How to Select a Good Alternate Path in Large PeertoPeer Systems

1
How to Select a Good Alternate Path in
LargePeer-to-Peer Systems?

• Teng Fei, Shu Tao, Lixin Gao, Roch Guerin
• Dept. Elec. Comput. Eng.,
• U. Massachusetts, Amherst
• Email tfei, lgao_at_ecs.umass.edu
• Dept. Elec. Sys. Eng., U. Pennsylvania
• Email shutao_at_seas.upenn.edu, guerin_at_ee.upenn.edu
• 
  Infocom 2006

2
Outline

• Introduction
• Related Work
• The Earliest-Divergence Rule
• Evaluation
• Measurement
• Conclusion

3
Introduction

• Background
• Performance degradation occur in Internet
• Application path switching is helpful when
  multiple paths exist.
• Peer-to-Peer network is an attractive platform
• Heterogeneous network
• Minimum infrastructure support requirement

4
Selecting an alternate path

• From thousands of IP address
• Choose one as relay
• Uncorrelated paths

5
Two challenges and approach

• There exist a large numbers of choices depends on
  the target destination
• The relay node may vary over time
• The approach is
• Most disjoint with the default path
• AS-level

6
Methodologies

• PlanetLab testbed
• Oregon Route Views and RIPE
• Two performance degradation metrics
• Delay degradation
• Loss path

7
Related Work

• RON and Tapestry correlation and may performance
  degradation at the same time as the default path
• A routing underlay dedicated to topology probing
• Use traceroute to obtain the IP-level path and
  latency Information, to estimate the most
  disjointness path
• Random choose k and picking the best one

8
The Earliest-Divergence Rule

• Disjoint overlay path
• Overall probability of performance degradation
  is
• Pdegrad PX1X2 PX2PX1X2 (1)
• X1 denote the event path 1 experiences poor
  performance
• X2 denote the event path 2 experiences poor
  performance
• Monitor the performance variations on all
  possible candidates simultaneously not a
  scalable solution

9

• Using path disjointness as a selection criterion
  may yield paths with significantly lower
  performance correlation
• i.e. PAB(x0,x1,,xn), PAO(y0,y1,,yn),
  POB(z0,z1,,zn).
• S as the overlap of PAB with PAO or POB,
  minimizeS.

10
AS-level or IP-level ?

• IP-level
• The Traceroute routine can be challenging for
  many source-destination pairs to obtain reliably
  and accurately IP-level path information
• The results of traceroute from the source node
  only, is also difficult.
• AS-level
• Also use traceroute for identifying the set of
  ASes
• Accurately infer the ASes even if return
  incomplete IP-level path information
• IP-level paths convert into AS-level paths

11
Comparison AS-level and IP-level
12
The Earliest-Divergence Rule

• In a system, N nodes, communicating node pairs
  O(NN), relay nodes N-2,O(NN) comparisons
• The earliest-divergence rule
• Select paths diverge at the earliest AS
• Among pre-select path set, select one

13

• PAB(x0,x1,,xi,,xn), PAO(x0,x1,,xi),
  POB(xi,xi1,,xn)
• Note x0,x1,,xn and z0,z1,,zj all represent ASes
• Link xi zj can be one of three types of inter-AS
  links
• Peering, customer-provider, or provider-customer
  

14
Selecting the Best Candidates

• Extend select nodes far from the direct path.
• Assume source node A knows the round-trip of
  delay nodes selected by ED rule (denoted as DAO)
• DAB denote the round-trip delay between source
  node and destination node
• Assume a upper bound denote as ß
• m (mlt10)

15
Evaluation

• Parallel paths diverging early have smaller
  overall overlapping
• Topology based study
• AS path data obtained from Oregon/RIPE and Planet
  lab
• Physically less overlapped paths have less
  correlated performance
• Measurement based study
• Active UDP probe using Planetlab nodes

16
(No Transcript)
17
Effect of Reducing Relay Node Numbers

• 
  PlanetLab

18
Effect of Reducing Relay Node Numbers
19
Effect of Selecting Disjoint Paths

• Retain good ones

20
Effect of Selecting Disjoint Paths

• Reject bad ones

21
Effect of Selecting Disjoint Paths

• Increases the likelihood of selecting good ones

22
Effect of Selecting Disjoint Paths

• Increases the likelihood of selecting good ones

23
Comparison with Random-4
24
Measurement

• Methodologies
• Use planetLab testbed test correlation
• Upenn and UMN, Abilene, commercial provider
  networks (Cogent for Upenn, Wiltel for UMN)

25
Performance Correlation Metrics

• Two main performance parameters
• Delay variation degradation
• Di denotes the round-trip delay, mean ,
  standard deviation ,k is a constant
• Using window size L, a threshold a

26
Performance Correlation Metrics

• Two main performance parameters
• Loss variation degradation
• Samples, average loss rate certain time period
• Window size L, threshold ?

27
Comparison Delay with Loss
28
Performance Evaluation of EDR

• Better than random in most cases
• Significantly better in many cases

29
Comparison With Other Schemes
30

• Compare our scheme with general random-k scheme

31
Case

• 9 PlanetLab nodes as potential relays to build
  alternate paths
• 1default path, 9 alternate paths, 20ms, 48hours
• Assuming path switching is used between the
  direct path and a selected overlay path
• Focus on end-to-end losses, most unambiguous
• Compute average loss rate for all paths every 10
  seconds(500 samples)
• If average loss rate gt1, the path in a loss
  period

32
Nodes information
33
Results

34
Conclusions

• Alternate path selection in a large P2P
  environment
• A good alternate path
• Most disjoint, AS-level
• ED rule to minimize overhead and ensure
  scalability
• Effectiveness, topology data and measurement
  traces
• Open question P2P system and underlying network
  grow, how the proposed heuristic would change