The Case for a Hybrid P2P Search Infrastructure

1
The Case for a Hybrid P2P Search Infrastructure

• Boon Thau Loo
• UC Berkeley
• (with Ryan Huebsch,
• Joseph Hellerstein, Ion Stoica)

IPTPS 2004 (26 Feb 2004)
2
Our Goal

• P2P Search - What is the best design?
• Target Environment
• Items stored in P2P network, queried by keywords
• Replicas of items follow a long-tailed
  distribution
• Popular items at head of distribution
• Rare items at tail of distribution
• Typical of P2P file-sharing environments

3
Design Choices

• Unstructured networks
• Queries are flooded for bounded number of hops
• E.g. Gnutella and Kazaa
• No guarantees on recall
• Structured Networks
• Inverted Indexes on Distributed Hash tables
  (DHTs)
• Inverted Lists indexed by keyword.
• ltBritney Doc4, Doc5, Doc6, Doc9, Doc11, gt
• ltQuantum Doc1, Doc3gt
• Query execution
• Routes query to all sites hosting keyword in
  query
• Intersection of multiple Inverted Lists.
• Guarantees perfect recall (absence of network
  failures)

4
Our Proposal Hybrid Solution
Flood-based Network (All items)
DHT (Index Rare Items)
5
Gnutella Network
Oct 2003 Crawl

• Crawl of Gnutella Network
• Based on multiple crawlers from 30 vantage points
  on PlanetLab
• ? 100,000 nodes, 20 million files
• Ultrapeer-based Topology
• Queries flooded among ultrapeers
• Leaf nodes shielded from query traffic

Ultrapeer nodes
Leaf nodes
gt100 Files
0 Files
0-100 Files
6
Gnutella Measurements

• Quality of Searches
• Recall ( of all relevant items retrieved)
• Response Time (Latency) to 1st result
• Software utilized
• Modified the LimeWire Gnutella Client
• Run as leaf or ultrapeer
• Sniff Gnutella traffic
• Inject queries and gather results

7
Gnutella Search Quality

• Reissue Gnutella queries
• 30 LimeWire Ultrapeers on PlanetLab
• 700 Gnutella queries at 3 different times
• Compute Query Recall
• Each query issued simultaneously from 30
  ultrapeers
• Union of results from 30 ultrapeers
• Union-of-30 is our approximation of perfect
  answer

8
Queries with Small Result Sets
9
Result Size CDF
10
Result Size CDF
Single Query
Union-of-30 Query
Large fraction of queries return few or no
results even when they exist
11
Query Latency
Queries that return few results have poor
response times
12
Summary of Measurements

• Searching on Gnutella
• Highly effective for popular items
• Less effective for rare items
• Significant opportunity to do better
• Large fraction of queries return few or no
  results even when they exist
• Bad response times for queries on rare items

13
DHT-based Search

• Advantages
• Avoid flooding query in network
• Guarantee recall (critical for small result sets)
• Disadvantages
• Constructing inverted lists is costly
• So is intersecting inverted lists at query time
• Back-of-envelope calculations shows infeasibility
  for large datasets (IPTPS 03)
• Feasible for querying rare items
• Queries over rare items ship 7x fewer inverted
  list entries compared to the average query

14
Hybrid Search

• Hybrid Best of both worlds
• Flooding techniques for searching popular items
• DHT for rare items
• Identifying rare items
• Query snooping
• Items from previous queries that return few
  results.
• Other techniques (ongoing work)
• Term Frequency Statistics (single term, pairs)
• Sample items on neighboring nodes

15
PlanetLab Deployment
L
L
Horizon of P1
U2
L
L
P2
P1
U1
L
P3
L
L
L
L
L
L
Gnutella Leaf
Gnutella Ultrapeer
U
Gnutella links
Hybrid Ultrapeer (PIER Gnutella)
P
PIER links
16
Results

• Improved Response Time
• PIER returns first result in 10 seconds
• 40 seconds in aggregate including 30 seconds
  timeout
• Gnutella queries returns first result in 65
  seconds
• 25 seconds (38) reduction in latency

17
Results

• Improved Recall 18 reduction in the number of
  queries that eventually receive no results from
  Gnutella
• Lots of room for improvement 66 potential
  reduction based on Union-of-30 query

18
Conclusion

• Hybrid Search Infrastructure
• Flooding for popular items, DHT for rare items
• Simple idea that works in practice

19
Questions?
20
Backup Slides
21
Gnutella Optimizations

• Ultrapeers and Leaf nodes
• Dynamic Querying
• More aggressive flooding for queries that return
  few results
• Query Routing Protocol (QRP) tables
• Bloom-filter-like structures
• Cache at neighboring nodes for directed flooding.

22
Hybrid Implementation

• PIER
• Fully decentralized relational query processor
  over DHTs
• Supports Select, Project, Joins, Group
  By/Aggregation.
• Support for keyword searching
• Intersection of Posting Lists -gt Joins.
• Hybrid Ultrapeer Node
• LimeWire Ultrapeer (Gnutella)
• PIER (DHT overlay)

23
For more details

• http//pier.cs.berkeley.edu
• LimeWire. http//www.limewire.org/
• Gnutella Developer Forum
• http//groups.yahoo.com/group/the_gdf

24
Handling Network Churn

• Only ultrapeers participate in DHT overlay.
  Ultrapeers publish items for leaves.
• Avoid publishing items from short-lived nodes.

25
Hybrid Implementation
26
Search Engine using PIER
Symmetric Hash Join
Join Index
Item(docID, filename, location, filesize,
) Posting(keyword, docID) using keyword as DHT
storage key.
27
Related Work

• Most file downloads are for popular items
  (SIGCOMM 2003)
• Downloads are only half the story.
• Rare items in aggregate can be substantial.
• Build Gnutella on a DHT? (HOTNETS 2003)
• Improved performance for floods and random walks
• Reduced maintenance overheads.

28
Outline

• Background
• P2P Search Today
• Unstructured Networks (Gnutella)
• Structured Networks or DHTs
• The Case for Hybrid
• Hybrid Design and Implementation
• Evaluation on PlanetLab