Clonal Selection Algorithm Parallelization with MPJExpress
Download as PPTX, PDF1 like522 views
The document presents research on parallelizing the Clonal Selection Algorithm (CSA) using MPJ Express to address optimization problems like the Traveling Salesman Problem (TSP). It details the design stages for building parallel algorithms, evaluates the CSA's performance through datasets, and compares communication strategies. The findings indicate that multicommunication models yield better performance in terms of execution time and solution quality compared to master-slave communication models.
![INTRODUCTION .. (1)
In an optimisation problem, we need to
find, among many alternatives, a best or good
enough solution.
However, if the problems arise on much bigger
scales, they are complex and need computer
algorithms to solve them.
The Travelling Salesman Problem (TSP) in one
type of optimisation problem, where there is a
salesman who needs to find the shortest possible
route to visit each city exactly once and return to his
origin city [1].
This problem is a combinatorial optimisation that
needs a sophisticated algorithm to solve it.](https://image.slidesharecdn.com/ceec-usingmpjexpresstoimplementmessagepasingmodel-160929094648/85/Clonal-Selection-Algorithm-Parallelization-with-MPJExpress-3-320.jpg)
![INTRODUCTION .. (2)
To solve an optimisation problem, there are exact methods
that need no constraints or unlimited resources, such as time
and memory.
But in the real world there are always limited resources, and
this limitation leads to the development of heuristic or
metaheuristic approaches as a major field in operation
research [2].
These approaches provide optimal or sub-optimal feasible
solutions in a reasonable time that significantly reduces
the time of the search process.
However, the high dimensions of many tasks will always pose
problems and result in time-consuming scenarios.
Therefore, parallelism is an approach that not only aims
to reduce the resolution time but also to improve the quality of
the provided solutions.](https://image.slidesharecdn.com/ceec-usingmpjexpresstoimplementmessagepasingmodel-160929094648/85/Clonal-Selection-Algorithm-Parallelization-with-MPJExpress-4-320.jpg)
Clonal Selection Algorithm Parallelization with MPJExpress
- 1. CLONAL SELECTION ALGORITHM PARALLELIZATION WITH MPJEXPRESS Ayi Purbasari Informatics Department, Pasundan University, Bandung Indonesia pbasari@unpas.ac.id Presented at CEEC 2016 Essex University, Colchester, United Kingdom
- 3. INTRODUCTION .. (1) In an optimisation problem, we need to find, among many alternatives, a best or good enough solution. However, if the problems arise on much bigger scales, they are complex and need computer algorithms to solve them. The Travelling Salesman Problem (TSP) in one type of optimisation problem, where there is a salesman who needs to find the shortest possible route to visit each city exactly once and return to his origin city [1]. This problem is a combinatorial optimisation that needs a sophisticated algorithm to solve it.
- 4. INTRODUCTION .. (2) To solve an optimisation problem, there are exact methods that need no constraints or unlimited resources, such as time and memory. But in the real world there are always limited resources, and this limitation leads to the development of heuristic or metaheuristic approaches as a major field in operation research [2]. These approaches provide optimal or sub-optimal feasible solutions in a reasonable time that significantly reduces the time of the search process. However, the high dimensions of many tasks will always pose problems and result in time-consuming scenarios. Therefore, parallelism is an approach that not only aims to reduce the resolution time but also to improve the quality of the provided solutions.
- 5. INTRODUCTION .. (3) Most papers on parallel metaheuristics do not explain in detail the stages of designing a parallel algorithm itself. They describe the process of transformation of sequential algorithms into parallel algorithms generally. The focus of research is on the end result and the performance of parallelism. In this paper, we will explain the stages and steps in building a parallel algorithm. These detailed stages will be a guide in designing parallel algorithms, so that the resulting one will have a good performance.
