The CIVA software developed by CEA LIST offers different modules for simulation and treatment for non-destructive testing (NDT). In this context, a high-order Spectral Finite Element code coupled with a domain decomposition approach has been implemented.

The subject of this thesis consists in parallelizing this code in order to reduce computation time which can be very important especially in 3D. For this, on one hand, different levels of parallelization shall have to be addressed: by distributing the computation of the different domains on several multi-core machines, and by performing intra-domain numerical computation with help of vectorized code, i.e., SIMD on CPU or MIC, SIMT on GPU. On the other hand, a computation time optimum shall try to be reached while maintaining a good level of precision. This shall require to adapt the different parameters of the simulation and to place the code execution on the different hardware compute devices with help of both a computation and memory cost model and an auto-tuning procedure at runtime start.

The thesis will take place as part of a collaboration with the Maison de la Simulation.

This position is open until it is filled.

Department: Département Imagerie Simulation pour le Contrôle (LIST)
Laboratory: Laboratoire Développement Informatique
Start Date: 01-10-2015
ECA Code: SL-DRT-15-0516
Contact: gilles.rougeron<στο>cea.fr