Gang Wang, Shining Wu, Guodong Wang,

1
.Experimental Study on Neighbor Selection Policy
for Phoenix Network Coordinate System

• Gang Wang, Shining Wu, Guodong Wang,
• Beixing Deng, Xing Li
• Tsinghua University

2
Outline

• Introduction
• Related work
• System design
• Performance evaluation
• Conclusion

3
Introduction

• Network Coordinate System (NCS)
• Distance(Latency) information is very important
  for large scale network applications P2P,
  Overlay Multicast, Overlay routing
• NCS maps the network into a mathematical space

Distance Estimation
Nearest neighbor awareness
others
Network
Mathematical space
4
Introduction

• Network Coordinate System (NCS)
• Network Coordinate System predicts
  End-to-End Links by measurement
  Scalability
• High accuracy and scalability
• Low overhead (Linear)

N
Measured Distance
Estimated Distance
N
5
Introduction

• NC System related Applications
• Google CDN (GNP NCS for sever selection)
• Vuze BitTorrent (NC for neighbor selection)
• SBON(NC for Data query)

6
Introduction

• Problem
• The recently proposed Phoenix NCS is a promising
  solution
• Avoids the Triangle Inequality Violation(TIV)
  problem
• High accuracy and convergence rate
• Robustness over measurement anomalies
• Phoenix NCS suffers disadvantage in certain
  applications such as Overlay Multicast
• The neighbor selection policy for Phoenix is a
  possible solution to this problem

7
Related Work

• Phoenix Network Coordinate System
• Each node will be associated to a Network
  Coordinate (NC)
• Is random neighbor selection is the best?

• For each new node m
• select any M existing hosts randomly
• m measures its RTTs to these M hosts as well as
  retrieves the NCs of these m hosts.
• NC can be calculated and updated periodically.

M
m
8
System Design
Random Policy
Closest Policy
Hybrid Policy

• Random Policy Randomly select M reference
  neighbors
• Closest Policy Choose M closest nodes as
  reference
• Hybrid Policy Mc Closest Nodes and Mr randomly
  selected nodes as reference

9
System Design

• Hybrid intuition
• Distant reference nodes to locate its position
• Nearby reference nodes to adjust it NC to reach
  high accuracy

Closest nodes
Accurate Location
Target node
Distant nodes
10
Performance Evaluation

• Experimental Set up
• Data set and Metrics
• Prediction accuracy
• Application on Overlay Multicast

11
Performance Evaluation

• Experimental Set up
• All of these three systems use 10-dimensional
  coordinates.
• Each node has M reference nodes (M32)
• All of these systems have10 runs on each data set
  and an average result is reported
• For Hybrid Mc 6 (The number of closest
  reference nodes) Mr M Mc 26

12
Performance Evaluation

• Datasets and Metrics
• The PlanetLab data set 226 hosts all over the
  earth
• The King data set1740 Internet DNS servers.
• Distance prediction Relative Error(RE)
• Nearest Neighbor Loss (NNL)
• the difference between the estimated nearest
  host by NCS and the true one

13
Performance Evaluation

• Prediction accuracy
• Mean RE
• Smaller RE indicates higher prediction accuracy
• Hybrid achieves lower RE than Random and Closest
  over both data set

14
Performance Evaluation

• Prediction accuracy
• NNL
• Smaller NNL indicates better ability to select
  nearest host
• Hybrid achieves lower NNL than Random and Closest
  over both data set

15
Performance Evaluation

• Application on Overlay Multicast
• What to do
• Multicast Tree constructed according the
  predicted distance by NCS
• The quality of the multicast tree is evaluated by
  tree cost (the sum of latencies of all tree
  links)
• The tree cost reflects the distance prediction
  accuracy of NCS
• Two kinds of multicast tree ESM MST

16
Performance Evaluation

• Application on Overlay Multicast
• Everage tree cost on PlanetLab and King

ESM-PlanetLab
ESM-King
17
Performance Evaluation

• Application on Overlay Multicast
• Everage tree cost on PlanetLab and King

MST-PlanetLab
MST-King

• Reduce the average tree cost by at least 20

18
Performance Evaluation

• Application on Overlay Multicast
• tree cost change as the tree size increases over
  King

ESM-King
MST-King

• Lower growth rate Lower tree cost

19
Conclusion

• Phoenix with Hybrid neighbor selection policy
  achieves
• Lower distance relative prediction error
• a better accuracy in selecting nearest host
• A better performance in the application of
  Overlay Multicast

20
Any Questions?

• Thank you

21
More NC Research
Simulator http//www.netglyph.org/wanggang/Phoen
ix_NCS_sim.zip Gang Wangs Homepage
http//www.net-glyph.org/wanggang/ More about
NC research in Tsinghua http//www.netglyph.org/
netcoord/