A Generalized Framework for Abstraction and Dynamic Loading of Numerical Solvers

Scientific software often relies on numerical solvers for tasks such as integration of differential equations, optimization and finding roots of linear or non-linear systems. Typically, the availability of only one solver for a certain task is not sufficient in generic software systems, since each solver usually only has a limited area of application. However, the inclusion of multiple solvers into complex industrial quality software is often cumbersome. One issue is the fact that solver codes are often implemented in different programming languages, other issues are related to a lack of standardization of the methods provided by the codes, fixed I/O routines and disregard of thread-safety concerns. This article describes how some of these issues were handled in the scope of Tornado, an advanced kernel for modelling and virtual experimentation (i.e., any evaluation of a model). Particularly important is that in view of the maintainability and extensibility requirements that are often imposed by complex scientific software, there is a need for abstraction and dynamic loadability of solvers.