Multi-objective software performance optimisation at the architecture level using randomised search rules

Abstract Architecture-based software performance optimisation can help to find potential performance problems and mitigate their negative effects at an early stage. To automate this optimisation process, rule-based and metaheuristic-based performance optimisation methods have been proposed. However, existing rule-based methods explore a limited search space, potentially excluding optimal or near-optimal solutions. Most of current metaheuristic-based methods ignore existing practical knowledge of performance improvement, and lead to solutions that are not easily explicable to humans. To address these problems, we propose a novel approach for performance optimisation at the software architecture level named Multiobjective performance Optimisation based on Randomised search rulEs (MORE). First, we design randomised search rules (MORE-R) to provide explanation without parameters while benefiting from existing practical knowledge of performance improvement. Second, based on all possible composite applications of MORE-R, an explicable multi-objective optimisation problem (MORE-P) is defined to enlarge search space and enable solutions explicable to architectural stakeholder. Third, a multi-objective evolutionary algorithm (MORE-EA) with an introduced do-nothing rule, innovative encoding and repair mechanism is designed to effectively solve MORE-P. The experiments show that MORE is able to achieve more explicable and higher quality solutions than two state-of-the-art techniques. They also demonstrate the benefits of integrating search-based software engineering approaches with practical knowledge.