Editing RSF School and Workshop, Vancouver 2006 (section)

=== Roscoe Bartlett: An Overview of the Thyra Interoperability Effort for Abstract Numerical Algorithms within Trilinos ===

The Trilinos Project is an effort to develop and implement robust parallel algorithms using modern object-oriented software design, while still leveraging the value of established numerical libraries such as PETSc, Aztec, the BLAS and LAPACK. It emphasizes abstract interfaces for maximum flexibility of component interchanging, and provides a full-featured set of concrete classes that implement all abstract interfaces.  The number of Trilinos packages continues to grow along with the variety and sophistication of the numerical algorithms (and other numerical support software) contained in these packages. It is clear that the standalone use of these packages and algorithms is insufficient to solve tomorrows challenging multi-physics analysis and design problems. The Thyra effort seeks to develop a uniform set of Trilinos standard software interfaces and protocols by which any reasonable combination of Trilinos algorithms and other tools may be automatically supported; even configurations that the individual package and algorithm developers never even imagined. In this talk I will provide an overview of the Thyra effort discussing its context, requirements, history, current status, and future plans. Thyra is both a Trilinos package (named thyra) and a more general collaboration between Trilinos developers. The primary focus of Thyra is the support for abstract numerical algorithms (ANAs) such as iterative linear solvers (e.g. Belos), eigen solvers (e.g. Anasazi), nonlinear equation solvers (e.g. NOX), stability and bifurcation analysis (e.g. LOCA), ODE/DAE solvers (e.g. Rythmos), and constrained optimization (e.g. MOOCHO). The primary foundation for Thyra's ANA support is a minimal set of fundamental operator/vector interoperability interfaces expressed as abstract C++ classes. Using the foundation of these basic operator/vector interfaces, more sophisticated mathematical interfaces are being developed to address higher-level capabilities such as preconditioner factories, linear solver factories, nonlinear model evaluators, and nonlinear solvers.