SPSO Parallelization Strategies for Electromagnetic Applications
Two parallelization techniques, GPGPU and Pthreads for multiprocessor architectures, are used to implement a SPSO algorithm in order to solve electromagnetic optimization problems. Several configurations for the GPGPU implementation are tested and a new f
- PDF / 780,283 Bytes
- 21 Pages / 439.37 x 666.142 pts Page_size
- 10 Downloads / 201 Views
Faculty of Electrical Engineering, Politehnica University of Bucharest, Bucharest, Romania {anton.duca,gabriela.ciuprina,daniel.ioan}@upb.ro 2 Faculty of Computer Science, Politehnica University of Bucharest, Bucharest, Romania [email protected]
Abstract. Two parallelization techniques, GPGPU and Pthreads for multiprocessor architectures, are used to implement a SPSO algorithm in order to solve electromagnetic optimization problems. Several configurations for the GPGPU implementation are tested and a new full parallel minimum branching implementation is proposed. The best GPGPU approaches are then compared with a Pthreads implementation in terms of speed up and solution quality. To test the efficiency of the parallelization techniques two electromagnetic optimization problems were chosen, namely the TEAM22 benchmark and Loney’s solenoid. In the end the paper provides suggestions regarding what parallelization technique should be used considering the implementation features of the optimization function. Keywords: SPSO GPGPU Pthreads Electromagnetic field Optimization
1 Introduction Electromagnetic optimizations problems are well known for their complex objective functions which for evaluation involve solving electromagnetic field equations. The objective function is most of the times multidimensional, with several local minimum and a wide search area, while the optimization variables often have to meet difficult constraints. For this reasons the evaluation of the objective function is usually computational intensive, requires a large number of subroutine calls (sometimes recursive), having a high level of branching, and many instructions [1, 2]. Since deterministic approaches like the gradient descent or conjugate gradient can not be used because of the multiple local minimum, in the past years stochastic methods based on, simulated annealing, tabu search, genetic algorithms, or swarm optimization, were widely adopted as standard methods for solving electromagnetic problems [3, 4]. The advantages of the heuristics based methods are their ability to find the global optimum, usually without knowing the objective function derivatives, and their robustness. The disadvantage of the stochastic methods is the large number of
© Springer International Publishing AG 2017 J.J. Merelo et al. (eds.), Computational Intelligence, Studies in Computational Intelligence 669, DOI 10.1007/978-3-319-48506-5_5
76
A. Duca et al.
evaluations for the objective function, essential for real world optimization problems which for the evaluation is often time consuming. To decrease the solving time there are the following options: to reduce the number of objective function evaluations by using more efficient stochastic methods [5, 6], to implement parallel and/or distributed optimization algorithm architectures [7], or to decrease the objective function evaluation time using methods specific to electromagnetic problems [8]. In this paper two different parallelization techniques, Pthreads (POSIX threads) for multiprocessor architect
Data Loading...
