Abstract

Performance is an important aspect of software quality. The goals of performance are typically defined by setting upper and lower bounds for response time and throughput of a system and physical level measurements such as CPU, memory, and I/O. To meet such performance goals, several performance-related activities are needed in development (Dev) and operations (Ops). Large software system failures are often due to performance issues rather than functional bugs. One of the most important performance issues is performance regression. Although performance regressions are not all bugs, they often have a direct impact on users’ experience of the system. The process of detection of performance regressions in development and operations is faced with challenges. First, the detection of performance regression is conducted after the fact, i.e., after the system is built and deployed in the field or dedicated performance testing environments. Large amounts of resources are required to detect, locate, understand, and fix performance regressions at such a late stage in the development cycle. Second, even we can detect a performance regression, it is extremely hard to fix it because other changes are applied to the system after the introduction of the regression.
These challenges call for further in-depth analyses of the performance regression. In this thesis, to avoid performance regression slipping into operation, we first perform an exploratory study on the source code changes that introduce performance regressions in order to understand root-causes of performance regression in the source code level. Second, we propose an approach that automatically predicts whether a test would manifest performance regressions in a code commit. Most of the performance issues are related to configurations. Therefore, third, we propose an approach that predicts whether a configuration option manifests a performance variation issue. To assist practitioners to analyze system performance with operational data, we propose an approach to recovering field-representative workload that can be used to detect performance regression.