On the Reusability and Numeric Efficiency of C++ Packages in Scientific Computing

In this paper, we discuss the reusability and numerical efficiency of selected Object-Oriented numerical packages, serial and parallel, for developing high performance scientific computing applications. We selected packages that can be obtained freely on the internet and most of them are open source. Since the authors did not have extensive previous experience with all the packages, we believe that we approximately reproduced the typical process that an experienced OO programmer undergoes to apply the fundamental OO concepts to component reusability and general programming in new and existing OO scientific computing applications. We attempted to apply these concepts in the selection of numerical containers, defined as class with the main purpose of holding objects, such as vectors and matrices that could be both efficient and well tested for dense matrix operations. Our results indicates that in general serial OO packages still lags behind well-tuned non-OO libraries that supply BLAS type operations. Surprisingly, simple containers from the Standard Template Library (STL) outperformed most of OO libraries that used more sophisticated techniques such as expression templates. Careful wrapping of non-OO libraries seems to be the best way to obtain good performance and reusability. Most parallel OO packages use this approach rather than reimplementing BLAS operations natively. Thus they are more attractive in terms of performance and reusability than their serial counterparts.