LocationAware Topology Matching in P2P Systems

1
Location-Aware Topology Matching in P2P Systems

• IEEE INFOCOM 2004

2
Outline

• Abstract
• Introduction
• Location-aware Topology Matching (LTM)
• Simulation
• Conclusion and Future Work

3
Abstract

• Peers randomly choosing logical neighbor without
  any knowledge about underlying physical topology.
• Topology mismatching
• P2P overlay network
• Physical underlying network

4
Abstract cont.

• Location-aware topology matching (LTM)
• Disconnecting low productive connections
• Choosing physically closer nodes as logical
  neighbors
• Still retaining the search scope and reducting
  response time for queries
• Scalable and completely distributed (does not
  require any global knowledge of the whole overlay
  network)

5
Mismatching Problem

• The same message may traverse the same physical
  link multiple times
• Multiple paths that are merged to the same peer
• The same query message may traverse the same
  logical link twice

6
Introduction

• Gnutella (Decentralized unstructured P2P)
• Generate 330 TB/month with 50,000 nodes
• Inefficient overlay topology
• Blind flooding
• Location-aware topology matching (LTM) scheme
• Issues a detector in a small region
• Record relative delay information

7
P2P Overlay
8
Unnecessary Traffic

• Unnecessary traffic
• A2 ltgt A3
• B1 ltgt B3

9
Topology Mismatching
10
Three Main Operations of LTM

• TTL2-detector flooding
• Low productive connection cutting
• Source peer probing

11
TTL2-Detector Flooding
12
Low Productive Connection Cutting Case 1
13
Low Productive Connection Cutting Case 2
S
N2
N1
N2P is in Ps Will-Cut List if Time(N2P) gt
Time(SN2),
N1P is in Ps Will-Cut List if Time(N1P) gt
Time(SN2),
P
14
Low Productive Connection Cutting Case 3

• Select
• S lt-gt P
• S lt-gt N1 lt-gt P or
• S lt-gt N2 lt-gt P randomly

15
Source Peer Probing

• Not receive S lt-gt P
• Create a new path
• S-P
• Probing
• Low productive connection cutting

16
An Example of LTM
17
Simulation Setup

• Topology generation
• 5 physical topologies with 22000 nodes
• Overlay networks with 2000 to 8000 nodes
• Avr. Number of neighbors 4 to 10
• Flooding search simulation
• Flooding search with BFS (i.e., Gnutella)
• Dynamic P2P environment
• 0.3 query /min
• Mean lifetime 10 min

18
Simulation Metrics

• Traffic Cost
• consumed BW
• Search Scope
• Number of peers reached
• Average neighbor distance
• Delay time between source and neighbors
• Response time
• query response time

19
Traffic Cost vs. Search Scpoe

• Static environment
• With a given traffic cost, LTM will increase its
  search scope

20
Traffic Reduction vs. Optimization Step

• Static environment
• With 8,000 peers
• The traffic cost reduction reaches to a threshold
  after 2-3 steps

21
Average Neighbor Distance vs. Optimization Step

• Static environment
• One-step LTM optimization reduces average
  neighbor distance by about 55, more steps to
  around 65

22
Average Response Time vs. Optimization Step

• Static environment
• LTM can shorten the query response time by about
  62

23
Effectiveness of Will-Cut List

• Dynamic environment
• Response loss problem
• The query success rate in dynamic environments is
  decreased by about 5
• The query success rate is decreased by 3040
  without W-C

24
Effectiveness of Cut List

• Dynamic environment
• The use of the cut list reduces traffic overhead
  by about 50

25
Total Traffic vs. LTM Frequency

• Dynamic environment
• LTM2 can reduct 75 traffic cost

26
Response Time vs. LTM Frequency

• Dynamic environment
• LTM2 can reduct 65 response time

27
Optimal LTM Frequency vs. Average Peer Lifetime

• LTM can be conducted less frequently if peer
  average lifetime is longer

28
Optimal LTM Frequency vs. Average Query Frequency

• LTM should be conducted more frequently if more
  queries are issued

29
Traffic Cost of Four Schemes

• Combining LTM and query index caching
• The traffic cost is reduced by about 10 times
  without shrinking the search scope

30
Average Response Time of Four Schemes

• Combining LTM and query index caching
• The average query response time is reduced by
  about 7 times

31
Conclusion

• LTM
• Location of near neighbors
• Match overlay network with physical topology
• To improve search efficiency and scope

32
Future Work

• To integrat LTM with other existing advanced
  search approaches
• To deploy and test an LTM based on Gnutella in
  PlanetLab