A comment on bio-inspired optimisation via GPU architecture: The genetic algorithm workload

Abstract

This paper characterizes a genetic algorithm based on the analysis of the workload of its operators. Different granular parallel implementations of a genetic algorithm in the GPU architecture are compared against the correspondent sequential version. With the help of three benchmark problems, a complete characterization of the relative execution times of the genetic operators, varying the population cardinality and the genotype size, is offered. The best speedups, obtained with large populations, are higher than one thousand times faster than the corresponding sequential version. The assessment of different granularity levels shows that the two-dimensional parallelism supported by the GPU architecture is valuable for the crossover operator. © 2012 Springer-Verlag.