Optimal design of master-worker architecture for parallelized simulation optimization

Abstract

This study formulates and solves the design problem for a master-worker architecture dedicated to the implementation of a parallelized simulation optimization algorithm. Such a formulation does not assume any specific characteristic of the optimization problem being solved, but the way the algorithm is parallelized. In particular, we refer to the master-worker paradigm, where the master makes sampling decisions while the workers receive solutions to evaluate. We identify two metrics to be optimized: the throughput of the workers in terms of the number of evaluations per time unit, and the lack of synchronization between the master and the workers. We identify several design parameters: number of workers (n), the buffer size for each worker and for the master and the sample size m, i.e., the number of solutions used by the master for sampling decisions at each iteration. Numerical experiments show optimal designs over randomly generated simulation optimization algorithm instances.