Boon Thau Loo1

1
Analysis and Improvements of P2P Search Networks
Boon Thau Loo1 PIER (Joseph Hellerstein1,2, Ryan
Huebsch1, Sam Mardanbeigi1, Sean Rhea1, Timothy
Roscoe2, Scott Shenker1,3, Ion Stoica1) UC
Berkeley1 Intel Research Berkeley2 ICSI3
Gnutellas Search Characteristics
Introduction

• - Unstructured Networks (Gnutella, KaZaA)
• 2-level hierarchy Ultrapeers, Peers
• Flooding-based
• Effective for highly replicated items.
• Inverted index over Distributed Hash Tables
  (DHTs)
• Effective for rare items
• Guaranteed recall but may use significant
  bandwidth.
• Our Goals
• Analysis of Gnutella.
• Exploring a hybrid search network.
• Querying Gnutella at scale.

• Observe queries from 20 ultrapeers on Planetlab.
• Dynamic flooding the fewer the query results,
  the further it travels in the network. TTL 3 to
  7.

• 30-40 of queries return less than 10 results.
  Around 10 of queries return no results. Average
  query recall 22.
• Can we use DHTs to improve query recall?

• 1.5 hrs crawl. 105715 nodes, 17766 ultrapeers
• Ultrapeer degree 75 leaves / 6 ultrapeers, 30
  leaves/ 32 ultrapeers

Hybrid Search Network

• DHT-based Boolean Keyword Searching
• PIER A P2P Relational Query Processor over
  DHTs
• For each file, publish FileInformation(docID,
  location, filename, ) and InvertedIndex(keyword,
  docID)
• Query Evaluation Strategies

• - Pier-Gnutella Gateway
• PIER nodes participating in Gnutella network as
  ultrapeers.
• Selectively publishes observed horizon of
  ultrapeer (files in leaf nodes and query
  results).
• Uses PIER for searching rare files. Rely on
  Gnutella for searching highly-replicated files.

• Live deployment on Planetlab
• 50 PIER-Gnutella ultrapeers.
• Publish query results of queries with fewer than
  20 results.
• Execute leaf queries that return no results
  after 60 seconds.
• 20 reduction in queries with no results.

Querying Gnutella with PIER
Demo K-Horizon Gnutella Crawl from multiple
ultrapeers

• PIER Recursive Queries
• Output of one operator tree is placed in a
  storage mechanism (DHT) and is the input to
  either that tree or another tree.
• Useful for querying network graphs (web
  hyperlinks, network topologies).

• - Schema Node(src, hop, ultrapeer, files),
  Link(src, dst)
• Avoid redundant work when horizon of ultrapeers
  overlaps.
• Increased aggregate bandwidth and throughput.
• Potential to exploit geographic proximity.