University of Illinois, Champaign-Urbana Progress Report 95/96
RECENT_96_ACCOMPLISHMENT
Improvements have been made in all phases of CAPSS, a fully parallel
solver for direct solution of large sparse positive definite systems.
Complementing the previous ordering, which was based on Cartesian
nested dissection using coordinate information, a new ordering
algorithm has been added which is based on graph compaction and does
not require coordinates. The use of two-dimensional matrix
partitioning and higher-level BLAS was investigated for enhancing the
performance and scalability of numeric factorization. A new selective
inversion scheme was developed that speeds up the triangular solution
phase by factor of 6 to 16 on typical test problems. This scheme,
which computes inverses of a sequence of submatrices of the factor and
uses these to replace substitution steps by more efficient distributed
matrix-vector multiplications, is a significant breakthrough for a
portion of the computation that was previously a substantial
performance and scalability bottleneck. Investiation also began on
improving the user interface to make the package easier to use in
applications.
FY1997_PLAN
The following new features will be integrated into a complete new
version of CAPSS: coordinate and non-coordinate orderings with improved
fill reduction and load balance, two-dimensional partitioning and
higher level BLAS in the factorization, and selective inversion in the
triangular solution. Communication will use
MPI for improved
portability, and performance will be made tunable for a given platform.
TECHNOLOGY_TRANSITION
CAPSS has been used in solving numerous problems in structural
mechanics, including shearing in foam-like materials and crack
propagation in extrusion processes.