VERROU: Assessing Floating-Point Accuracy Without Recompiling

As an industrial facility relying on numerical simulation to improve the safety and efficiency of its electricity production units, EDF is committed to ensure that all the numerical simulation codes it develops and uses are correctly validated and verified. Within this context, the accuracy of floating-point operations has progressively become one of the important topics to study, especially since computing codes are exploited on ever more powerful hardware to solve ever larger problems. The Verification and Validation (V&V) process should therefore include the monitoring of inaccuracies introduced by floating-point arithmetic, as well as the verification that they are kept within acceptable limits. In this paper, we present the verrou tool, a valgrind-based system which uses Monte Carlo Arithmetic (MCA) to monitor the accuracy of floating-point operations without needing to instrument the source code or even recompile it. This tool is therefore well-suited to be part of an industrial V&V process. It has been successfully tested both on small-scale, well understood numerical applications, and on large-scale, more complex industrial computing codes.