Using trace sampling techniques to identify dynamic clusters of classes
Philippe Dugerdil · Proceedings of CASCON · 2007
In dynamic analysis (i.e. execution trace analysis), an important problem is to cope with the volume of data to process. However, in the literature, no definitive solution has yet been proposed. Generally, the techniques start by compressing the execution trace before proceeding with the analysis. In this paper we propose a way to process the uncompressed execution trace using a sampling technique. Then, we present the concept of temporally omnipresent class that is the analogy of the "noise" in signal processing. During analysis, the omnipresent classes can be filtered out to concentrate only on relevant ones. Next, we present the extension of our sampling technique to the dynamic clustering of classes. This is a way to recover the components of a legacy system. We finally show the application of this approach to a medium size industrial software system, as well as the tool that supports it. As a conclusion, we suggest that our noise reduction and clustering techniques are both efficient and scalable.