EXPERIMENTAL STUDY ON SKIF MSU CHEBYSHEV OF THE COMBINED MPI+THREADS IMPLEMENTATION OF THE ALGORITHM FOR SOLVING SYSTEMS OF LINEAR EQUATIONS ARISING IN FLOWVISION WHEN SIMULATING COMPUTATIONAL FLUID DYNAMICS PROBLEMS |
| 1 | |
| 2010 |
| scientific article | 519.612 | ||
| 202-209 | system of linear equations, iterative methods, incomplete triangular factorization, parallel computations, MPI, common memory, computational threads |
| Simulation of CFD problems in FlowVision by implicit numerical methods requires the solution of systems of
linear equations described by large unstructured sparse matrices. In FlowVision the solution algorithm for such systems
of linear equations includes the computation of the high order incomplete triangular factors (ICH2/ILU2) of the
coefficient matrix followed by calculations using the algorithm of the preconditioned Krylov subspace iterative method.
In order to reduce the solution time, the algorithm is a priori parallelized over the distributed memory using the
MPI standard, while over the common memory it is parallelized by using Intel® Threading Building Blocks (TBB)
with automatic dynamic load balancing. For extracting explicit parallelism in dependent computations and for solving
parallel systems of equations with triangular matrices, the nested dissection (ND) ordering algorithm is used. The
numerical experiment results on scalability of the proposed algorithms performed on SKIF MSU Chebyshev demonstrate
their efficiency in simulating CFD problems in FlowVision. The results of numerical experiments performed
on SKIF MGU Chebyshev demonstrate the efficiency and parallel scalability of the proposed algorithms when
simulating sample CFD problems in FlowVision. |
 |