|
|
|
| A parallel sweeping preconditioner for heterogeneous 3D Helmholtz equations | |
|
Next: Parallel Sweeping Preconditioner (PSP)
Up: Parallel sweeping preconditioner
Previous: Parallel preconditioned GMRES(k)
In order to implement the previously discussed techniques for scalable
multifrontal factorizations and solves (via selective inversion), an
open-source distributed multifrontal solver named Clique was built on top of
Elemental (30), a library for distributed-memory dense
linear algebra. In addition to being designed to support the techniques
we discussed above: selective inversion, subtree-to-submesh mappings, and
two-dimensional frontal matrix distributions, it was also written with
a strong emphasis on memory scalability. This is because the
sweeping preconditioner requires large numbers of factorizations
of relatively small sparse matrices, and so it is crucial that the per-process
memory usage for each subdomain factorization decreases inversely with the
total number of processes.
We note that Clique was designed specifically to provide a memory-scalable
multifrontal implementation for our parallel sweeping preconditioner, and so
there is not yet support for pivoting. We plan to add a pivoted LU factorization
in the near future.
2014-08-20