Colloquium - Diwan

Many modern commercial and open source applications include tens of millions of lines of code each. These large applications are (i) difficult to write, (ii) hard to understand and evolve, and (iii) resource intensive. My research addresses these three challenges by (i) rethinking how and what we teach computer scientists, (ii) building tools that support software engineering tasks, and (iii) developing automatic and semi-automatic techniques for improving the resource usage of applications. In my talk I'll focus on Vertical Profiling.

Vertical profiling is a tool-supported methodology for understanding the performance of modern applications. Modern applications, such as ones written in Java, run on top of many components, such as dynamic compilers, garbage collectors, and middleware. These components interact with each other and with the application. Thus, one cannot understand the performance of such applications by looking at only the application or hardware-level profiles. Instead, vertical profiling collects traces from all the components (including the application and hardware) and uses statistical and visualization techniques on these traces to explain performance phenomenon. This talk will demonstrate vertical profiling by applying it to several previously unexplained performance phenomenon in popular Java applications.