Tools for the Development of Memory-Efficient Sparse Linear Solvers

University Of Colorado At Boulder, Boulder CO

Investigators

Abstract

Tools for the Development of Memory-Efficient Sparse Linear Solvers Recent advances in both hardware and algorithm development have increased the effect of memory access on the runtime of large scale linear algebra applications. The PI proposes to create tools that will provide the code developer with information about memory costs of codes in preparation as well as with software building blocks designed to lessen those costs. First, a framework for memory-efficient implementation of the sequences of sparse Basic Linear Algebra Subprograms that typically constitute a sparse linear solver will be provided. Second, a collection of metrics for the prediction of memory usage of linear algebra software will be developed. Working from those metrics, a tool to automate the prediction process will be built. Finally, keeping in mind that the performance of a code ultimately depends on how it compiles, a comprehensive study of compilers and the performance of the executables they produce will be carried out.