Parallel Iterative Solvers for 3D Magnetic Field Problems

In this talk, we discuss a general framework for the formulation and implementation of parallel iterative solvers. The parallelism is based on distributing the initial data, e.g., elements of a mesh, providing a distribution of derived objects, e.g., matrices. In this way, most of sequential iterative algorithms can be formulated and implemented in parallel very naturally with minimal communication overhead providing high parallel efficiency. Moreover, using operator overloading of C++ allows to use identical code sequences, e.g., for the cg-solver, for both, sequential and parallel versions. Numerical results obtained on a SGI Origin 2000 with 60 processors are presented.