Running krill herd algorithm on Hadoop: A performance study

Abstract

The krill herd algorithm was introduced mimicking the herding behavior of krill individuals. The krill herd algorithm obtains the optimum solution by considering two factors, namely the density-dependent attraction of the krill, and the areas of high food concentration. The iterative optimization procedure is based on the time-dependent position of the krill individuals of (1) movement induced by other krill individuals, (2) foraging activity, and (3) random diffusion. In this paper, the krill herd algorithm is investigated when run on a Hadoop cluster. The algorithm was parallelized using MapReduce. Four different sets of experiments are conducted to evaluate the krill herd algorithm in terms of speed and accuracy. The first set of experiments investigates the execution time and speedup of the krill herd algorithm applied to six different benchmark functions. The second set of experiments investigates the varying dimensions of the Alpine benchmark function with regards to the effect on the execution time. The third set of experiments analyzes the effect of different krill sizes and numbers of maximum iterations on the execution time but also on the objective value. The fourth set of experiments compares the speed gain obtained when running the MapReduce-enabled version of the krill herd algorithm compared to the normal non-parallelized krill herd algorithm.