Parallelization strategies and experiences with the dynamically defined reaction path (DDRP) method

摘要

The possibilites and advantages of parallel realizations of the "Dynamically Defined Reaction Path (DDRP) method on large (Single Instruction Multiple Data-SIMD) and small (Multiple Instruction Multiple Data-MIMD) computer architectures. together with some runtime estimates and simulations, are discussed, preceded by a short theoretical introduction referring to the basic mathematical concepts. a description of the algorithm and a numerical realization of the general DDRP procedure. The main difficulty in getting optimal runtimes when using the method of small steps was found to be in the storage strategy of SCF data.