Developing high quality software is of eminent importance to keep the existing customers satisfied and to remain competitive. One of the most important software quality characteristics is performance, which defines how fast and/or efficiently a software can perform its operation. While several studies have shown that field problems are often due to performance issues instead of feature bugs, prior research typically treats all bugs as similar when studying various aspects of software quality (e.g., predicting the time to fix a bug) or focused on other types of bug (e.g., security bugs). There is little work that studies performance bugs. In this thesis, we perform an empirical study to quantitatively and qualitatively examine performance bugs in the Mozilla Firefox and Google Chrome web browser projects in order to find out if performance bugs are really different from other bugs in practice and to understand the rationale behind those differences. In our quantitative study, we find that performance bugs of the Firefox project take longer time to fix, are fixed by more experienced developers, and require changes to more lines of code. We also study performance bugs relative to security bugs, since security bugs have been extensively studied separately in the past. We find that security bugs are re-opened and tossed more often, are fixed and triaged faster, are fixed by more experienced developers, and are assigned more number of developers in the Firefox project. Google Chrome project also shows different quantitative characteristics between performance and non-performance bugs and from the Firefox project. Based on our quantitative results, we look at that data from a qualitative point of view. As one of our most interesting observation, we find that end-users are often frustrated with performance problems and often threaten to switch to competing software products. To better understand, the rationale for some users being very frustrated (even threatening to switch product) even though most systems are well tested, we performed an additional study. In this final study, we explore a global perspective vs a user centric perspective of analyzing performance data. We find that a user-centric perspective might lead to a small number of users with considerably poor performance while the global perspective might show good or same performance across releases. The results of our studies show that performance bugs are different and should be studied separately in large scale software systems to improve the quality assurance processes related to software performance. / Thesis (Master, Computing) -- Queen's University, 2012-04-30 01:28:22.623

Modern computer software systems are complicated. Developers can change the behavior of the software system through software configurations. The large number of configuration option and their interactions make the task of software tuning, testing, and debugging very challenging. Performance is one of the key aspects of non-functional qualities, where performance bugs can cause significant performance degradation and lead to poor user experience. However, performance bugs are difficult to expose, primarily because detecting them requires specific inputs, as well as specific configurations. While researchers have developed techniques to analyze, quantify, detect, and fix performance bugs, many of these techniques are not effective in highly-configurable systems. To improve the non-functional qualities of configurable software systems, testing engineers need to be able to understand the performance influence of configuration options, adjust the performance of a system under different configurations, and detect configuration-related performance bugs. This research will provide an automated framework that allows engineers to effectively analyze performance-influence configuration options, detect performance bugs in highly-configurable software systems, and adjust configuration options to achieve higher long-term performance gains. To understand real-world performance bugs in highly-configurable software systems, we first perform a performance bug characteristics study from three large-scale opensource projects. Many researchers have studied the characteristics of performance bugs from the bug report but few have reported what the experience is when trying to replicate confirmed performance bugs from the perspective of non-domain experts such as researchers. This study is meant to report the challenges and potential workaround to replicate confirmed performance bugs. We also want to share a performance benchmark to provide real-world performance bugs to evaluate future performance testing techniques. Inspired by our performance bug study, we propose a performance profiling approach that can help developers to understand how configuration options and their interactions can influence the performance of a system. The approach uses a combination of dynamic analysis and machine learning techniques, together with configuration sampling techniques, to profile the program execution, analyze configuration options relevant to performance. Next, the framework leverages natural language processing and information retrieval techniques to automatically generate test inputs and configurations to expose performance bugs. Finally, the framework combines reinforcement learning and dynamic state reduction techniques to guide subject application towards achieving higher long-term performance gains.