The International Conference for High Performance Computing, Networking, Storage and Analysis
Floating-Point Robustness Estimation by Concrete Testing.
Student: Wei-Fan Chiang (University of Utah)
Supervisor: Ganesh Gopalakrishnan (University of Utah)
Abstract: Analyzing floating-point imprecision is required for dealing with the performance-precision trade-off of numerical programming. However, without specific mathematical knowledge, this is hard to be done by general developers.
We focus on solving one of the many floating-point precision issues, robustness estimation, by concrete testing. Programs without sufficient robustness tend to generate unexpected outputs, but developers cannot tune their programs without knowing the robustness degrees of candidate routines. Our testing driven framework estimates robustness by enumerating the inputs causing non-robust behavior. The number of problematic inputs indicates robustness degree and provides a comparison basis between candidate routines. We evaluated our framework with two implementations of a geometric primitive: one of the two implementations is pointed out to be more robust than another.
This empirical result matches the robustness argument in the previous context.
Thus we strongly believe that it is worthy to further investigate concrete testing based robustness estimation.