Collaborative Research: CIRC: Medium: New: Workbench for Reliable and Efficient Numerics

University Of Utah, Salt Lake City UT

Investigators

Abstract

This project aims to build a workbench for scientists and engineers to address numerical issues in real-world applications. Numerical issues are issues caused by the gap between mathematical (real) numbers and the number representations used on computers, like floating point. Ultimately, this gap makes it difficult for scientists and engineers to develop software that does numerical computation accurately and runs reliably and efficiently on a variety of hardware and software platforms. Over the years, the research community has studied these issues and developed a number of tools that make developing numerical software easier, but these tools have become difficult to use together. In this project, the investigators will develop a set of standards, benchmarks, and user interfaces to make these existing tools interoperable and thus easier to use in concert. The project’s novelties are a set of standards where floating-point computations can be connected to the hardware and software platforms they run on, along with observed bad inputs or bugs. The project’s impacts are its potential improvements to real-world software packages, making them faster and more reliable across a wide range of hardware and software. Additionally, the investigators plan a variety of community-building initiatives including a community meeting, workshops, and REUs to build further ties within the numerical research community and between that community and practitioners in industry, national laboratories, and academia. The project builds on the existing FPBench standardization and interoperability effort. That standardization effort largely focuses on unambiguously describing floating-point computations, but real-world numerical workflows must track much more information: representative inputs; platform characteristics; pointers into codebases; and error bounds, observed or proven. This project will extend the FPBench standard with new formats to record and transmit this additional information, and update a variety of existing, widely-used numerical tools to use the new format. The investigators will then collect more benchmarks from real-world applications, recording rich metadata descriptions using the new formats, and distribute the extended tools and new benchmarks. To tie these standards, tools, and benchmarks together, the investigators will develop a novel, task-oriented user interface for scientists and engineers dealing with numerical issues. This interface will dispatch individual tools and collect the information they generate (in the new standard formats) in a single database, transparently passing the necessary information to every tool and informing the user when running additional tools would be useful. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.