SFI: A Refinement Based Layered Software Architecture

Refinement calculus [1,5] is a formal framework for reasoning about program correctness and correctness preserving program refinements. It serves as the foundation for an object-oriented software architecture and construction method that we refer to as stepwise feature introduction (SFI) [3]. Characteristic for this approach is that each software module is described in terms of thin layers. Each layer extends the software with some new feature, in a way that preserves the features that have been introduced by earlier layers. This amounts to requiring that the new layer is a superposition refinement [4] of the layers below. The modules are interconnected using interface specifications, usually providing a more abstract view of the module state than what will actually be implemented. The implementation is required to be a data refinement [9,6] of the interface specification. SFI is based on structuring software with these two basic mechanisms, modularization and extension, while the refinement calculus provides the formal framework for reasoning about the correctness of software constructed in this way. We use UML [8] to describe the software architecture, with refinement calculus providing a rigorous mathematical semantics for the UML constructs employed. This will also give us a formal framework for reasoning about UML class diagrams, in essence using these as proof schemes when arguing about software properties. SFI goes well in hand with an extreme programming (XP) [7] approach to constructing software, with feature introductions corresponding to the iteration cycles in XP. We can look at SFI as providing a general software architecture for XP, and we can look at XP as providing a software process for SFI. We will report on some experiments where these two approaches have been combined [2]. We will also discuss in more detail how the notions of unit tests in XP can be generalized to requirement specifications in SFI.