Spark Clustering Computing Platform Based Parallel Particle Swarm Optimizers for Computationally Expensive Global Optimization

The increasing demands on processing large-scale data from both industry and academia have boosted the emergence of data-intensive clustering computing platforms. Among them, Hadoop MapReduce has been widely adopted in the evolutionary computation community to implement a variety of parallel evolutionary algorithms, owing to its scalability and fault-tolerance. However, the recently proposed in-memory Spark clustering computing framework is more suitable for iterative computing than disk-based MapReduce and often boosts the speedup by several orders of magnitude. In this paper we will parallelize three mostly cited versions of particle swarm optimizers on Spark, in order to solve computationally expensive problems. First we will utilize the simple but powerful Amdahl’s law to analyze the master-slave model, that is, we do quantitative analysis based on Amdahl’s law to answer the question on which kinds of optimization problems the master-slave model could work well. Then we will design a publicly available Spark-based software framework which parallelizes three different particle swarm optimizers in a unified way. This new framework can not only simplify the development workflow of Spark-based parallel evolutionary algorithms, but also benefit from both functional programming and object-oriented programming. Numerical experiments on computationally expensive benchmark functions show that a super-linear speedup could be obtained via the master-slave model. All the source code are put at the complementary GitHub project for free access.