System Software R&D
System Software R&D
The objective of this task is to develop a robust, scalable, and portable parallel Adaptive Mesh Refinement (AMR) software library on unstructured 2-D and 3-D meshes for unstructured scientific and engineering computing applications. Targeted applications for this tool include finite-element and finite-volume based numerical applications, both inside and outside the NASA community, and some advanced scientific visualization applications using unstructured grids.
A set of robust and efficient parallel AMR algorithms was designed for parallel adaptive refinement of triangular meshes and tetrahedral meshes. A robust scheme for mesh quality control was incorporated into the parallel adaptive refinement process and has been successfully implemented and tested on triangular meshes. An efficient parallel scheme was designed and implemented for mesh migration and dynamic load-balancing. The software framework of our unstructured parallel AMR library is based on Fortran 90; its core data structure is a set of user-defined data types in Fortran 90. An efficient parallel graph partitioner, the ParMeTiS software developed at the University of Minnesota, is utilized for mesh partitioning and mesh remapping in the dynamic load-balancing process.
A complete parallel AMR software library for triangular meshes has been completed and is now available. The library contains modules for operations in a typical parallel AMR process, including mesh input and output, parallel adaptive refinement, mesh quality control, parallel mesh partitioning and migration for dynamic load-balancing. This 2-D library has been successfully tested on the CRAY T3E system at Goddard Space Flight Center using a few finite-element meshes from electromagnetic scattering problems and semiconductor device modeling problems. A sequential AMR library was completed for tetrahedral meshes and successfully tried on the CRAY T3E using a few test meshes. The ParMeTis package has been applied successfully to tetrahedral meshes for parallel partitioning. A Web site has been established to reflect the current state of this work.
Adaptive mesh refinement (AMR) is an important numerical technique for efficiently solving a variety of large-scale scientific and engineering computing problems. The benefits of using this advanced numerical technique to dramatically speed up the time to solution of numerical simulation problems are being increasingly recognized by scientific and engineering communities. The AMR technique, when applied successfully, not only saves a tremendous amount of computing time, but also makes the use of computer memory much more efficient. This makes it possible to solve numerical simulation problems of sizes that would otherwise be impossible to solve on a uniform mesh, even on massively parallel multiprocessor systems. The primary impact of this work will be the availability of a practical parallel unstructured AMR software tool implemented in a programming language designed mainly for scientific computing applications (Fortran 90).
The parallel unstructured AMR library for triangular meshes is complete. A sequential version of AMR on tetrahedral meshes is also complete. A parallel version of AMR for tetrahedral meshes is currently being developed. Our goal is to finish the parallel implementation of AMR for tetrahedral meshes, and integrate the triangular and tetrahedral parts into a unified AMR software library package with consistent data structures and interfaces to 2-D and 3-D application codes. Some performance analysis and optimization of the parallel AMR software will then be conducted. We also plan to make the parallel AMR package available on other multiprocessor systems, such as the HP Exemplar, IBM SP2, and PC or workstation cluster systems (e.g. Beowulf systems).
John Z. Lou
NASA Jet Propulsion Laboratory
lou@acadia.jpl.nasa.gov
818-354-4870
Charles D. Norton
NASA Jet Propulsion Laboratory
Charles.D.Norton@jpl.nasa.gov
818-393-3920