P2P Systems

1
P2P Systems technologies

• Zacharioudakis Giorgos

2
Presentation overview

• P2P architectures typical systems
• Technical issues
• Popular P2P Systems
• Research areas
• Project JXTA technology
• Vision about SeLene project

3
What is Peer-to-Peer?

• Definition Nodes of equal roles exchanging
  information and services directly
• Scale millions (billions?) of peers
• Nature of peers PCs
• Application lightweight semantics (e.g.,
  file-sharing)
• Is this a new idea?
• IP routing
• DNS, NTP
• Distributed Databases

4
P2P vs. Distributed DBMS

• Traditional DDBMS Issues
• Transactions
• Network Partitions
• Distributed Query Optimization
• Interoperation of heterogeneous data sources
• Reliability/failure of nodes
• Complex features do not scale

• Example P2P application file-sharing
• Simple data model query language
• No complex query optimization
• Easy interoperation
• No guarantee on quality of results
• Individual site availability unimportant
• Local updates
• No transactions
• Network partitions OK
• Simple Amenable to large-scale network of
  PCs

5
P2P Applications

• File sharing
• Napster, Gnutella
• Instant Messaging
• Jabber
• Distributed Computation
• SETI_at_home
• Web services
• Akamai

• Distributed storage
• Freenet
• Anonymity, censorship resistance
• Mixmaster remailers
• Red Rover, Publius
• Cooperative work
• Groove
• Other ...

6
Technical issues

• scalability
• fault tolerance
• speed
• bandwidth consumption
• processing cost
• security
• anonymity

• publishing/retrieval
• metadata
• semantic querying
• availability of results
• interoperability
• ...

7
Metadata and Interoperability

• Metadata
• System metadata (e.g filename, bitrate, filesize
  etc)
• Resource metadata (e.g relations, hierarchies
  etc)
• Currently, queries are in the form of keyword
  matching
• We would like to perform queries in more
  expressive languages, taking advantage of
  semantic knowledge metadata
• Technologies
• Programming interfaces
• XML-RPC, SOAP, HTTP, JXTA
• Data and metadata representation - common
  ontologies and format
• XML, RDF

8
Different Approaches to Distributed Search

• Network topology based architectures
• Relies on the organization of peers within the
  network to route requests
• These approaches focus on how to reduce the
  diameter of the graph representing the
  distributed networks
• Content based approaches
• Message content is used in either the
  organization of the network or the routing of
  messages or both
• These approaches focus on how to reduce the query
  path-length of the access structure they use

9
Spectrum of Purity

• Hybrid
• Centralized index, P2P file storage and transfer
• Napster, SETI_at_home
• Super-peer
• A pure network of hybrid clusters
• Morpheus, e-donkey
• Pure
• functionality completely distributed
• Freenet, Gnutella

10
Publishing/Requesting/Responding

• hybrid
• central indexing
• each node registers to a central index
• queries are performed to the central index
• retrieval is done from other peer nodes

• pure
• each peer manages its own index about local
  (remote) resources
• queries are typically performed with broadcasts
• retrieval is done from responding peers that
  hold the requested resource

• super-peers
• some nodes act as coordinators and manage indices
  for a subset of nodes
• each node registers to its local coordinator
• queries are performed to the coordinators, which
  in turn communicate as in a distributed p2p
  system with other super-peers
• retrieval is done from other peers that hold
  the requested resource

11
Representative P2P Systems

• Network topology based architectures
• Napster
• Gnutella
• Morpheus
• Content based architectures
• Chord
• P-Grid

12
Napster (hybrid)

• Membership Each client joins a server, where he
  registers its local files to the central index
• Query A client make queries to the central
  server which returns references to the clients
  that actually hold the resources
• Retrieval The client connects to other peer
  clients and retrieves the resource. The selection
  is performed by the user but it could be done
  automatically based on bandwidth, load or other
  criteria

13
Napster (hybrid)

14
Gnutella (pure)

• Gnutella is not a system it is a protocol, with
  various existing gnutella clients that implement
  it.
• Membership Through a predefined static list
  with addresses or through host caches, a peer
  can connect to a set of gnutella clients. After
  connection a client expands its list of known
  addresses with the lists obtained from other
  peers.
• Query A peer broadcasts a query to its known
  peers these forward the query to their known
  peers and so on until a max TTL (packets Time To
  Live) is reached, which is the depth limit of the
  query.
• Retrieval Peers that hold the requested resource
  respond to the peer that issued the query.
  Through the reverse path of the query, the
  originating peer finally discovers a list of
  peers having the resource and then obtains it
  from one of them.

