Scientific Software Sustainability: The Numerical Reproducibility Challenge

Published

October 15, 2015

Author(s)

Walid Keyrouz, Michael V. Mascagni

Abstract

Experimental reproducibility is a cornerstone of the scientific method. The ease of achieving its counterpart in computing, numerical reproducibility, was one of the core assumptions underpinning the growth of scientific computing over the past several decades to become a powerful tool for scientific inquiry that is now widely considered as the ``third leg of science''. The other core assumption was the deterministic behavior of computer hardware. Unfortunately, these assumptions are currently being challenged by hardware developments over the past several decades as discussed and documented in recent reports and workshops. In this position paper, we are advocating for the following actions: - Redefine numeric reproducibility by considering numeric results as computational measurements and treat them as the equivalent of physical measurements. - Further develop and disseminate algorithms that sustain the newly redefined reproducibility. - Bring approaches used in stochastic algorithms and Monte Carlo Methods to bear on problems more traditionally associated with deterministic approaches. - Carry out simulations as ensemble runs as is becoming the norm in weather forecasting. In this context, we are organizing a workshop on ``Numerical Reproducibility at Exascale'' at SC'15 in Austin, TX, with the goals of bringing a community together to address the issue of Numerical Reproducibility.

Citation

Computational Science & Engineering Software Sustainability and Productivity Challenges (CSESSP Challenges)

Pub Weblink

https://www.nitrd.gov/csessp/

Pub Type

Websites

Keywords

Numerical Reproducibility, Software Sustainability, Uncertainty Quantification

Citation

Keyrouz, W. and Mascagni, M. (2015), Scientific Software Sustainability: The Numerical Reproducibility Challenge, Computational Science & Engineering Software Sustainability and Productivity Challenges (CSESSP Challenges), [online], https://www.nitrd.gov/csessp/ (Accessed November 21, 2024)