Parallel Optimization Methods

This chapter introduces common techniques used in developing evolutionary algorithms for distributed systems, providing a survey of different methods. Most of these methods are evaluated against a common suite of benchmark equations. Chapter 8 provides some common example equations, and we refer the reader to Tang et al. for a comprehensive overview (Tang et al., Benchmark functions for the cec2008 special session and competition on large scale global optimization, in Nature Inspired Computation and Applications Laboratory (USTC, China, 2007)). Strategies for single-population, multiple-population, and cellular algorithms are presented, and then particular emphasis is placed on the implementation used for optimizing FDTD, which uses asynchronous updates of the population to aid in scalability and improve performance. Results are presented for the performance of this implementation using CPU, GPUs, and hybrid CPU–GPU strategies for evaluation of the objective functions.