15
Gnutella (pure)
Breadth-First Search (BFS)
16
Gnutella (pure)

• Each peer maintains a small minimum number of
  simultaneous active connections
• These peers are selected from a locally
  maintained host catcher list containing the
  addresses of all known peers
• Peer discovery
• watching PING-PONG messages
• noting the addresses of peers initiating queries
• receiving connections from previously unknown
  hosts
• out-of-band channels (IRC, Web)
• host caches
• Query propagation upon receiving a query a peer
  broadcasts it to all peers that is currently
  connected to, and so on as a chain letter
• If a peer has a file that matches the query,
  sends an answer back (though it still forwards
  the query). This process continues to a maximum
  depth (search horizon)

17
Morpheus (Super-Peer)

• Self organizing network
• Neither search requests nor actual downloads pass
  through any central server
• The network is multi-layered, so that more
  powerful computers get to become search hubs
  ("SuperNodes")
• Any client may become a SuperNode, if it meets
  the criteria of processing power, bandwidth and
  latency
• Network management is automatic - SuperNodes
  appear and disappear according to demand

18
Morpheus (Super-Peer)
SN2
SN4
SN4 12.34.56.78
SN3
SN1
19
Morpheus (Super-Peer)

• Intelligent downloads
• Morpheus implements a type of fail-over system
  that attempts to locate another peer sharing the
  same file, and automatically resume the download
  where it left off at the failed host
• When Morpheus search engine finds that more than
  one active peer is serving a particular file, it
  associates the list of peers with the file for
  later reference
• If the user instructs Morpheus to download the
  file, it can distribute the download task over
  this list of peers
• SuperNodes act like local search
  hubs
  and proxy search requests
  on
  behalf of their connected peers

20
Chord (content based search)

• Chord is a lookup service, not a search service
• Based on binary search trees
• Provides just one operation
• A peer-to-peer hash lookup
• Lookup(key) ? IP address
• Chord does not store the data
• Uses Hash function
• Key identifier SHA-1 (key)
• Node identifier SHA-1 (IP address)
• Both are uniformly distributed
• Both exist in the same ID space
• How to map key IDs to node IDs?
• A key is stored at its successor node with next
  higher ID (modulo N)

M
0
21
Chord (content based search)

• The goal of Chord is to provide the performance
  of a binary search which means O(log N) query
  path-length
• In order to manage a maximum path-length O(log N)
  each node maintains a routing table (called
  finger table) with at most m entries (where
  mlogN)
• The ith entry in the table at node n contains
  the identity of the first node s that succeeds n
  by at least 2i-1 on the identifier circle (all
  arithmetic modulo 2m)
• i.e., s successor(n 2i-1), 1 i m
• Note that the first finger of n is its
  immediate successor on the circle

existing node
not existing node, but a possible value in ID
space
22
Chord (content based search)

• Important characteristics
• Each node stores info only about a small number
  of possible IDs (at most logN)
• Knows more info about nodes closely following it
  on the identifier circle
• A nodes table does not generally contain enough
  info to locate the successor of an arbitrary key
  k

0
1
7
6
2
5
3
4
23
Chord (content based search)
Finger Table Allows Log(n)-time Lookups

• How do we locate the successor of a key k?
• If n can find a node whose ID is closer than its
  own to k, that node will know more about the
  identifier circle in the region of k than n does
• Thus n searches its finger table for the node j
  whose ID most immediately precedes k, and asks j
  for the node it knows whose ID is closest to k

N5
N10
N110

• By repeating this process, n learns about nodes
  with IDs closer and closer to k
• Gradually we will find the immediate predecessor
  of k

