EAGER: Collaborative Research: Lighthouse: A User-Centered Web System for High-Performance Software Development

University Of Colorado At Boulder, Boulder CO

Investigators

Abstract

The role of computing in science and engineering has been growing steadily in recent years, leading to ever larger and more complex problems. Their solution requires high-performance resources, which has resulted in many efforts to produce new mathematical approaches, programming models and languages, software libraries, and runtime environments for advanced parallel computers. This vast and growing number of solution methods presents a challenge to scientists who do not want to invest the time needed to locate and learn new computational tools. The PIs are developing Lighthouse, a taxonomy-based system, that equips practitioners with easy discovery and use of high-performance software solutions to a variety of common problems arising in scientific and engineering applications. Lighthouse also integrates information about the performance of multiple implementations of numerical algorithms. It thus allows users to readily find and start using the best method for addressing a particular challenge with the computers available to them. Less time spent in learning numerical packages and faster scientific simulations directly increases scientific productivity. Lighthouse is a web-based framework that offers a solution to two main research challenges. First, it aids in the creation of complex high-performance applications that leverage the latest advances in applied mathematics and computer science research. Second, it facilitates effective intra- and cross-domain communication for computer scientists, applied mathematicians, computational scientists, and students as they tackle scientific and engineering computing problems. Lighthouse incorporates an extensible taxonomy that enables different types of searches for numerical solutions depending on a user?s level of expertise. It then provides a mechanism for generating code templates based on the search results. Lighthouse thereby reduces the substantial learning curves associated with using state-of-the art high-performance software libraries. The Lighthouse system integrates taxonomy information for an expanding number of high-performance numerical libraries that presently comprises the linear algebra packages LAPACK, PETSc, and SLEPc. Lighthouse employs machine learning methods to automatically classify routines based on their performance attributes. In this way, it helps to provide more accurate search results that take into account not only functionality but also problem scale, performance, and available computational resource requirements. Finally, Lighthouse improves communication within and between communities by providing an accessible web-based interface to taxonomies of numerical packages that includes documentation, references, and performance information.