With significant research going into the development of scientific software over the years, there exist a plethora of toolkits using different algorithms to solve the same problem. But the performance of these toolkits are very much problem specific and depend on multiple factors, including experimental setup and hardware configurations. This makes it very difficult to choose a suitable software beforehand without testing them for specific problems while the wrong choice of software contributes to severe performance downgrade. In this thesis, we address this selection challenge by proposing a faster and reliable model-based approach instead of empirically running time-consuming experiments to select suitable software toolkits. In specific, we would be looking at selection for two classes of algorithms that solve parallel graph processing applications and systems of linear equations.Appropriate metrics have also been introduced to evaluate the quality of selection techniques.