UnTRIM parallelization with message passing

Jacek A. Jankowski. UnTRIM parallelization with message passing, 2nd UnTRIM Workshop, Sirmione, 23-25 May 2005.

The main ideas behind the parallelisation of UnTRIM applying the domain decomposition method and the message passing (MPI) are discussed. The issues concerning the mesh partitioning, point-to-point communication over the mesh-overlapping areas (halos) for purely triangular and mixed (triangles and quadrangles) meshes and the problems connected with the Lagrangian advection scheme are thoroughly described. The problem of the necessity of the code changes — mostly extending the field lengths to accommodate the halo cells/edges/nodes — is compared to the seemingly simpler OpenMP approach. The main idea of swapping the information about the halo cells/edges/nodes via the message passing is discussed in detail. The final part concerns the user-friendly way of running the simulations, including the stages of the mesh partitioning and result merging, with conclusions that the working prototype of the parallel code potentially needs only a few improvements presently not needed to introduce into the code for a daily project routine. Additionally, a few potential side projects are mentioned: weighted mesh partitioning, alternative “upwind-based” advection schemes and iterative linear equation solvers more adequate to the domain decomposition method.

The second part of the developments see the post from the next UnTRIM Workshop. For the parallel streamline tracking parallelisation presentation see the post from 2007.

Transparencies are available.