A Survey of Peer-to-Peer Content Distribution Technologies

Transcription

1 A Survey of Peer-to-Peer Content Distribution Technologies Stephanos Androutsellis-Theotokis and Diomidis Spinellis ACM Computing Surveys, December 2004 Presenter: Seung-hwan Baek Ja-eun Choi

2 Outline Overview of P2P P2P Motivation P2P Characteristics & Benefits P2P Application Types P2P Classification Unstructured: Gnutella, Kazaa, Napster Structured: Freenet, Chord, CAN, Tapestry Other Aspects Conclusions 2/50

3 P2P Motivation Client/Server Architecture: Well known, powerful, reliable server is a data source Clients request data from server Very successful model WWW (HTTP), FTP, Web services, etc. 3/50

4 P2P Motivation (Cont d) Client/Server Limitation: Scalability is hard to achieve Presents a single point of failure Requires administration Unused resources at the network edge P2P systems try to address these limitations 4/50

5 P2P Characteristics P2P Computing: P2P computing is the sharing of computer resources and services by direct exchange between systems. These resources and services include the exchange of information, processing cycles, cache storage, and disk storage for files. P2P computing takes advantage of existing computing power, computer storage and networking connectivity, allowing users to leverage their collective power to the benefit of all. 5/50

6 P2P Characteristics (Cont d) P2P Characteristics: All nodes are both clients and servers Provide and consume data Any node can initiate a connection No centralized data source Nodes collaborate directly with each other (not through well-known servers) Network is dynamic Nodes enter and leave the network frequently 6/50

7 P2P Benefits Ease of administration Nodes self-organize adaptively No need to deploy servers to satisfy demand (c.f. scalability) Built-in fault tolerance, replication, and load balancing Scalability Consumers of resources also donate resources Aggregate resources grow naturally with utilization Reliability Geographic distribution No single point of failure 7/50

8 P2P Application Types Direct real-time communication: instant messaging Combine processing power of multiple distributed machines to perform complex computations: analysis of SETI data, prime computation Distributed database systems Store and distribute digital content: mp3 file sharing (Content Distribution) 8/50

9 P2P Classification Architecture Types: Unstructured Structured Loosely structured Here, By structure, we refer to whether overlay network is created non-deterministically or whether it s created based on a specific rules 9/50

10 Centralization P2P Classification (Cont d) Data organization Unstructured Loosely Structured Highly Structured Hybrid Napster, IM Partial Kazaa, Gia None Gnutella Freenet Chord, CAN 10/50

11 Unstructured Architectures Placement of content is unrelated to overlay topology Search mechanism is required. Appropriate for case of highly-transient node population Degrees of centralization: Purely Decentralized Partially Centralized Hybrid Decentralized 11/50

12 Purely Decentralized Purely Decentralized No central coordination Users (servents) connect to each other directly. Gnutella architecture Query: Flooding Send messages to all neighbors Response: Route back Scalability Issues With TTL, virtual horizon Without TTL, unlimited flooding E.g., Gnutella, FreeHaven query request registration reply query query registration query download registration reply query query registration registration 12/50

13 Partially Centralized Partially Centralized Supernodes Indexing & caching files of small subpart of the peer network Peers are automatically elected to become supernodes. Advantages Reduced discovery time Normal nodes will be lightly loaded. E.g., Kazaa, Edutella, Gnutella (later version) query reply query registration reply request download 13/50

14 Hybrid Decentralized Hybrid Decentralized Central directory server Advantages User connection info. File & metadata info. Simple to implement Locate files quickly and efficiently Disadvantages Vulnerable to technical failure Inherently unscalable E.g., Napster, Publius request download resigtration query reply 14/50

15 Outline Overview of P2P P2P Motivation P2P Characteristics & Benefits P2P Application Types P2P Classification Unstructured: Gnutella, Kazaa, Napster Structured: Freenet, Chord, CAN, Tapestry Other Aspects Conclusions 15/50

16 Structured Architectures Features Mapping of content and location Scalable solution for exact-match queries Examples Freenet Chord CAN Tapestry

17 Freenet Loosely Structured System Chain mode propagation Each node Local data store Dynamic routing table Each file ( node address, file key ) Unique binary key

18 Freenet (Cont d) Messages Node ID, Timeout, Src ID, Dst ID Message types Data insert : key, data Data request : key Data reply : file Data filed : failure location, reason

19 Freenet (Cont d) Data Insert Calculates a binary key Sends a data insert message to itself Receiving a Data Insert message If not taken Store the data Forwards to the closest key s owner If taken Returns the preexisting file

