Abstract: Kernels in the ADER-DG method, when solving the elastic wave equations, boil down to sparse and dense matrix multiplications of small sizes. At the example of the Earthquake simulations code SeisSol, we will investigate how this routines can be implemented and speeded-up by the code generation tool LIBXSMM. A long these lines we will analyze different tradeoffs of switching from sparse to dense matrix multiplication kernels and report performance with respect to time-to-solution and energy consumption. Bio: Alexander Heinecke studied Computer Science and Finance and Information Management at Technische Universität München, Germany. In 2010 and 2012, he completed internships at Intel in Munich, Germany and at Intel Labs Santa Clara, CA, USA. In 2013 he completed his Ph.D. studies at TUM and joined Intel’s Parallel Computing in Santa Clara in 2014. His core research topic is the use of multi- and many-core architectures in advanced scientific computing applications. In 2014, he and his co-authors were selected as Gordon Bell finalists for running multi-physics earthquake simulations at multi-petaflop performance on more than 1.5 million of cores.

Abstract: We developed an unstructured finite element mesh generation method capable of modeling multiple-material complex geometry problems for large-scale seismic analysis of soil-structure systems. We used an octree structure to decompose the target domain into small subdomains and use the multiple material marching cubes method for robust and parallel tetrahedralization of each subdomain. By using the developed method on a 32 core shared memory machine, we could generate a 594,168,792 tetrahedral element soil-structure model of a power plant in 13 h 01 min. The validity of the generated model was confirmed by conducting a seismic response analysis on 2,304 compute nodes of the K computer at RIKEN. Although the model contains a small approximation in geometry (half of the minimum octree size) at present, we can expect fast and high quality meshing of large-scale models by making geometry correction in the future, which is expected to help improve the seismic safety of important structures and complex urban systems.

Abstract: The efficiency and robustness of preconditioned parallel iterative solvers, based on domain decomposition for ill-conditioned problems with heterogeneous material properties, are evaluated in the present work. The preconditioning method is based on the BILUT(p,d,t) method proposed by the author in a previous study, and two types of domain decomposition procedures, LBJ (Localized Block Jacobi) and HID (Hierarchical Interface Decomposition), are considered. The proposed methods are implemented using the Hetero3D code, which is a parallel finite-element benchmark program for solid mechanics problems, and the code provides excellent scalability and robustness on up to 240 nodes (3,840 cores) of the Fujitsu PRIMEHPC FX10 (Oakleaf-FX) at the Information Technology Center, the University of Tokyo. Generally, HID provides better efficiency and robustness than LBJ for a wide range of values of parameters.