Abstract

Software systems have become increasingly complex, which makes it difficult to detect the root causes of performance degradation. Software tracing has been used extensively to analyze the system at run-time to detect performance issues and uncover the causes. There exist several studies that use tracing and other dynamic analysis techniques for performance analysis. These studies focus on specific system characteristics such as latency, performance bugs, etc. In this thesis, we review the literature to build a catalogue of performance analysis patterns that can be detected using trace data. The goal is to help developers debug run-time and performance issues more efficiently. The patterns are formalized and implemented so that they can be readily referred to by developers while analyzing large execution traces. The thesis focuses on the traces of system calls generated by the Linux kernel. This is because no application is an island and that we cannot ignore the complex interactions that an application has with the operating system kernel if we are to detect potential performance issues.