Section: New Results

Timing refinement for multidimensional dataflow models using MARTE Time Model

Participants : Frédéric Mallet, Julien Deantoni, Jean-Vivien Millo.

Extensions of dataflow process networks have been proposed (as multidimensional SDF) to combine task parallelism (as in traditional process networks) with intensive data parallelism (as proposed in the Array-OL/Gaspard2 formalism developed in the DaRT EPI, for instance). The prospect of scheduling (seen as precise time cycle allocation) is here more complex, because of possible trade-offs between the granularity of treatments at task level vs. the size of data arrays that are handled uniformally in parallel inside each task. We considered how these phenomena could be represented (if not solved) inside the framework of MARTE Time Model and logical clocks, so as to handle such design issues in a well-defined MDE approach. Additionally, we used the MARTE platform description to specify how the previous models are refined through mapping allocation. The resulting modeling framework was presented in a journal article [19] . This work was conducted jointly with P. Boulet, from DaRT EPI, and C. Glitia, former DaRT PhD and Aoste postdoc student.