A Portable Scalable Library for Sparse Matrix Linear Algebra on Distributed Memory Systems

This proposal describes the development and analysis of a prototype scalable library for sparse matrix operations in a distributed memory environment. The library is to be object-oriented; hence, will be implemented in C++. It is designed to be portable, and therefore will adhere to the Message Passing Interface (MPI) standard, but will use P4 from Argonne National Laboratory for the prototype. Finally, the class library will not sacrifice performance, and will take advantage of prescheduled communication to minimize communication costs. The class library will supports creation, deletion, updating, and remapping of matrices, matrix-vector multiply, dot product, SAXBY and DAXBY vector operations. A PCG BOlver and a GMRES solver will be implemented using the libraries, and their performance will be compared to existing solvers. Anticipated Benefits/Potential Applications - A portable, object-oriented linear algebra library will permit users to develop easy to use, machine-independent linear solvers for use in a wide variety of scientific computing applications and environments.