Characterizing Fault Tolerance of Genetic Algorithms in Desktop Grid Systems

Characterizing Fault Tolerance of GeneticAlgorithms in Desktop Grid Systems

Daniel Lombrana Gonzalez Juan Luis Jimenez LaredoFrancisco Fernandez de Vega Juan Julian Merelo

Guervos

April 8, 2010

D. Lombrana, JJ. Jimenez, F. Fernandez, JJ. Merelo Evocop 2010

Outline

1 Introduction

2 Motivation

3 Methodology

4 Experiments and Results

5 Conclusions


Introduction


Introduction

Parallel Genetic Algorithms (PGA)

Sometimes Evolutionary Algorithms (EAs) require largeexecution times.One solution is to use:

Parallel Computing andDistributed Platforms.


Introduction

Parallel algorithms can be run in


Introduction

Parallel algorithms can be run in


Introduction

Failures in distributed platforms

Distributed platforms are prone to errors.Failures are expected events rather than catastrophicexceptions.


Introduction

Fault Tolerance

Fault Toleranceis the ability of a system to behave in a well-defined manneronce a failure occurs.


Introduction

Fault Tolerance

Different techniques have been developed to cope with failures:

Redundancy,S. Ghosh. Distributed systems: an algorithmic approach. Chapman & Hall/CRC, 2006.

Checkpointing,E. Elnozahy, L. Alvisi, Y. Wang, and D. Johnson. A survey of rollback-recovery protocols inmessage-passing systems. ACM Computing Surveys (CSUR), 34(3):375–408, 2002.

Rejuvenation frameworks,A. T. Tai and K. S. Tso. A performability-oriented software rejuvenation framework for distributedapplications. In DSN ’05, pages 570–579, Washington, DC, USA, 2005. IEEE Computer Society.

etc.


Introduction

Fault Tolerance

The use of a fault tolerance technique mandates that:the application has to be modified, and eventhe parallel algorithm.

Thus, this modification can represent a heavy burden for thedeveloper.


Motivation


Motivation

Parallel EAs and Fault Tolerance

To the best of our knowledgethere has been little research about the fault tolerance featuresof PEAs in general and of PGA applications in particular.


Motivation

Previous Works

We firstly studied the Fault-Tolerance nature of ParallelGenetic Programming (PGP) on:

Real World Desktop Grid Systems.Concluding that PGP is fault-tolerant by default.

Daniel Lombrana Gonzalez, Francisco Fernandez de Vega, and Henri Casanova.Characterizing fault tolerance in genetic programming.Future Generation Computer Systems, 2010.DOI: 10.1016/j.future.2010.02.006.Daniel Lombrana Gonzalez, Francisco Fernandez de Vega, and Henri Casanova.Characterizing fault tolerance in genetic programming.In Workshop on Bio-Inspired Algorithms for Distributed Systems,pages 1–10. Barcelona, Spain, 2009. ISBN 978-1-60558-564-2.


Motivation

Proposal

Based on this insightThis work builds on top of the previous ones, and extends thestudy of fault-tolerance in EAs to PGAs, using the samemethodology.


Methodology


Methodology

Master-Worker


Methodology

Desktop Grid platforms (DGs)

DGs exhibit large numbers of failures.DGs failure behavior has been studied in literature.DGs are low-cost when compared to clusters ofcomparable scale.And, PGA applications are loosely coupled and thuswell-suited to DGs.


Methodology

Desktop Grid Platforms

DGs are very promising for PGA applications, andtheir high failure rate make them a great test case forstudying and characterizing the fault tolerance abilities ofPGA.


Methodology

Experiments

In order to characterize the fault-tolerant nature of PGA werun two kind of experiments:

a failure-free environment, andreplaying and simulating failure traces from real-world DGplatforms.


Methodology

DG traces

We perform simulations of DG platforms and of hostavailability based on three real-world traces:

entrfin,ucb,xwtr.


Methodology

DG traces

Trace Hosts Venue TimeEntrfin 275 San Diego 1.0 monthsUcb 85 UC Berkeley 1.5 monthsXwtr 100 Univeriste Paris-Sud 1.0 months


Methodology

Using the traces

We consider two cases:hosts that become unavailable never become availableagain (worst case assumption),and the complete host-churn (unavailable hosts can bere-acquired afterwards).

For two different days of each trace.