K19
N20
N99
N32
N80
N60
24
Chord Autonomy

• When new keys are inserted the system is not
  affected. It just finds the appropriate node and
  stores it
• When nodes join or leave, the finger tables must
  be correctly maintained and also some keys must
  be transferred to other nodes
• Also, every key is stored only in one node, which
  means that if that node becomes unavailable the
  key is also unavailable
• This incurs an O(log2N) cost for maintaining the
  finger tables and assuring correctness of the
  system while nodes join/leave the system
• This imply a restricted autonomy of the system
• The only replicated information is (implicitly)
  the finger tables, because each node has to
  maintain its own

25
P-Grid

• Basic characteristics
• Based on building distributed, binary prefix
  trees
• Use of randomized algorithms for constructing the
  access structure, updating the data and
  performing the search
• Scale gracefully, equally for all nodes
• Access structure
• We assume that the index terms are binary
  strings, built from 0s 1s
• The search space is partitioned into intervals
• Every peer takes over responsibility for one
  interval
• As each key corresponds to a path in the binary
  prefix tree the peer is also responsible for one
  path of the search tree
• Each peer stores the peers responsible for the
  other branches of the path for routing
• Search requests are either processed locally or
  forwarded to the peers on the alternative branches

26
P-Grid

• P-Grid construction
• Initially, all peers are responsible for the
  whole search space
• Whenever peers meet, they try to make a
  refinement to the access structure
• they split the search space into two parts and
  each take the responsibility for the one half
• They also store the reference to the other peer
  in order to cover the other part of the search
  space
• The same happens whenever two peers meet, that
  are responsible for the same interval at the same
  level
• To avoid overspecialization of peers, we restrict
  the maximal length of paths that can be
  constructed to a defined maxlength

27
P-Grid
Key intervals Level 0
001
0010
01
0100
100
1001
1011
110
28
P-Grid
queries
Key intervals Level 0
0
1
Key intervals Level 1
01
11
00
10
Key intervals Level 2
001
0010
01
0100
100
1001
1011
110
29
P-Grid Autonomy

• The system implies that peers eventually meet,
  but does not examine how does this occur, i.e. it
  is possible that they never meet
• As many peers can be responsible for the same key
  the general problem is how to find all those
  peers in case of an update
• Proposed solutions
• multiple BFS or DFS searches for a key and
  propagating the update to them
• Creating lists of buddies for each peer (i.e.
  other peers that share the same key) and
  propagate the update to all buddies
• These imply that although the system is
  decentralized and peers does not rely to central
  authorities, the construction and update of the
  access structure may impose some performance
  issues, especially when updating a key

30
P-Grid Autonomy

• When a new node enters the system, assumes that
  he is responsible over the whole prefix namespace
  interval
• When he meets with other nodes they split the
  interval and each maintain a reference to the
  other node
• When a node leaves abruptly, the other nodes have
  incorrect references and as soon as they are
  aware of it they resume responsibility over
  that prefix interval
• The replicated information in this system is the
  multiple references to the same keys and the
  buddies lists (when used) in order to face the
  update problem

31
P2P comparison
32
P2P performance metrics

• Bandwidth
• Storage (replication)
• Processing cost
• Path-length (required hops)
• Quality of Results
• Number of results
• Satisfaction (true if results gt X, false
  otherwise)
• Time to satisfaction

33
Hybrid p2p

• Advantages
• Simple to manage and availability of results -due
  to central indexing
• Less (aggregated) bandwidth consumption
• Small processing cost for peers
• Idle nodes that do not offer resources does not
  downscale systems performance

• Disadvantages
• Does not scale
• Single point of failure
• Great processing cost for server
• Vulnerable to censorship

34
Pure p2p

• Advantages
• Efficiency harnessing unused resources
• Self-organizing
• Robustness and availability through replication
• Anonymity/legal protection/censorship resistant

• Disadvantages
• Difficult to manage and poor results due to lack
  of central indexing
• Bandwidth consuming
• Idle nodes downscale the overall performance
• Higher processing cost for peers

35
Super peers

• Advantages
• Scalable
• Fault tolerant
• Adaptable and self-organizing
• Efficient
• Low path-length

• Disadvantages
• Hard to manage/maintain
• Complex topology, difficult to evaluate its
  metrics (through simulation or trace driven
  analysis)

36
Content-based searching architectures

• Advantages
• Low search cost ( O(logN) )
• Harnessing the content information into queries.
• Good approach for content that can be described
  with simple attributes.
• Less messages per query than a random graph.
• Load balancing.

