Performance Multi-objective Evolutionary Optimization At Software Architecture Level Based On Random Search Rule

In Architecture-based software performance optimization can help to find performance problems and mitigate their negative effects at the early stage.To automatize this optimization process,some rule-based and metaheuristic-based performance optimization methods at software architecture(SA)level have been proposed in recent years.However,in the most existing rule-based methods,the count and the order and improvement amplitude of each rule have not been fully considered in search process.So the search space is limited and better solutions may be excluded.For the current most metaheuristic-based methods,software architects are confused about how to optimize the performance.These methods are lack of interpretability because performance improvement knowledge is ignored.To address these problems,we propose the performance multi-objective evolutionary optimization at software architecture level based on random search rule,named RRM04P0.Our research contents are as follows.Firstly,random search rules are designed in the views of deployment,structure and behavior.These rules enlarge the performance improvement space by randomizing the action part of the performance improvement strategy at SA level while maintaining interpretability.Then the random search rules based performance optimization model(RRPOM)is presented.And the RRPOM characterizes the relationship between rules,SA and optimization goals such as system reponse time,availability and improvement cost.Then the search space is enlarged further by trading off availability and improvement cost.Furthermore,the multi-objective evolutionary algorithm for performance optimization(EA4PO)with constraints checking and repairing,which can search for Pareto optimal solution in large space efficiently,is designed for solving RRPOM.Finally,a series of comparative experiments,including both RRM04PO vs PB method(a typical rule-based method)and RRM04P0 vs PCM method(a typical metaheuristic-based method),are conducted on six different software cases.In comparison with PCM method in the three indicators of contribution,generational distance and hypervolume,RRM04P0 is significantly better.Meanwhile better interpretability can be obtained.In contrast with PB method,RRM04P0 can obtain the better solutions by using fewer rules and fewer number of modification to SA elements.