Evaluating GUESS and NonForwarding PeertoPeer Search

1
Evaluating GUESS and Non-Forwarding Peer-to-Peer
Search

• IEEE ICDCS 2004

2
Outline

• Abstract
• Introduction
• GUESS architecture
• Policies
• Experiment / result
• Conclusion

3
Abstract

• Unstractured peer-to-peer network
• Forwarding-based techniques
• Inherent performance drawbacks
• Overhead
• Lack of central control
• In this paper, promote GUESS
• Non-forwarding search mechanism
• Be over an order of magnitude more efficient

4
Introduction

• Peer to Peer (P2P) Networks
• structured
• Unstructured
• Gnutella
• GUESS

5
GUESS protocol-Basic architecture
6
Maintaining State

• Introduction protocol
• P ping Q
• Q cache P, then pong P
• Maintains peers link cache
• Periodically selecting an entry and sending a
  Ping message to the neighbors
• A Pong message contains a small number of IP
  addresses selected from the peers own link cache

7
Query Propagation

• GUESS peer simply iterates through the entries in
  its link cache
• Performs a uni-cast query, or probe

8
Policies types

• Determine how entries in the pong cache are used
  and maintained
• Five types of policies
• QueryProbe
• The order in which peers in the link and query
  caches are probed for queries
• QueryPong
• PingProbe
• The order in which peers in the link cache are
  pinged
• PingPong
• CacheReplacement

9
Policies

• Random (Ran)
• Most Recently Used (MRU)
• Least Recently Used (LRU)
• Most Files Shared (MFS)
• Most Results (MR)
• A peer that has worked well for me will continue
  to work well

10
Experimental Setup

• The primary goal of the GUESS
• Probe traffic
• Proportion of queries that go unsatisfied
• Wasted
• Sent to peer that have already left the network

11
Experimental Setup cont.
12
Result - Cache Size

• As the size of the cache increases, there are
  more peers to probe.

13
Result - Cache Size cont.
Although larger cache sizes result in a larger
number of probes, they do not translate to more
satisfied queries
14
Result Average Query Cost

• GUESS
• 6 unsatisfied
• 99 probes
• MFS
• 8 unsatisfied
• 17 probes
• Gnutella
• 6 (8)unsatisfied
• 1000 (540) probes

15
Result Query Probe Policy
MFS QueryPong and LFS CacheReplacement policies
are the most efficient.
16
Result CacheReplacement Policy

• CacheReplacement policy can reduce cost by over a
  factor of 5

17
Result QueryPong Policy
18
Result Ranked Distribution

• The load is spread more evenly with the
  Random/Random policy

19
Result Limited Capacity

• The number of refused probes increases with
  network size. (overloaded)

20
Result Max Probes Per Second

• The network may be self-sufficient. (more query
  with more servers)

21
Conclusion

• Promote the concept of non-forwarding search
  mechanisms
• Demonstrate how the policies used to determine
  the order of
• probes
• pongs
• cache replacement
• MR policy presents the best tradeoff between
  efficiency and robustness.