On the reproducibility of numerical experiments with FEniCS

Paul Garlick (Tourbillion Technology Ltd, 🇬🇧)
Monday session 3 (Zoom) (17:00–18:30 GMT)
View slides (pdf) (available under a CC BY 4.0 license)

A key tenet of the scientific method is that an experiment conducted in one laboratory may be reproduced in another. For numerical experiments the process should be straightforward; simply a case of supplying the same inputs to the same software application. However, for software applications with complex hierarchies the essential step of re-creating the same software environment can represent a significant challenge. Difficulties can emerge in installing library dependencies on computer systems with different architectures. Even on the same system changes to low-level libraries over time can prevent compilation or execution of higher-level code. To address these issues methods from functional programming have been developed to guarantee bit-identical software installations. Package definitions have been created for FEniCS and its dependencies within the GNU Guix framework. Used as a package manager, GNU Guix provides the tools to manage a dedicated file structure that sits alongside the host operating system. This presentation shows how the underlying mechanism works, how FEniCS fits in and how to create software environments that can be reproduced identically in space and time.