Searching for high-performing software configurations with metaheuristic algorithms
Chong Tang, Kevin Sullivan, Baishakhi Ray · 2018
Modern systems often have complex configuration spaces. Research has shown that people often just use default settings. This practice leaves significant performance potential unrealized. In this work, we propose an approach that uses metaheuristic search algorithms to explore the configuration space of Hadoop for high-performing configurations. We present results of a set of experiments to show that our approach can find configurations that perform significantly better than defaults. We tested two metaheuristic search algorithms---coordinate descent and genetic algorithms---for three common MapReduce programs---Wordcount, Sort, and Terasort---for a total of six experiments. Our results suggest that metaheuristic search can find configurations cost-effectively that perform significantly better than baseline default configurations.