We have developed an object-based linear algebra package, currently
with emphasis on sparse Krylov methods, driven primarily by needs of
the Los Alamos National Laboratory parallel unstructured-mesh casting
simulation tool
TELLURIDE.
Support for a number of sparse storage formats,
methods, and preconditioners have been implemented, driven primarily
by application needs.
We describe our object-based
Fortran 90 approach, which enhances maintainability, performance, and
extensibility, our parallelization approach using a new portable
gather/scatter library (
PGSLIB), current capabilities and future
plans, and present preliminary performance results on a variety of platforms.
