Slicing UML's Three-layer Architecture: A Semantic Foundation for Behavioural Specification

Abstract

One of the main notational contexts in which model-driven software development has been studied is the Unified Modeling Language (UML), the de facto standard in software modelling. The current trend in software development is not just towards the use of models, but the use of executable models. In 2006, the Object Management Group issued a Request for Proposal (RFP), soliciting the definition of an Executable UML Foundation, with a fully specified executable semantics. The purpose of such a version of UML is to make the advantages of executable models available to UML users by enabling "a chain of tools that support the construction, verification, translation, and execution" of models.

An oft-voiced criticism of UML is its lack of a formal, unambiguous description of its semantics. In an effort to improve the support for model-driven development, especially with respect to executable modelling, the UML 2 specification introduced a novel three-layer semantics architecture. This architecture provides a stratification of the description of UML models that clearly separates 'low-level' behavioural specification mechanisms, such as actions, from 'high-level' behavioural formalisms, such as activities, state machines and interactions. Although UML describes the effect of actions, it does not provide either the concrete syntax or the formal semantics of an action language.

Our research focuses on a top-to-bottom slice of the three-layer architecture. We formally define the execution semantics of two-thirds of UML actions, including the most complicated actions---invocation actions. Our formal definition is expressed in terms of state changes to a global state machine representing an executing UML model. Our work provides an alternate formalization to that of the current submission to the RFP and could be used to enhance that submission.

To validate our formal semantics and to determine the usefulness of the three-layer architecture, we have created an interpreter for UML actions and activities. This interpreter was designed in accordance with the complex token passing semantics of UML and provides analysis capabilities that have been successfully used to identify problems even in published activity diagrams. In effect, we have created a tool that supports the construction, verification and execution of a subset of UML models, namely activities. Our handling of this slice of the three-layer architecture is a preliminary step to realizing the grander vision of general executable (and analyzable) models.