are more loaded Some nodes are more important Less resource consumption due to limits of number of answers 29/55

30 Plan 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 30/55

31 A case study : Freenet Ian Clarke, University of Edinbourgh, (1999) Keywords A peer-to-peer le sharing system Provide anonymity for authors and readers A web of Freedom Principle Files are referenced by key The key is obtained by SHA-1 on the le The key is routed to localize the le dacosta@irit.fr 31/55

32 Content Driven routing algorithm Routing table contains a set of key/node pairs Take the nearest key in the routing table to obtain the next node to consult. Nearest key = by lexical comparison Request abc a abb acd 1 node c node b 3 8 c Data b d 4 7 e dacosta@irit.fr 32/55

33 On the path of the answer File is replicated on the path in the cache Cache : variant of Last Recently Used Routing tables are updated the graph evolves (new links = new entries) c d a b e abb node c acd node b abc node d Old links New links (entries) dacosta@irit.fr 33/55

34 Anonymity Reader Writer Impossible to know if a user is forwarding or initiating the request Impossible to know if a user is the last to receive a le Once in the system, the writer can disconnect Impossible to know if someone insert some le or forward it dacosta@irit.fr 34/55

35 Some properties Self-organization of the graph Nodes specialize in les with close keys (learning process) Good properties (Small World) File are automatically replicated in function of their popularity Hot-spots are limited Tolerant against attacks 35/55

36 Drawbacks Counterpart Files might disappear (LRU cache) The network is heavily loaded Dicult to update a value Impossible to know what is hosted locally dacosta@irit.fr 36/55

37 Plan 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 37/55

38 Pastry Principle Each le has a key Each node has an identier Node with identier Id manages keys whose values are near Id Queries Content driven queries Sux forwarding dacosta@irit.fr 38/55

39 Pastry II Links to the neighbor F4B E98 Table of the node x098 x198 x298 x398 x498 x598 x698 x798 xx08 xx18 xx28 xx38 xx48 xx58 xx68 xx78 xxx0 xxx1 xxx2 xxx3 xxx4 xxx5 xxx6 xxx CA 2BB D598 Neighbors of Id are chosen as to have the sux of their identier in common with Id dacosta@irit.fr 39/55

40 Pastry III Pros ln(n) messages guarantee Good path redundancy Cons Dicult to keep a synchronized neighbor table Problem of data redundancy No adaptation to data dynamicity dacosta@irit.fr 40/55

41 Plan 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 41/55

42 Current state of Peer to Peer systems A lot of redundant systems Typically File Sharing Common basic component Distributed index (Key, Value) Key is typically the lename Value is typically the le content or where to obtain it Each Key is associated with a node dacosta@irit.fr 42/55

43 Generic Interface Node Id : k-bit identier (unique) Key : k-bit identier (unique) Value : bytes (can be a le, an IP,...) Generic DHT (Distributed Hash Table) put(key, value) Stores (key, value) on the node responsible of key value = get(key) Retrieves the data associated with key dacosta@irit.fr 43/55

44 Current implementations Software Kadmelia Chord CAN Usage File sharing Naming Chat service Databases 44/55

45 Still limited Fundamental Problems Complex request Data coherence Request with several answer Implementation diculties Distribute workload evenly Keys Requests Only local information Dynamic information 45/55

46 Chord structure 0 Nodes are distributed on a circle Keys are assigned to the node with Id just before their value 192 Key in store of 61 62, dacosta@irit.fr 46/55

47 Neighbors 0 Log(N) neighbors Neighbors are nodes Id + 1, Id + 2, Id + 4,..., Id + 2 i,..., Id + 2 k 1 (modulo 2 k ) dacosta@irit.fr 47/55

48 Routing algorithm Forward to the neighbor which is prior to the key Query needs at most Log(N) messages Node responsible of Id between 35 and 50 Query for dacosta@irit.fr 48/55

49 Chord characteristics Ecient If a (key, value) exists, the query will nd it Fast : Log 2 ( ) = 23 Small neighbors table Log 2 (N) dacosta@irit.fr 49/55

50 Chord characteristics Some problems Security and privacy Attack How to test and evaluate such system? Real performance (instead of number of messages) 50/55

51 Physical overlay Logical topology mapped in the physical network : Query N2 Answer N1 dacosta@irit.fr 51/55

52 Plan 1 What is P2P 2 First generation systems 3 Self-organized systems 4 Structured systems 5 Distributed Hash Table 6 Conclusion dacosta@irit.fr 52/55

53 Conclusion Peer to Peer systems are ecient for several uses (using border resources) Recent systems are scalable Low cost alternative to Client/Server Field old enough to be used in real cases Still not perfect Trust & certication Anonymity Security Performance Layers fees 53/55

54 When to use Peer to Peer systems Limited budget Large audience Trusted users Dynamic system, but not too much Do not need guarantee Do not need control 54/55

55 Vision of the future User centered No more servers All content provided and served by users Only cooperation of peers Wikipedia Social networks Youtube Good Ol' Time web-pages 55/55