Genetic Optimization of Parallel FDTD Computations

In this paper we present and compare two algorithms to transform and optimize the finite difference time domain (FDTD) method computations in irregular computational areas for multiprocessor systems. Both algorithms are based on a genetic approach with different design of chromosomes. In the first case, a chromosome represents a complete merging algorithm as a sequence of merging functions that are used to reduce a number of macro nodes. In the second case, a chromosome corresponds to an assignment of each macro node to specified processor. We have tested a mixed fitness function, which can switch between two sub functions: cut min value and difference between the maximally and minimally loaded macro nodes in an individual. As our experiments have shown, it allows to adjust the behavior of a genetic algorithm to the requirements of the computational system. We have tested both algorithms for multiprocessor systems with shared and distributed memory with MPI and RDMA communication.