On peculiarities of using heterogeneous cluster architecture for solving continuum mechanics problems

The use of high-performance computing systems built on the basis of NVIDIA graphics processors is discussed. Several models of parallelism are compared, for using in the iterative Krylov subspace methods intended for solving large sparse systems of linear algebraic equations that arise in the numerical solution of a wide class of continuum mechanics problems described by partial differential equations. The numerical computations are based on the application of such technologies as MPI, OpenMP, Boost, and CUDA. The proposed methods are tested on the supercomputer "GraphIT!" installed at the Research Computing Center of Moscow State University when solving a problem of nonstationary flow of a fluid to the wells with complex trajectories in a three-dimensional domain.