Agenda Motivation and Related Works ALS The pALS framework - - PDF document

19/04/2010 1

pALS: an Object Oriented Framework for Developing Parallel Cooperative

Harold Castro, Andrés Bernal COMIT (COMunications and Information Technology) Systems and Computing Engineering Department Universidad de los Andes

Developing Parallel Cooperative Metaheuristics

Universidad de los Andes Bogotá, Colombia

Agenda

 Motivation and Related Works  ALS  The pALS framework  Achievements and Results  Conclusions

NIDISC'10, Atlanta, April 19-23 2

19/04/2010 2

Motivation

 Research on Optimization (COPA Research Group)  JG2A: Java Framework for Genetic Algorithms on a Grid

 Evolution for JGA

 Instances parallelization  Individual parallel evaluation

 Uses standard tools: Condor, Globus, Java  Allows different models: Master/Slave, Cellular, Island,

Hierarchical

G d l b

 Good results, but

 Only works for GA  Lacks of statistics

3 NIDISC'10, Atlanta, April 19-23

JG2A

Pa Parallel i rallel instances ances Parallel E allel Evaluation uation

NIDISC'10, Atlanta, April 19-23 4

19/04/2010 3

Current frameworks

Sequential Parallel Single Metaheuristic

JGA, SFERES, ECJ 18 JG2A, GALib

NIDISC'10, Atlanta, April 19-23 5

Metaheuristics

HotFrame, EasyLocal, iOpt, PISA, Templar ParadisEO, HeuristicLab

Candidates

Pa Para radisEO disEO HeuristicLab euristicLab

 C++

 Templates

 Four different modules

 Paradis-EO  Paradis-MO

P di MOEO

 C#  Multiple algorithms

 GA  Evolutionary strategies  Ant Colony  Tabu search

NIDISC'10, Atlanta, April 19-23 6

 Paradis-MOEO  Paradis-PEO

 GT4 + MPI

 Difficult to use  No grid enabled

 Security  Heterogeneity

19/04/2010 4

The ALS framework

 ALS, Adaptive Learning Search provides a conceptual frame to

analyze Metaheuristics

 B

d th t t

 Based on the operator concept

 Any specific operation  Represents an algorithm, a set of algorithms or a single Metaheuristic

• peration

 4 operators are common to any Metaheuristic Initialize Iterate <while non stop condition>

NIDISC'10, Atlanta, April 19-23 7

Sampling, selects elements Learning, extract information Diversification, search for new solutions Intensification, try to improve current solution Replace, replace the old solution End

Validating ALS

Variable Neighborhood Search Simulated Annealing Genetic Algorithms Tabu Search 8 NIDISC'10, Atlanta, April 19-23

19/04/2010 5

pALS proposal

 To develop a framework making easy the

implementation of different metaheuristics on parallel environments, helping with

 Parallel design  Callback services  Statistic services  Cooperative strategies

Operator

NIDISC'10, Atlanta, April 19-23 9

Transformation Statistics Cooperator Callbacks Solutions Solutions

pALS architecture

NIDISC'10, Atlanta, April 19-23 10

19/04/2010 6

 Contains classes for modeling

metaheuristics

 Solution transformation

based process

 Most classes are abstract

classes classes

 Behavior defined by

configuration files

 Define operators and their

parameters

 Every operator has its own

configuration file (or inherit from its parent) M k f ALS h b

Solution 11

 Makes of pALS a white box

framework

pALS Core Layer

Representation Results

1 1

F.O. 1 = 10 NIDISC'10, Atlanta, April 19-23

 Represents a transformation

• f a set of solutions

 May be used to model

complex (i.e. metaheuristics)

• r simple (i.e. crossover

Solutions pALS Genetic

p

• perator) operations

 An execution is a hierarchy of

• perators

Solutions 12

Operator

Operator Solutions ListOperator ObjectOperator Genetic Algorithm Mutation Crossover Evaluation Selection Solutions

