An Evaluation and Comparison of Current PeertoPeer FullText Keyword Search Techniques

1
An Evaluation and Comparison of Current
Peer-to-Peer Full-Text Keyword Search Techniques

• Ming Zhong, Justin Moore, Kai Shen
• University of Rochester
• Amy Murphy
• University of Lugano

2
Current P2P Full-Text Keyword Search Techniques

• Document-based partitioning Gnutella, KaZaA
• Keyword-based partitioning Bhattacharjee03,
  Gnawali02, Reynold03, Suel03
• Hybrid indexing Tang04
• Semantic search Ganesan04, Li04, Tang03
• There is no comprehensive quantitative
  performance evaluation and comparison of these
  techniques! (Li03)

3
Our Work

• Quantitative performance evaluation results on
  real, large datasets (3.8 million web pages from
  www.dmoz.org and 6.8 million web queries from
  AskJeeves).
• Performance metrics
• Total storage consumption
• Communication overhead
• Search latency
• Search quality
• Performance evaluation results linearly projected
  to 1 billion web pages.

4
Evaluation Setup

• 8-byte page ID (the MD5 of page URL).
• Each entry in inverted lists has 10 bytes (8 byte
  ID 2 byte term frequency).
• Each query only retrieves the top 20 most related
  page IDs, ranked by TF.IDF term weighting scheme.
  
• The underlying topology Chord.
• Network settings.
• p2p overlay link latency 40ms.
• The maximum per-query bandwidth consumption
  1.5Mbps.
• The maximum whole-system bandwidth consumption
  1Gbps 0.26 of US Internet backbone bandwidth
  in 2002.

5
Document-Based Partitioning

• Each node holds a partition of documents. A query
  is broadcast to all nodes and each node returns
  top 20 most relevant documents.
• Tree-based (log n depth) message broadcast and
  aggregation.
• Total storage consumption 4.24 GB for our data
  set
• Total communication cost 300n bytes for our data
  set (n, the network size).
• Search latency 0.08 X log(n) secs.
• Search Quality

6
Baseline Keyword-Based Partitioning

• Each node holds the inverted lists of some
  keywords (randomly distributed). To save the
  communication overhead of the inverted list
  intersection, a k-word query visit k peers in the
  ascending order of the inverted list sizes.
• NO quality degradation for baseline keyword-based
  partitioning.
• Equal storage consumption compared with doc-based
  partitioning.
• Average comm. cost per query 96.61 KB.
• Max comm. Cost per query 18.65 MB
• Search latencylt0.14 X log(n) 0.52 secs.

7
Improved Keyword-Based Partitioning (I)
The average comm. cost is reduced to 0.137 times
that of the baseline keyword-based partitioning.
8
Improved Keyword-Based Partitioning (II)
san ca diego francisco puerto tx austin rico
antonio earthquake jose juan vallarta lucas rican
luis cabo fransisco bernardino
9
Improved Keyword-Based Partitioning (III)
10
Hybrid Indexing

• Each page ID in inverted lists has some metadata
  about the corresponding page.
• A naive approach each page ID has the complete
  term list of the corresponding page. Too much
  storage consumption!!!
• Tangs approach An inverted list of x only
  contains those pages that has x as its top terms.
  Query expansion also used.
• Storage consumption 13 times that of
  keyword-based partitioning.
• Latency lt 0.16 X DHT diameter sec.
• Comm. cost per query 7.5 KB.
• Search quality

11
Semantic Search

• Documents and queries are mapped into points in a
  semantic space and hence keyword search becomes
  nearest point search.
• pSearch LSI, rolling index.
• Storage consumption 30.06 GB, 7.09 times that of
  keyword-based partitioning if p 10.
• Comm. cost per query 1.29 MB.
• (P 10, K 160)
• Search latency DHT comm. latency
• network transmission time
• Search quality

12
Scaled Performance on 109 pages and n peers
13
Examples of Choosing Techniques

• Application 1 105 peers, 107 pages.
• Use keyword-based partitioning
• 29.64 MB storage per peer
• lt12.22 KB communication overhead per query
• lt2.4 sec search latency
• 100 search quality
• Application 2 103 peers, 109 pages.
• Use document-based partitioning
• 1.14 GB per peer.
• 300 KB communication overhead per query
• 0.8 sec search latency
• 75 95 search quality