Professor Lu Xiao gave a research seminar in CIS
Dr. Lu Xiao, Assistant Professor at Stevens Institute of Technology, visited the CIS Department (virtually) and gave a research seminar on data-driven and architecture-centric software performance optimization
Dr. Lu Xiao, Assistant Professor at Stevens Institute of Technology, visited the Computer and Information Science Department (virtually) and gave a research seminar on data-driven and architecture-centric software performance optimization.
Software performance is a critical quality attribute measured by the timeliness, responsiveness, and resource consumption of a system at run-time. Many real-world cases revealed that ensuring performance is as important as achieving the functional goals for a successful software system. In the past decades, Moore's Law has greatly benefited software performance by offering exponentially more powerful hardware resources. Unfortunately, the advancement on the hardware side is reaching physical limitations in recent years. Therefore, the need for transforming performance engineering techniques on the software side becomes more critical and urgent.
In this seminar, Dr. Xiao presented her research team’s effort in three related studies. In the first study, the team developed a novel approach to automatically identify performance issue reports from the issue tracking databases. The approach is the combination of a set of 80 linguistic patterns---manually extracted from thousands of real-life performance issue reports---and classic machine and deep learning models. The long-term goal of this direction is to build open, large-scale datasets of real-life performance issues for supporting related future studies. In the second study, the team conducted an empirical study of 200 real-life performance issues---investigating how these issues are caused and resolved. This study specifically focused on the relationship between performance issue resolution and software design. It revealed that a non-trivial (30%) amount of the real-life performance issues require design-level optimization that involve groups of related source files. This study illuminates future work in revealing performance optimization tactics at different levels. Finally, the team contributed a new architecture model, namely, the Butterfly Space modeling, which provides a new perspective to systematically investigate whether and how performance improvement opportunities are architecturally connected to each other. This study showed that the Butterfly Space modeling can potentially be used for more efficient performance optimization under two different scenarios, compared to classic profiling-based approaches.