On Improving (Non)Functional Testing

Software testing is commonly classified into two categories, nonfunctional testing and functional testing. The goal of nonfunctional testing is to test nonfunctional requirements, such as performance and reliability. Performance testing is one of the most important types of nonfunctional testing, one goal of which is to detect the phenomena that an Application Under Testing (AUT) exhibits unexpectedly worse performance (e.g., lower throughput) with some input data. During performance testing, a critical challenge is to understand the AUT’s behaviors with large numbers of combinations of input data and find the particular subset of inputs leading to performance bottlenecks. However, enumerating those particular inputs and identifying those bottlenecks are always laborious and intellectually intensive. In addition, for an evolving software system, some code changes may accidentally degrade performance between two software versions, it is even more challenging to find problematic changes (out of a large number of committed changes) may lead to performance regressions under certain test inputs. This dissertation presents a set of approaches to automatically find specific combinations of input data for exposing performance bottlenecks and further analyze execution traces to identify performance bottlenecks. In addition, this dissertation also provides an approach that automatically estimates the impact of code changes on performance degradation between two released software versions to identify the problematic ones likely leading to performance regressions. Functional testing is used to test the functional correctness of AUTs. Developers commonly write test suites for AUTs to test different functionalities and locate functional faults. During functional testing, developers rely on some strategies to order test cases to achieve certain objectives, such as exposing faults faster, which is known as Test Case Prioritization (TCP). TCP techniques are commonly classified into two categories, dynamic and static techniques. A set of empirical studies has been conducted to examine and understand different TCP techniques, but there is a clear gap in existing studies. No study has compared static techniques against dynamic techniques and comprehensively examined the impact of test granularity, program size, fault characteristics, and the similarities in terms of fault detection on TCP techniques. Thus, this dissertation presents an empirical study to thoroughly compare static and dynamic TCP techniques in terms of effectiveness, efficiency, and similarity of uncovered faults at different granularities on a large set of real-world programs, and further analyze the potential impact of program size and fault characteristics on TCP evaluation. Moreover, in the prior work, TCP techniques have been typically evaluated against synthetic software defects, called mutants. For this reason, it is currently unclear whether TCP performance on mutants would be representative of the performance achieved on real faults. to answer this fundamental question, this dissertation presents the first empirical study that investigates TCP performance when applied to both real-world faults and mutation faults for understanding the representativeness of mutants