Rémy Dautriche - Techniques d’Interaction Multi-Echelles Pour la Visualisation Interactive de Traces d’Exécution

Developing streaming multimedia applications on embedded systems becomes increasingly complex over time. New multimedia standards reach the market to support better resolutions and overall improved quality delivered to the end-user. Consequently, hardware platforms complexify and developing the software to fully exploit them becomes harder at each new generation. The traditional debugging method for streaming applications is the usage of execution traces. However, the amount of data generated by modern software largely increases and existing tools do not allow an efficient debugging process as they become unable to tackle large amounts of data. In this thesis, we focus on new interactive visualization techniques enriched by results of data mining algorithms for a more efficient analysis of execution traces for multimedia applications.
First, we introduce Slick Graphs, a binning and smoothing technique for time series visualization. Slick Graphs mitigate the quantization artifacts, introduced by the traditional smoothing techniques, by using the smallest possible binning intervals, i.e. pixels. We compared Slick Graphs to traditional smoothing techniques in a user study and show that the Slick Graphs are significantly faster and more accurate when working with periodic data. We then propose a novel interaction visualization framework, TraceViz, to explore the execution traces at different level of details and integrate the Slick Graphs to provide a global overview of the trace. With TraceViz, we also introduce a fast back-end to support the interactive browsing of huge traces. We perform a performance analysis to show that the TraceViz back-end outperforms the back-end used in state-of-the-art debugging tools for execution traces. Execution traces contain meaningful information that can be computed using data mining techniques.
A wide range of patterns can be computed and provide valuable information: for example existence of repeated sequences of events or periodic behaviors. However, while pattern mining approaches provide a deeper understanding of the traces, their results is hard to understand due to the large amount of patterns that have to be examined one by one. We propose a novel visual analytics method that allows to immediately visualize hidden structures such as repeated sets/sequences and periodicity, allowing to quickly gain a deep understanding of the trace. Finally, we also show how our method can be applied with different types of data than execution traces.