Using Google's Compute Engine Service Pricing as a Reference for Comparison Between Master-Slave and Island Model-Based Fully Distributed Genetic Algorithm

Abstract

Common Master-Slave based Parallel Genetic Algorithms can efficiently utilize multiple computational nodes working on a single process, which leads to finding solutions in less execution time. However, Master-Slave model causes high communication traffic between the nodes. On the other hand, Island-model based parallel implementation can reduce the amount of traffic between the nodes while reducing the efficiency in finding solutions. However it is discussed deeply in the literature, the tradeoff between execution time, traffic size and quality of result have not been estimated based on actual money cost. This paper presents a comparison between Master-Slave and Island-model based fully distributed implementations for Genetic Algorithm. The comparison is based on the actual money cost when running on Google Cloud Compute Engine Service. We implemented Genetic Algorithm in Master-Slave model and in Island model running in different migration rates in order to find a reasonable migration rate that can help find high quality results will the least possible cost.