• Disadvantages
• More restrictions than topology-based
  architectures when nodes join/leave, rehashing
  and content migration needs to be performed.
• A peer needs to know what is looking for, to map
  it to an address.
• Not practical for content described by multiple
  attributes.
• Storage and routing are closely connected

37
Conclusions about p2p systems

• Benefits
• efficiency harnessing unused resources
• Self-organizing
• Sharing cost of ownership
• Robustness and availability through replication
• Anonymity/legal protection

• Challenges
• No authority to enforce behavior
• Cooperation
• Unreliability of individual peers
• Efficiency of distributed operations (absolute
  resources)

• Imposed research issues
• Resource Management
• Security
• Efficient Search

38
Resource Management

• Resource
• Storage/information
• CPU processing
• Bandwidth
• Issues
• fairness
• load balancing

39
Security

• Issues
• Reputation
• Trust
• Accountability
• Information Preservation Quality
• Denial of service attacks
• Problem Detecting and punishing bad behavior

40
Efficiency of Search

• Problem finding needle in haystack
• Efficiency measured in terms of absolute
  resources consumed
• Bandwidth
• Processing cost
• Several factors
• Purity
• Control
• Query expressiveness

41
Project JXTA

• JXTA is a set of protocols which allow peers to
  discover and communicate with each other
• Protocols are defined in terms of XML messages
  exchanged between peers
• JXTA is platform (e.g Windows), language (e.g
  Java) and transport (e.g TCP/IP) independent

42
JXTA Concepts

• Concepts
• Peer - a node that speaks the JXTA protocols
• Peer Group - a collection of cooperating peers
• Message - a datagram containing an envelope,
  protocol headers and bodies
• Pipe - an async communication channel for
  sending/receiving messages
• Advertisement - an XML document that publishes
  the existence of a resource (peer, peer group,
  pipe, service)

43
JXTA Model
44
JXTA Protocols

• Peer Discovery Protocol - used between any peers
  to find other peers, peer groups, or
  advertisements
• Peer Information Protocol - used to learn about
  another peer's properties
• Peer Resolver Protocol - 'foundation protocol'
  for the Peer Discovery Protocol and the Peer
  Information Protocol. Can be used to build other
  protocols as well. Defines send/receive 'generic
  queries' and responses to be sent from one peer
  to another

• Peer Membership Protocol - used to find out
  about, join and leave groups
• Pipe Binding Protocol - used to bind a pipe to an
  actual endpoint
• Peer Endpoint Protocol - used to provide routing
  information for paths between peers (if a direct
  connection is not possible)

45
JXTA Search

• JXTASearch is a framework for searching in
  distributed networks
• A protocol for registration, query and response
• A series of services for interacting via this
  protocol

46
JXTA Search

• Advantages
• Supports very dynamic networks
• Reduce publishing and query response latency
• Centralized control (centralized implementation
  of security, accounting, membership, )

• Disadvantages
• Single point of failure
• Scalability
• Centralized control 

47
Towards a Super-Peer Architecture for SeLene
48
References

• http//www.internet2.edu/presentations/20020131-P2
  P-Kan.htm
• http//softwaredev.earthweb.com/java/article/0,,12
  082_783281,00.html
• http//www.cs.vu.nl/pub/globe/cp2pc/notes/allnotes
  /jxta.overview
• http//wiki.cs.uiuc.edu/cs427/P2PArchitecture
• http//www.stanford.edu/class/cs347/handouts/p2p.p
  pt
• http//cv.uoc.es/grc0_000228_web/Marques/Tesi_JM.
  htm
• http//iew3.technion.ac.il/spektory/098223/presen
  tations/fastTrack.ppt