SeisSol Optimization, Scaling and Synchronization for Local Time Stepping

The SeisSol code is a solver for the elastic wave equations developed in the past three years. It uses a novel technique, essentially a DG (Discontinuous Galerkin) method, called ADER-DG which is a high-order scheme using tetrahedra as computational cells. The code has been extended to many special cases interesting for seismologists e.g. anisotropy, viscoelasticity, poroelasticity and finite sources. It is written in FORTRAN 95 and uses meshes created externally with a variety of commercial software, mostly Gambit and ICEM CFD. Mesh partitioning is performed using the Metis software. Parallelization has been implemented using MPI (Message Passing Interface) libraries and the code is successfully working in different local clusters as well as the Munich Supercomputer HLRB2. A novel asynchronous explicit time stepping scheme can strongly reduce the overall number of iterations but makes the number of iterations performed by each processor time-dependent. We study the time evolution of the workload for each processor for a large-scale simulation setup and speculate on possible solutions.

@incollection{id1421,
  author = {de la Puente, Josep and K\"aser, Martin and Cela, Jose Maria},
  booktitle = {Science and Supercomputing in Europe},
  editor = {Monfardini, S.},
  isbn = {978-88-86037-22-8},
  language = {en},
  pages = {300-3022},
  publisher = {CINECA, Italy},
  title = {SeisSol Optimization, Scaling and Synchronization for Local Time Stepping},
  year = {2009},
}

%O Book Section
%A de la Puente, Josep
%A Käser, Martin
%A Cela, Jose Maria
%B Science and Supercomputing in Europe
%E Monfardini, S.
%@ 978-88-86037-22-8
%G en
%P 300-3022
%I CINECA, Italy
%T SeisSol Optimization, Scaling and Synchronization for Local Time Stepping
%D 2009