MACHINE LEARNING-ASSISTED LOAD TESTING

The increasing worldwide demand for software systems involved in society has led to the need where not only functionality is fundamental and addressed, but end-user satisfaction in terms of availability, throughput, and response time is essential and should be preserved. Thus, systems must be evaluated at preset load levels to assess the non-functional quality problems from the closest perspective of real application use. In this context, where the problem involves a high and complex search space, a search-based approach for load test generation is required. This thesis proposes and evaluates an evolutionary search-based approach for load test generation using multi-objective optimization methods consisting of selection, crossover, and mutation operators. In this thesis, load testing is addressed as a multi-objective optimization problem by using four different evolutionary algorithms: Non-dominated Storing Genetic Algorithm II (NSGA-II), Pareto Archived Evolution Strategy (PAES), The Strength Pareto Evolutionary Algorithm 2 (SPEA2), Multi-Objective Cellular Genetic Algorithm (MOCell) as well as a Random Search algorithm. Additionally, this study demonstrates the applicability of the proposed approach by running several experiments, aiming to compare the algorithms’ efficiency based on different quality indicators such as hypervolume, spread, and epsilon. Experimental results show that evolutionary search-based methods can be used to generate effective workloads. Since, all algorithms have found the optimal workload, having the hypervolume values to zero, we believe that the objectives of the problem could be combined as a single objective, hence scalarization techniques can be applicable. Based on the other quality indicators (Spread and Epsilon respectively), NSGA-II and MOCell tend to perform better compared to other algorithms. Finally, the study concludes that multi-objective evolutionary algorithms can be used for load testing purpose, obtaining better results in generating optimal workloads than an existing (adapted) model-free reinforcement learning approach