Seminar details

Abstract: Distributed software is often required for new systems, because of efficiency and physical distribution and sensors and actuators. Ensuring correctness of a distributed implementation is hard due to the interleaving of actions belonging to distinct processes. This thesis proposes a method for generating a correct and efficient distributed implementation from a high-level model of an application. The input model is described as a set of components communicating through prioritized multiparty interactions. Such primitives change the state of all components involved in an interaction during a single atomic execution step. We assume that a distributed implementation is a set of processes communicating through asynchronous message-passing. The main challenge is to produce a correct and efficient distributed implementation of prioritized multiparty interactions, relying only on message-passing. The method relies on a rigorous design flow refining the high-level model of the application into a low-level model, from which code for a given platform is generated. All intermediate models appearing in the flow are expressed using the same semantics as the input model. Complex interactions are replaced with constructs using simpler interactions at each step of the design flow. In particular, the last model obtained before code generation contains only interactions modeling asynchronous message passing. The correctness of the implementation is obtained by construction. Using multiparty interaction reduces drastically the set of reachable states, compared to an equivalent model expressed with lower level primitives. Essential properties of the system are checked at this abstraction level. Each transformation of the design flow is simple enough to be fully formalized and proved by showing observational equivalence or trace equivalence between the input and output models. The obtained implementation is correct with respect to the original model, which avoids an expensive a posteriori verification. Performance can be optimized through adequate choice of the transformation parameters, or by augmenting the knowledge of components. The latter solution requires to analyze the original model to compute the knowledge, that is reused at subsequent steps of the decentralization. The various transformations and optimizations constituting the design flow have been implemented using the BIP framework. The implementation has been used to evaluate the different possibilities, as well the influence of parameters of the design flow, on several examples. The generated code uses either Unix sockets, MPI or pthreads primitives for communication between processes.



Je vais soutenir ma thèse lundi 16 septembre dans l'amphi du CTL à 14h. La soutenace sera en anglais.
I'll defend my PhD on Monday, september 16 in the CTL amphitheater. The defence will be in english.