Computation operations caching for numerical repeatability

How can we track numerical variations during a simulation execution? How can we guarantee identical results when trying to repeat a simulation? These questions are fundamental in any viable attempt to solve computational results repeatability. The current solutions approach these concerns through approximation methods that are meant to keep the final result within a certain error scale. In this paper, we propose a centralized cache based approach that is the actual dump of all the numerical operations computed during the simulation execution. Thus, this cache can be used as a witness to track all the numerical variations down to operations level. This allows the finest grain of debuggability in case of numerical errors, which is essential. Furthermore, we can guarantee identical results by simply reloading all previous operations results when the new execution attempts to recompute them again.