Abstract: |
The considered non-conforming FEM and MIC(0) factorization are robust for problems with possible jumps of the coefficients. The numerical tests confirm robustness of the point-wise MIC(0) preconditioner. Estimates for the parallel times, speed-ups and efficiencies are derived. The algorithm is implemented in C++ using the Message Passing Interface (MPI) library. Experiments on several distributed computing systems are presented. The results agree well with the theoretical analysis. It is shown that computations and communications can be overlapped in time, thus improving the parallel efficiency. |