Efficient Parallelization of MATLAB Stencil Applications for Multi-core Clusters

This paper presents the automatic parallelization of Stencil codes written in MATLAB for distributed systems. The compiler translates MATLAB source into C code and automatically parallelizes using MPI. For clusters of multi-cores, also a hybrid approach is presented where the generated code combines MPI and OpenMP parallelization. The parallelization concepts are evaluated with two stencil computation benchmarks. The automatically generated code not only outperforms the MATLAB code, but shows also a very good scaling.