SHF: Small: Collaborative Research: Automated Numerical Solver EnviRonment (ANSER)

University Of Oregon Eugene, Eugene OR

Investigators

Abstract

The computational science community is tackling ever larger and more complex applications. The solution of the underlying mathematics problems requires using high-end parallel computing resources effectively, and delivering performance without degrading productivity is critical for the success of scientific computing. Converting mathematics from algorithms to high-quality implementations, however, is a difficult process, whether an application is developed from scratch or by leveraging existing software libraries. Modern numerical packages provide numerous solutions with widely varying performance. Selecting among these possibilities requires expertise in numerical computation, mathematical software, compilers, and computer architecture, but even such broad knowledge does not guarantee the selection of the best-performing method for a particular problem. In response to these challenges, ANSER (Automated Numerical Solver EnviRonment) automates the selection and configuration of algorithms such as sparse linear solvers, eigensolvers, and graph methods in the context of large-scale scientific and engineering applications. The overall approach is generalizable to any situation involving multiple solutions whose performance varies with input problem properties. ANSER increases developer productivity and promotes effective use of modern parallel architectures to solve large-scale scientific and engineering problems. This work also impacts the training of the next-generation scientific workforce by involving graduate and undergraduate students in this model-guided development of high-performance software. ANSER, the Automated Numerical Solver EnviRonment, is an open-source web-based platform that supports the development of both scientific applications and high-performance libraries. It selects, configures and, in some cases, generates implementations of high-performance numerical algorithms. ANSER defines a methodology for automating the process of identifying problem features, creating performance models (based on combining analytical and machine learning approaches), and employing them in creating and configuring numerical software. ANSER initially targets widely used numerical packages for nonlinear partial differential equations and solution of eigenvalue problems, but it is designed to be extensible to other types of numerical methods, such as graph computations and n-body simulations. In addition to traditional dissemination methods (open-source software releases and publications), ANSER integrates semantic analysis of scientific computing literature to discover numerical methods similar to those provided by the target libraries and to identify and connect with our users. ANSER provides multiple interfaces to support different types of users, including students, computational scientists, and numerical library developers.