Alternating Direction Implicit Methods on Distributed and Shared Memory Parallel Computers
High Performance Computing Systems and Applications, 1998
Global scale simulation codes capable of modeling the interaction of the solar wind with the Eart... more Global scale simulation codes capable of modeling the interaction of the solar wind with the Earth’s magnetosphere are being developed by the international space community. We shall describe different implementations of a 3D alternating direction implicit (ADI) magnetohydrodynamic (MHD) simulation code on distributed and shared memory parallel and parallel/vector computers. The specific architectures considered are the IBM SP and Silicon Graphics Origin, and techniques appropriate to Cray-class computers. On the distributed memory IBM SP, the ADI code was run under PVMe and MPI, and gave good scalability under a single program multiple data (SPMD) implementation. Scalability on the SGI Origin was approximately linear up to the available configuration size of 128 processors, with a change in slope at 32 processors. On Cray-class computers, very high efficiencies are possible when vectorization is combined with parallel processing. It is shown that while this leads to near optimum performance, it is achieved at the expense of increased memory requirements.
Uploads
Papers by Robert Rankin