PlanetP: Using Gossiping to Build Content Addressable PeertoPeer Information Sharing Communities

1
PlanetP Using Gossiping to Build Content
Addressable Peer-to-Peer Information Sharing
Communities

• Jianzhi Wang

2
Outline

• Introduction
• Architecture
• Performance
• Conclusion and Future Work

3
Introduction

• PlanetP is a content addressable
  publish/subscribe service for unstructured
  peer-to-peer(P2P) communities.
• PlanetP is composed of two components
• an infrastructural gossiping layer
• an approximation of a state-of-the-art text-based
  search and rank algorithm.
• membership directory.
• an extremely compact content index. (Bloom
  Filter)

4
Outline

• Introduction
• Architecture
• Gossiping
• Content Search and Retrieval
• Document Locating
• Local Index
• Global Index
• Ranking algorithm
• Performance
• Conclusion and Future Work

5
Gossiping

• Demers algorithm
• Rumor
• Periodically push change to a randomly chosen
  peer
• Anti-entropy
• Pull data structure summary from a randomly
  chosen peer

6
Gossiping (Cont.)

• Extended Gossiping Algorithm
• Rumor
• Anti-entropy
• High dynamic, less accurate
• Partial anti-entropy
• Advantages
• Reduce cost
• Reduce the propagation time

7
Document Locating

• Uses the global index to derive the set of peers
  that have these terms.
• t p
• Forwards the query to these peers and asks them
  to return URLs for any documents that are
  relevant to the query
• t D

8
Local Index (Bloom Filter)

• Bloom filter is an array of bits used to
  represent a set of strings in our case, the set
  of terms in the peers local index.

Autonomic cat
Hash Functions

1
1
1
1
1
1
1
XML snippet (inversed index)
Yes
No
D1
9
Advantages of Bloom Filter

• It is extremely high compacted
• Update of a peers bloom filter could be spread
  to the whole community in constant time,
  regardless how many changes it includes
• The cost of replicating the global index can be
  reduced by simply decreasing the gossiping rate,
  a tradeoff between propagation time and bandwidth
  usage

10
Global Index

• The global index is replicated on every peer in
  the community by gossiping
• The global index maintains all of the information
  about the documents in the community.
• (peer IP, peer status, bloom filter)

11
Content Ranking Algorithm

• Ranking Document (TFxIDF)
• For a query, the rank of a document is defined as
  the similarity between them
• The weight of a term for a document is determined
  by its appear frequency in this document
• The weight of a term for a query follows the
  inverse of how often it shows up in the entire
  collection (IDF)

12
Content Ranking Algorithm (Cont.)

• Document-Query Similarity
• Question
• How to get the number of documents and the number
  of times that a specific term appears in the
  collection?
• Answer
• We need an approximating algorithm

13
Content Ranking Algorithm (Cont.)

• Mapping terms to peers
• Ranking Peers
• Inverse Peer Frequency (IPF)
• Peer rank for a query
• Select Peers (Q,k)
• Ranking peers for Q
• Contact peers from top to bottom of the ranks
• Each contacted peers return the relevant
  documents with the query-document similarity
  rank.
• Stop contacting when p consecutive peers fail to
  contribute document to the top k ranked documents

14
Outline

• Introduction
• Architecture
• Performance
• Search Efficacy
• Storage Cost
• Gossiping Performance
• Conclusion and Future Work

15
Search Efficiency

• Two accepted information retrieval metrics
• Recall (R)
• The capability of getting the relevant documents
• Precision (P)
• The capability of getting the correct documents
• The characteristic of the collections used
• Number documents
• Number unique terms
• Number of queries
• Documents-to-peers distribution
• Uniform
• Weibull

16
Evaluation Result

• PlanetP has very close performance to the
  centralized implementation.
• The performance of PlanetP is independent from
  how the documents are distributed in the
  community.
• Reducing the percentage of unique term indices
  dos not sacrifice the performance much.

17
Evaluation Results (Cont.)

• PlanetPs adaptive stopping heuristic is critical
  to its performance. (How to determine p)
• Dynamic heuristic allows the search to contact
  more peers when documents are widely distributed.

18
Storage Cost

• PlanetP could be easily scaled to several
  thousand peers

19
Gossiping Performance

• The reliability of PlanetPs gossiping algorithms
• Whether the change can be propagated to all
  on-line nodes or not?
• What kind of changes will be gossiped?
• changes in a Bloom filter
• joining of a new member (peer)
• rejoin of a previous off-line member (peer)
• the leaving of a peer will not be gossiped

20
Gossiping of Bloom Filter Changes

• Simulates the addition of 1000 new terms to some
  peers
• This is a not a small change to the simulation
  collection

21
Gossiping of Bloom Filter Changes (Cont.)

• Gossiping new information is very scalable
• The way to decrease the communication overhead is
  very simple
• PlanetP gossiping significantly overperforms
  ones that only use anti-entropy.

22
Joint of New Members (worst case)

• Joining process is much more bandwidth intensive
  than Bloom filter changes gossiping
• PlanetP does quite well in DSL or higher
  bandwidth network
• PlanetP requires modification for mixed network
  with lower bandwidth

23
Modified Gossiping Algorithms

• The members with lower network bandwidth could be
  the bottleneck of the community

24
Dynamic Operation (Common Case)

• Gossiping with partial anti-entropy performs
  better than the one only with anti-entropy

25
Dynamic Operation (Cont.)

• By applying the modified gossiping algorithm for
  a mixed network, the fast peers of the community
  do not lose much performance.
• The normal case does not need much bandwidth

26

• Introduction
• Architecture
• Performance
• Conclusion and Future Work

27
Conclusion

• PlanetP is a powerful P2P publish/subscribe
  information sharing infrastructure
• PlanetP is able to robustly disseminate the new
  information throughout the community.
• PlanetPs extreme compact global index does not
  effect its ranking accuracy and performance
• PlanetPs does not require high cost in a scale
  of several thousand peers.

28
Future Work

• What is the performance of PlanetP in community
  with the scale of millions or billions of peers?

29
Question?