- 6. RESEARCH METHOD • The algorithms: CSA • The problem: TSP • Immune Engineering Input • Parallel Designing • Evaluating Process • Parallel CSA Model • Parallel CSA Implementation using MPJExpress Output
- 7. CLONAL SELECTION ALGORITHM (CSA) AS BIO-INSPIRED COMPUTING 7
- 8. BIO-INSPIRED COMPUTING Artificial Immune System Genetic Algorithm Neural Net Anti Colony Optimization Bee Colony Swarm Intelligent 29/09/2016ayipurbasari 8
- 9. CSA .. (1) A Clonal Selection Algorithm (CSA) is an algorithm that inspired by a clonal selection mechanism to provide an immune response. This inspiration results in some algorithms in the class of Artificial Immune System (AIS) algorithms. Fig. 1 Clonal Selection
- 10. IMMUNE SYSTEM AND ARTIFICIAL IMMUNE SYSTEM • Immune Network Theory (Jerne) • Clonal Selection Theory (Burnet) • Negative Selection • Danger Theory (Matzinger) Immune System • Immune Network Algorithm • Clonal Selection Algorithm (de Castro & Von Zuben) • Negative Selection Algorithm • Dendrit Cell Algorithm Artificial Immune System 29/09/2016 10 ayipurbasari,icose2016
- 11. CSA .. (2) Clonal Selection Algorithm is one of the population-based heuristic search algorithms that has been able to solve combinatorial problems, such us the Travelling Salesman Problem (TSP). Fig. 2 Clonal Selection Algorithm
- 12. MAPPING CSA PARAMETERS AND THE TSP PROBLEM Clonal Selection Algorithm step TSP Problem Population initialization Set of randomly generated tours. There are (n-1) possibilities that the tours may be raised. This population is part of the whole tours. The number of tours is generated by the specified population size. Affinity evaluation Evaluation of affinity checks each tour that has occurred and finds the cost required to form the tour. Selection: affinity maturation Affinity is how close the cost of a tour is to the optimal/best cost. The closer, the higher affinity will be selected. Cloning and Hypermutation Cloning is a process to copy a selected tour, so the number of copies depends on the clone factor: . Cloned/copied tour will be mutated according to hypermutation probability mutate factor: RandomReplacement After this mutation, we will have the best tours, which will be replaced by the worst tours in the initial population. The number of worst tours replaced will depend on some random size replacement d. Stop condition Clonal selection process will be repeated until a stop condition is obtained. Stopping criteria could be the number of generations, numbers of populations (tours) evaluated, or the best cost found.
- 14. PARALLEL DESIGNING Starting with a problem specification, we do parallel design by engaging in partition, determining communication requirements, agglomerating tasks and, finally, mapping tasks to processors.
- 15. PARALLEL DESIGNING In this research, we focus on the first and second stages: Partitioning. The computation that is to be performed and the data operated on by this computation are decomposed into small tasks. Practical issues such as the number of processors in the target computer are ignored, and attention is focused on recognising opportunities for parallel execution. Communication. The communication required to coordinate task execution is determined and appropriate communication structures and algorithms are defined.
- 16. PARALLEL DESIGNING Parallel Designing: Partition domain/data decomposition and functional decomposition. Communication Master-slave communication Multi- communication
- 17. PARALLEL DESIGNING: PARTITION .. (1) There are two approaches in partition: domain/data decomposition and functional decomposition. Clonal selection deals with two sets of data: matrix of the dataset and population of tours. Since the algorithm stages need the whole matrix of a dataset to evaluate the affinity, we couldn’t do matrix decomposition. Since population of tours are dependent on each other, we could do population partition.
- 18. PARALLEL DESIGNING: PARTITION .. (2) Functional decomposition in CSA deals with stages of • evaluation, • selection, • clone and mutation, and • randomreplacement. Since all of these stages are dependent each other, we couldn’t do a functional decomposition for each one.
- 19. PARALLEL DESIGNING: PARTITION .. (2) There is some parallelism potential in each stage. It is possible to do functional decomposition. Then we have to calculate the granularity for each potential and compare this to whole process. Granularity is a size of task for each stage. Since we need much generation to evaluate the population, we will need G* (NL+NlogN+nNL+NlogN+NlogN), where G is the number of generation. We can conclude that functional decomposition in each stage is not recommended since their granularities are too small compared to the whole process. Since we couldn’t do functional decomposition for CSA loop, we did population decomposition. Each processing element is assigned to a population partition that will be computed for the CSA stages.
- 20. GRANULARITY FOR EACH STAGE Stages Granularity Compare to whole process Evaluation NL NL/ G*(NL+NlogN+nNL+NlogN+NlogN) Selection NlogN NlogN/ G*(NL+NlogN+nNL+NlogN+NlogN) Clone and hypermutati on nNL nNL/ G*(NL+NlogN+nNL+NlogN+NlogN) RandomRep lacement NlogN+Nlog N (NlogN+NlogN)/ G*(NL+NlogN+nNL+NlogN+NlogN)
- 21. PARALLEL DESIGNING: COMMUNICATION Master-slave communication Multi-communication Process 1 Process 2* Process n Process 0 Process 2 Process 1 Process n Process 0 Process 2 Process 0*
- 23. MPI AND MPJEXPRESS We use the Message-Passing Interface (MPI), the de facto message-passing standard for concreteness and MPJExpress, as one of the libraries that implemented MPI. MPI is not a library but rather the specification of what such a library should be. MPJExpress is a library that is implemented with MPI’s specification.
