Big optimization with genetic algorithms: Hadoop, Spark, and MPI

Abstract

Solving problems of high dimensionality (and complexity) usually needs the intense use of technologies, like parallelism, advanced computers and new types of algorithms. MapReduce (MR) is a computing paradigm long time existing in computer science that has been proposed in the last years for dealing with big data applications, though it could also be used for many other tasks. In this article, we address big optimization: the solution to large instances of combinatorial optimization problems by using MR as the paradigm to design solvers that allow transparent runs on a varied number of computers that collaborate to find the problem solution. We study and analyze the MR technology, focusing on Hadoop, Spark, and MPI as the middleware platforms to develop genetic algorithms (GAs). From this, MRGA solvers arise using a different programming paradigm from the usual imperative transformational programming. Our objective is to confirm the expected benefits of these systems, namely file, memory, and communication management, over the resulting algorithms. We analyze our MRGA solvers from relevant points of view like scalability, speedup, and communication vs. computation time in big optimization. The results for high-dimensional datasets show that the MRGA over Hadoop outperforms the implementations in Spark and MPI frameworks. For the smallest datasets, the execution of MRGA on MPI is always faster than the executions of the remaining MRGAs. Finally, the MRGA over Spark presents the lowest communication times. Numerical and time insights are given in our work, so as to ease future comparisons of new algorithms over these three popular technologies.