New MP-SoC profiling tools based on data mining techniques

LAGRAA Sofiane
under the direction of Frédéric PETROT (TIMA) and Alexandre TERMIER
Funding: BQR DMSoCProfile, Grenoble INP (Sep 10 - 14)

The goal of this thesis is to define and design new profiling tools for the evaluation of performance and power for applications mapped on Multi-Processor System-on-Chip. As the number of processors and tasks increases in embedded applications, being able to automatically extract hot spot or spot dysfunctional behaviors from gigabytes of traces becomes both an interesting scientific challenge and a practical issue. We believe that an approach based on Frequent Pattern Mining techniques,coming from the Data Mining community, is the appropriate choice for digging into the large amount of data to process. Frequent Pattern Mining allows to discover patterns having a complex structure (sequences, trees or graphs) occurring frequently in data. The thesis will study the applicability of the Frequent Pattern Mining techniques to trace data extracted from simulated program executions on the SoC and propose solution to make it practical. It also targets to adapt these techniques in order to discover problematic access patterns such as memory contention or bad temporal locality. This thesis will also study the evolution of the frequent patterns found in the traces after modification of the application code or of the SoC architecture, in order to help determining the impact of the modification.