PACT-08-workshop-churn-p2p-ea

ADDRESSING CHURN IN A PEER-TO-PEER EVOLUTIONARY ALGORITHM

J.L.J. Laredo, P.A. Castillo, A.M. Mora, C. Fernandes,

J.J. Merelo

Dept. of Architecture and Computer Technology

University of Granada

Scope

P2P systems are large networks of volatile

resources, the dynamics are known as

CHURN.

1. P2P EA viable?

2. Is it a P2P EA FAULT TOLERANT under

CHURN ?

Outline

Introduction

P2P EAs Issues

Experimental Setup

Results

Conclusions

Introduction

P2P EAs Issues

Experimental Setup

Results

Conclusions

P2P EA

Evolvable Agent Model

Outline

Introduction

P2P EAs Issues

Experimental Setup

Results

Conclusions

Decentralization

Scalability

Fault Tolerance

Outline

Outline

Introduction

P2P EAs Issues

Experimental Setup

Results

Conclusions

Modelling Churn

MMDP

Some Considerations

Introduction

P2P EAs Issues

Experimental Setup

Results

Conclusions

Outline

Introduction

P2P EAs Issues

Experimental Setup

Results

Conclusions

Outline

Overlay Network

Physical Network

Introduction

P2P EA

Evolvable Agent Model

• Volunteer Computing

• Convergence P2P GRID

• Application Level Networks

•Dynamic neighborhood

(i.e. Small World)

Evolvable Agent

St Initialize

DO

Sols Selection

St +1 Recombination(Sols,Pc )

Evaluation (St +1)

If St +1 better than St

St St + 1

Introduction

P2P EA

Evolvable Agent Model

P2P EAs Issues

Decentralization

Scalability

Fault Tolerance

P2P EAs Issues

Decentralization

Scalability

Fault Tolerance

P2P EAs Issues

Decentralization

Scalability

Fault Tolerance

• Cuadratic number of edges

• High clustering coefficient

• Logarithmic number of edges

• High clustering coefficient

Peer Session length Inter-arrival

1 2

2 1

Decentralization

Scalability

Fault Tolerance

P2P EAs Issues

T0 T1 T2

T3T4T5

Peer Session length Inter-arrival

1 2

2 1

Decentralization

Scalability

Fault Tolerance

P2P EAs Issues

T0 T1 T2

T3T4T5

Modelling Churn

MMDP

Some Considerations

Experimental Setup

Peer Session length Inter-arrival

1 2

2 1

•Stutzbach and Rejaie

•Weibull Distribution

• s = 0.40

• λ= 1, 5, 10, 50

sUX1

))ln((

CHURN

Modelling Churn

MMDP

Some Considerations

Experimental Setup

1 0 1 0 0 0 1 1 1 1 0 1 1 0 1 1 0 0

1, 2… k

Massively Multimodal Deceptive Problem

Instances k = 2, 4, 8, 16, 32, 64

GA EvAg Newscast

Population

Size

Bisection0.98 of SR

Experimental Setup

EA Population Size Є (40, 400)

Wrong assumption!!!

• Population Size depends on the problem instance complexity

• Larger Instances require larger population sizes

• Practitioners tackle large instances with few individuals due to a

lack in resources

• We have many resources in P2P systems

Modelling Churn

MMDP

Some Considerations

Results

Results

Results

k=64

λ=1

0.4% of the

initial population

Conclusions

Agent-based approach for dEA on P2P

Algorithmically viable, 0.98 SR

Massively scalable

Resilient to churn

THANK YOU!