A parallel sweeping preconditioner for heterogeneous 3D Helmholtz equations |
For performance reasons, it is beneficial to solve as many right-hand sides
simultaneously as possible: both the communication latency and the costs of
loading the local data from frontal and sparse matrices from main memory
can be amortized over all of the right-hand sides. Another idea is to extend
the so-called trsm
algorithm for triangular solves
with many right-hand sides (i.e., more right-hand sides than processes),
which is well-known in the field of dense linear
algebra (30), into the realm of sparse-direct solvers via
the dense frontal triangular solves.
This approach was not pursued in this paper due to the modest storage space
available on Lonestar and is left for future work.
Another performance improvement might come from exploiting block variants of
GMRES (36), which can potentially lower the number of
required iterations.
A parallel sweeping preconditioner for heterogeneous 3D Helmholtz equations |