Massively parallel computation on anisotropic meshes

In this paper, we present developments done to obtain efficient parallel computations on supercomputers up to 8192 cores. While most massively parallel computation are shown using regular grid it is less common to see massively parallel computation using anisotropic adapted unstructured meshes. We will present here two mains components done to reach very large scale calculation up to 10 billions unknowns using a muligrid method over unstructured mesh running on 8192 cores. We firstly focus on the strategy used to generate computational meshes and in particular anisotropic ones adapted to capture a quite complicated test function. Then we will briefly describe a parallel multigrid method. Performance test over a large range of cores from 512 to 8192 cores is then presented using the French national supercomputers Jade and Curie. The last section will present a calculation done on smallest number of cores on our own cluster, but using more realistic data obtain directly from experimentation. The goal is to be able to realize such kind of simulation on really complex micro structure obtain by tomography at a larger scale