- 24. MPJEXPRESS The Java Grande forum defined MPI-like API for Java as mpiJava 1.2 API, which is the Java equivalent of the MPI 1.1 specification document. MPJ Express is one type of a reference implementation of the mpiJava 1.2 API. MPJ Express is a message-passing library that can execute parallel Java applications on computed clusters or networks of computers used by application developers. It is like a middleware that supports communication between individual processors of clusters. MPJ Express supported Single Program Multiple Data (SPMD).
- 25. MPJEXPRESS Therearethreewaysto configureMPJExpress: multicore configuration, cluster configuration and hybrid configuration. The Multicore Configuration is used to execute MPJ Express user programs on laptops and desktops. In this research, we use multicore configuration
- 26. MPJEXPRESS METHOD: SEND Synopsis: public void Send(java.lang.Object buf, int offset, int count, Datatype datatype, int dest, int tag)throws MPIException Example in MPJExpress: MPI.COMM_WORLD.Send(objectP, 0, objectP.length, MPI.OBJECT, dest, tag);
- 27. MPJEXPRESS METHOD: RECV Synopsis: public void Recv (java.lang.Object buf,int offset,int count,Datatype datatype,int source,int tag)throws MPIException Example in MPJExpress: MPI.COMM_WORLD.Recv(objectP, 0, objectP.length, MPI.OBJECT, 0, tag);
- 28. MPJEXPRESS METHOD: ALLGATHER Synopsis: public void Allgather(java.lang.Object sendbuf, int sendoffset,int sendcount,Datatype sendtype, java.lang.Object recvbuf,int recvoffset, int recvcount,Datatype recvtype) throws MPIException Example in MPJExpress: MPI.COMM_WORLD.Allgather(sendBuff, 0, unitSize, MPI.OBJECT, recvBuff, 0, unitSize, MPI.OBJECT);
- 30. RESULT With datasets from TSPLIB, we implemented two model of Clonal Selection Algoritm and result 1) comparison two model, 2) best cost for several dataset and 3) execution time for several dataset.
- 31. RESULT - 20.000 40.000 60.000 80.000 100.000 120.000 140.000 160.000 1 60 119 178 237 296 355 414 473 532 591 650 709 768 827 886 945 BestCost Generation Comparing BestCost Model-1 Model-2 - 50.000 100.000 150.000 200.000 250.000 300.000 350.000 1 2 3 4 5 6 7 Bestcost Dataset Best cost Model-1 Model-2
- 32. RESULT Using dataset burma14, berlin52, ulysses22, tsp225, ch150, kroa100, and pr76 from TSPLib, we can see that model-2 gained lower best cost for each dataset. - 5.000 10.000 15.000 20.000 25.000 30.000 35.000 40.000 1 2 3 4 5 6 7 Executiontime(ms) Dataset Execution time Model-1 Model-2
- 33. DISCUSSION We described how to implement MJExpress to parallel the clonal selection algorithm, from parallel designing to implementation. The most important part of these steps is parallelism exploitation analysis, especially each of the process granularities. For a population-based optimisation algorithm like this clonal selection algorithms, the best paralleling is achieved by doing their clonal operation at the same time for different population data and then communicating with each other. We have two-way communications, by master-slave communications or by multicommunication. In their implementations, we need a different method library. Using master-slave communication is less simple, or just using point-to- point communication, but it needs synchronisation between population- generation controlled by the master process. Using multicommunication, we could use a global communication, allgather, which is more sophisticated but we need to be more careful to define which data has to be sent between processes. From experiment result, we can see that using multicommunication model, we gained lower best cost dan lower execution time
- 34. CONCLUSION To parallelise population-based algorithms, we need to exploit and define their granularity for each stage. After that we do data or functional partition, if necessary, and use the communication model. Using a library for the message-passing model, we should define appropriate methods to implement process communication. MPJExpress, as one of the implementations of the message-passing interface specification, has several communication methods that fit in with the SPMD model. MPJExpress also has a MPI.OBJECT datatype to support object- oriented programming. This research finds results of pseudo-code using MPJExpress that can be implemented using Java Language within their environments. Two kinds of communication between processes can be executed for several datasets from the TSP problem, and then we can compare their performances using execution time and the best cost achieved. We conclude that multicommunication model will have better performance than master slave model.
- 35. THANK YOU