Methodology

Host availability for 1 day of the ucb trace

0

5

10

15

20

25

0 50 100 150 200 250 300

Com

pute

rs

Time StepOriginal Trace Trace without return


Experiments and Results



Problems

We conduct experiments with a 3-trap instance:

trap(u(−→x )) =

{ az (z − u(

−→x )), if u(−→x ) ≤ z

bl−z (u(

−→x )− z), otherwise(1)



GA Parameters for 3-Trap instance

Trap instanceSize of sub-function (k ) 3

Number of sub-functions (m) 10Individual length (L) 30

GA settingsGA GGA

Population size 3000Selection of Parents Binary Tournament

Recombination Uniform crossover, pc = 1.0Mutation Bit-Flip mutation, pm = 1

L



Population size vs. generation

0

500

1000

1500

2000

2500

3000

3500

4000

0 10 20 30 40 50

0

25

50

75

100

Indiv

iduals

% o

f L

oss

Generations

entrfin 1entrfin 2

ucb 1ucb 2

xwtr 1xwtr 2



Obtained Fitness for 3-Trap Day1

Error Free fitness = 23.56

Trace Fitness Wilcoxon Test Significantly different?Entrfin 23.30 W = 6093, p-value = 0.002688 yesEntrfin 10% 23.47 W = 5408.5, p-value = 0.2535 noEntrfin 20% 23.48 W = 5360, p-value = 0.3137 noEntrfin 30% 23.49 W = 5283.5, p-value = 0.4271 noEntrfin 40% 23.57 W = 4923.5, p-value = 0.8286 noEntrfin 50% 23.59 W = 4910.5, p-value = 0.7994 no

Ucb 23.22 W = 6453, p-value = 6.877e-05 yesUcb 10% 23.27 W = 6098.5, p-value = 0.002753 yesUcb 20% 23.37 W = 5837.5, p-value = 0.02051 yesUcb 30% 23.40 W = 5664, p-value = 0.06588 noUcb 40% 23.51 W = 5186.5, p-value = 0.6004 noUcb 50% 23.42 W = 5623, p-value = 0.08335 no

Xwtr 23.56 W = 5056, p-value = 0.8748 noXwtr 10% 23.57 W = 4923.5, p-value = 0.8286 noXwtr 20% 23.68 W = 4474, p-value = 0.1245 noXwtr 30% 23.73 W = 4259.5, p-value = 0.02812 yesXwtr 40% 23.68 W = 4502, p-value = 0.1466 noXwtr 50% 23.71 W = 4356.5, p-value = 0.05817 no



Obtained fitness for 3-Trap Day2

Error Free fitness = 23.56

Trace Fitness Wilcoxon Test Significantly different?Entrfin 23.57 W = 4979.5, p-value = 0.9546 noEntrfin 10% 23.69 W = 4397.5, p-value = 0.07682 noEntrfin 20% 23.67 W = 4522.5, p-value = 0.1645 noEntrfin 30% 23.70 W = 4405, p-value = 0.08086 noEntrfin 40% 23.69 W = 4453.5, p-value = 0.11 noEntrfin 50% 23.75 W = 4162.5, p-value = 0.01234 yes

Ucb 23.09 W = 6672.5, p-value = 7.486e-06 yesUcb 10% 23.12 W = 6826, p-value = 6.647e-07 yesUcb 20% 23.14 W = 6654, p-value = 7.223e-06 yesUcb 30% 23.26 W = 6371, p-value = 0.0001507 yesUcb 40% 23.37 W = 5893.5, p-value = 0.01316 yesUcb 50% 23.32 W = 6108, p-value = 0.002166 yes

Xwtr 23.60 W = 4806, p-value = 0.5791 noXwtr 10% 23.62 W = 4765, p-value = 0.5002 noXwtr 20% 23.69 W = 4453.5, p-value = 0.11 noXwtr 30% 23.60 W = 4806, p-value = 0.5791 noXwtr 40% 23.63 W = 4688.5, p-value = 0.3695 noXwtr 50% 23.77 W = 4065.5, p-value = 0.004877 yes



Obtained fitness with host-churn

Table: Day1Error Free fitness = 23.56

Trace Fitness Wilcoxon Test Significantly different?Entrfin 23.52 W = W = 5222, p-value = 0.5322 noUcb 21.31 W = 9708.5, p-value < 2.2e-16 yesXwtr 23.64 W = 4640, p-value = 0.2982 no

Table: Day2Error Free fitness = 23.56

Trace Fitness Wilcoxon Test Significantly different?Entrfin 23.58 W = 4931, p-value = 0.8452 noUcb 23.03 W = 7038.5, p-value = 4.588e-08 yesXwtr 23.7 W = 4405, p-value = 0.08086 no


Conclusions


Conclusions

Summary of Results

PGA applications are fault-tolerant by nature in DGplatforms.PGA features the well-known fault-tolerant techniqueknown as graceful degradation in DG platforms.We provided a new method to mitigate the effect of failuresby increasing the initial population.


Conclusions

Conclusions

We have studied and characterized the behavior of PGAapplications running in distributed platforms with highfailure rates.We have tested the PGA fault-tolerance using threereal-world DG traces.Our main conclusion is that PGA inherently providesgraceful degradation.


Conclusions

Questions

[email protected]@geneura.ugr.es

[email protected]@geneura.ugr.es

Icons from Tango Desktop project and Gnome Desktop (Creative Commons & GPL License)


Characterizing Fault Tolerance of Genetic Algorithms in Desktop Grid Systems

Education