A new parallel direct solver for 2D frequency-domain modeling

The parallel direct solver is important for 2D frequencydomain modeling. It can effectively reduce the memory requirement and computing time. And multi-shot seismic data can be simulated efficiently in this way. However the traditional parallel sparse LU direct solvers, such as MUMPS and superLU, have poor speed-up performance. In this paper, we introduce a new parallel direct solver called banded-block cyclic reduction(BBCR), which is developed to get a better speed-up performance and reduce the computing time. Numerical experiments for Marmousi model show that the speed-up ratio of BBCR is obviously superior to the speed-up ratio of MUMPS, and the frequency-domain wavefields got by BBCR and MUMPS are almost the same. Besides, since BBCR makes full use of banded matrix operation, it can almost reduce the computing time of traditional block cyclic reduction(BCR) solver by half.