CSE Software design for scientific software is an intricate process composed of multiple steps - including determination of what capabilities are needed to meet the targeted scientific goals to understanding how to balance tradeoffs between modularity, functionality and performance in the resulting software product.
Scientific software incorporates a broad range of expertise ranging from domain knowledge and applied mathematics to computer science issues such as performance and portability. It is necessary to untangle various complexities so that experts can focus on what they know best. Therefore, the process of software design in the context of scientific software involves determining a union of capabilities needed to achieve scientific objectives, the decomposition of needed capabilities into components that encapsulate a functionality and an ability to compose these components as needed by the application. Decomposition can be functional and data/spatial domain. There are several examples of application codes and frameworks where sections of the code that pertain to the science of the application are written and verified separately and then plugged into a composing framework with a wrapper. The internals of the framework handle the data management and other infrastructural concerns such as parallelization if needed.
Therefore, one of the very first choices to make in designing a software is which abstractions to use, and how to provide footholds for the abstractions in the framework so that they maintain separation of concerns. There are many long standing obstacles to meeting requirements above. There has always been a trade-off between modularity and performance. Modularity is necessary for maintainable and reusable code, but it compromises performance. Some of the conflicting requirements require hard-nosed trade-offs in design. For example, most composable codes run relatively slowly because they sacrifice performance for multiphysics and composability. Sometimes this means using suboptimal options for individual components. The framework should be able to facilitate such unorthodox approaches and therefore should provide hooks for being able to make these choices.
Another important consideration in scientific software design is its dynamic nature. Codes designed for one problem are routinely modified to use other similar problems, or need modification because growing understanding places more demands on the current model. Therefore, extensibility in also a very important aspect of scientific software design.
