Generic programming for scientific computing in C++, Java, and C#

Parameterized types (generics) have been announced for the Java(TM) and C# programming languages. In this paper, we evaluate these extensions with respect to the realm of scientific computing and compare them with C++ templates. At the heart of our comparison are the set and relation classes from the Janus framework which provides abstraction for the efficient representation of meshes, sparse graphs and associated matrices. As an example, we compare the performance of the Bellman-Ford single source shortest path algorithm when using parameterized Janus data structures implemented in C++, Java(TM), and C#. Our measurements suggest that both Java(TM) and C# generics introduce only little run time overhead when compared with non-parameterized implementations. With respect to scientific application, C# generics have the advantage of allowing value types (including builtin types) as parameters of generic classes and methods.