20 Freenet (Cont d) Data Request Chain mode propagation Receiving a Data Request If locally stored The search stops and the data is forwarded back If not Forwards to the closest key s owner

21 Freenet (Cont d) Data Fail Timeout (hops-to-live) Receiving a Data Failed Message Forwards the request to the next best node After failed through all neighbors, Sends back data filed message to the request sender

22 Freenet (Cont d) Data Reply Includes the actual data Passed back through the chain The data is cached in all intermediate nodes A subsequent request w/ the same key served immediately A request for a similar key forwarded to the node that previously provided the data

23 Freenet (Cont d) Indirect Files A special class of lightweight files Named according to search keywords Contain pointers to the real file Multiple files w/ the same key

24 Indirect Files Freenet (Cont d)

25 Freenet (Cont d) Properties Nodes specialize in searching for similar keys Nodes store similar keys Similarity of keys does not reflect similarity of files Routing does not reflect the underlying network topology

26 Chord Nodes and Files are identified by keys m-bit identifiers a deterministic hash function Mapping File ID onto Node ID Nodes store (key, data item) pairs

27 Chord (Cont d) A Chord Identifier Circle

28 Simple Key Location Chord (Cont d)

29 Chord (Cont d) Scalable Key Location

30 Chord (Cont d) Simple Key Location Routing Information: Successor pointer O( n ) Scalable Key Location Routing Information: Finger Table O( logn )

31 Chord (Cont d) Node Joining Certain keys assigned to its successor are reassigned to it Node Departing Keys are reassigned to its successor

32 Chord (Cont d) Node Joining N26 joins the network

33 CAN Content Addressable Network Hash Table Maps file names to their location ( key K, value V ) pairs stored Each node storing a part of the hash table A zone

34 CAN (Cont d) Virtual coordinate space A zone corresponds to a segment of space Key K is mapped onto a point P A deterministic function ( K, V ) is stored at the node responsible for P

35 CAN (Cont d) Virtual coordinate space

36 Retrieve CAN (Cont d) Map K to P Retrieve the value from the node covering P Routing Request is routed to the node covering P Nodes maintain a routing table Addresses of Nodes holding adjoining zones Following the straight line path in the space

37 Routing CAN (Cont d)

38 CAN (Cont d) Node Joining Allocatedits own portion of the space By splitting the zone of an existing node Node Departing Hand over hash table entries to one of its neighbors

39 Tapestry Location and Routing Infrastructure Self Administeration Fault Tolerance Stability By bypassing failed routes and nodes Plaxton Mesh Routing mechanism Location mechanism

40 Tapestry (Cont d) Routing Mechanism Neighbor Maps Local routing maps Incrementally route messages Multiple levels Level l node ID matched w/ l digits Multiple entries The number equals to the base of the ID Pointer to the closest node in the network

41 Tapestry (Cont d) Neighbor Map of Node w/ ID 67493

42 Tapestry (Cont d) Routing Path from to xxxx7 xxx67 xx567 x

43 Tapestry (Cont d) Location Mechanism Root node Provide a guaranteed node from which the object can be located Assigned when an object is inserted A globally consistent deterministic algorithm When inserted Server node Ns, object O, root node Nr Message routed to Ns to Nr (O, Ns) stored along the routing path

44 Tapestry (Cont d) Location Mechanism Location query Messages destined for O Initially routed toward to Nr Meet a node containing (O, Ns) mapping

45 Tapestry (Cont d) Advantages of Plexton Mesh Simple fault-handling Routing by choosing a node w/ a similar suffix Scalability w/ the only bottleneck (root nodes) Limitations The need for global knowledge Assigning and identifying root nodes The vulnerability of the root nodes

46 Tapestry (Cont d) Extending Plaxton mesh s Design Plaxton mesh assumes a static node population Tapestry adapts it to the transient population Adaptibility Fault tolerance Optimizations

47 Tapestry (Cont d) Optimizations Back-pointers for dynamic node insertion Flexible concept of distance between nodes Maintain cached content for failures Multiple roots to each object Adapt to environment changes

48 Other Aspects Content Caching, Replication and Migration Security Provisions for Anonymity Provisions for Deniability Incentive Mechanisms and Accountability Resource Management Capability Semantic Grouping of Information

49 Conclusions Study of P2P Content Distribution Systems Properties Design features Location and routing algorithms Two Categories Unstructured system Structured system Remains Open Research Problems