## Section: New Results

Keywords : High-Performance Computing (HPC) (high-performance computing, high-performance computing, hpc), parallel computing, Message-Passing Interface (MPI), seismic wave equation.

### High Performance Computing for the seismic wave equation at very high resolution

Participants : Dimitri Komatitsch, Roland Martin, David Michéa, Nicolas Le Goff, Pieyre Le Loher.

With the very rapid evolution of personal computers, computer clusters, and supercomputers, nowadays the seismic wave equation can be solved with very good accuracy using very precise techniques implemented based on parallel computing in the context of so-called High-Performance Computing (HPC). This has been a central part of our research activity in the last few years and even increasingly more in 2008. In particular with some colleagues from the CINES supercomputing center in Montpellier (France) we have performed some very large scale calculations that are currently being analyzed and published.

Using the high-order finite-element method implemented in our SPECFEM3D software package, which we reviewed in [29] , in the context of a collaboration with some colleagues in Taiwan we for instance studied the influence of topography modeled at very high resolution on seismic wave propagation in the region of Taipei in Taiwan [23] , [24] , [22] .

In the context of collaborations with some colleagues in Barcelona (Catalonia, Spain), in San Diego, Caltech and Princeton (USA) and in Japan, we also applied the technique to model seismic wave propagation at very high frequency in the whole Earth [20] , [30] . In November 2008 our SPECFEM3D software package was again among the six finalists of the pretigious Gordon Bell Prize of the SuperComputing'2008 conference in the USA [34] for a calculation performed in parallel on 150,000 processor cores, reaching a sustained performance level of 0.16 petaflops.

In the context of a collaboration with Gordon Erlebacher from Florida State University (USA) who visited us for a month in May-June 2008 we ported our modeling algorithm to a NVIDIA graphics video card (Graphical Processing Units – GPU) using the CUDA language on top of a C implementation of our code. This technique is known as General-purpose Processing on Graphical Processing Units (GPGPU) and had never been used before for a high-order finite-element technique, which induces significant technical problems in particular regarding memory accesses. In [21] we used it to improve the speed of our code by a huge factor of 25.

Finally, in [16] we used a finite-element code to model stress redistribution in Island following a large earthquake that occurred there in June 2000 and study how changes in the stress field could have had an influence on the triggering of a second earthquake that occurred in the same region a few days later.