Flip bits in a solution Select best n solutions

NIDISC'10, Atlanta, April 19-23

19/04/2010 7

 Cooperator i takes

solutions from (i-1)%m

Solutions Cooperator Solutions

 Common with small m

13

Ring Topology Cooperator Ring Topology Cooperator

p Operator Solutions Solutions NIDISC'10, Atlanta, April 19-23 Callback

• Defines conditional behaviors after

ending an operator execution

• Defined as a condition and an action

Statistic Context Builder Soluciones Operator Soluciones

• About transformations inside
• perators (time, best solution, etc.)
• Bounded by the InitStatistic and

EndStatistic methods

14

Support classes Support classes

InitStatistic EndStatistic CheckCallbacks

value = 10 vector = 1 2 3 4 matrix = 1 2 3 4 ,\ 5 6 7 8

ReadContext

• Allows to use external parameters

from a file or an external Java class

NIDISC'10, Atlanta, April 19-23

19/04/2010 8

• Captures operators’ execution logic
• Offers flexibility and transparency to

the execution (centralized, distributed, etc.)

ExecutionPrototype

• Every operator has its own

ExecutionPrototype defined by the user in a configuration file

• The ExecutionPrototype class can be

easily exteneded

Solutions Operator Solutions 15

ExecutionLayer

LocalDelegate GridDelegate ThreadDelegate Thread Pool Delegate NIDISC'10, Atlanta, April 19-23

Thread/Grid delegates

Solutions Files

Operator

Libs Certs GT4 GRAM

 Takes advantage of multi-

core platforms

 Launch a new thread for  Manages security

Operator Thread 1 Thread 2 Solutions

Files Operator Libs Certs

Operator Execution Prototype CondorPool Condor Master NIDISC'10, Atlanta, April 19-23 16

the operator’s execution Manages security

 Transparent for users

Thread/Grid Delegate

Solutions Operator Solutions

19/04/2010 9

Execution models

Multi-Boot Parallel

Execution

Mix

17 NIDISC'10, Atlanta, April 19-23

Life cycle

pALS pALS

Operator

Global Settings Settings Cooperator

Read

Context Builder Init Statistics Execution Prototype End Statistics Callback Coperator Publish Operator

18 NIDISC'10, Atlanta, April 19-23

19/04/2010 10

pALS offer

Flexibility Functionality

Operator Cooperator Context B ild

Wid / bilit

pALS

p p Callback Statistic Builder Settings 19

Wide scope / reusability Performance

Execution Prototype Thread Delegate Local Delegate Grid Delegate Genetic Algorithm Tabu Search Simulated Annealing VNS NIDISC'10, Atlanta, April 19-23

pALS experimentation

Boolean satisfiability problem (SAT) Tabu Search Boolean satisfiability problem (SAT) Tabu Search

NIDISC'10, Atlanta, April 19-23 20

Knapsack Genetic Algorithm with island deployment Knapsack Genetic Algorithm with island deployment

19/04/2010 11

Conclusions

 pALS is a framework implementing the metaheuristic

concepts behind ALS

 Not a set of metaheuristics  Allows for different parallel execution, transparent to users  White-box framework, ready to be used  Enhanced tools for research: callbacks, statistics  Self-incrementing  Public a ailable at http //sistemas uniandes edu co/~comit

NIDISC'10, Atlanta, April 19-23 21

 Public available at http://sistemas.uniandes.edu.co/~comit

 Needs further testing and comparisons

Configuration file example



# SAT Tabu Search Algorithm settings file



execution_instance = BasicTabuSearch



• bjectives

= max



• bjectives

max



representation_length = 50



execution_instance_settings_file = data/satisfiability.properties



neighborhood_size = 200



tabu_list_length = 5



representation = BinaryArrayRepresentation



init_solution = RandomPopulationGenerator



neighborhood = SinglePointFlipOperator



binary_flip_operator_rate= 0.5



record_tabu = ArrayIndexesTabuRecordOperator

NIDISC'10, Atlanta, April 19-23 22



tabu_check = ArrayIndexesTabuCheckOperator



evaluation = CNFSatisfiabilityOperator



selection = BestSolutionsSelectionOperator



solutions = 1



max_iterations = 1



cnf_file = data/aim-50-2_0-yes1-1.cnf

19/04/2010 12

Grid delegate execution

NIDISC'10, Atlanta, April 19-23 23