Book Description

The recent technological advancements and market trends are causing an interesting phenomenon towards the convergence of the high-performance and the embedded computing domains. Critical real-time embedded systems are increasingly concerned with providing higher performance to implement advanced functionalities in a predictable way. OpenMP, the de-facto parallel programming model for shared memory architectures in the high-performance computing domain, is gaining the attention to be used in embedded platforms. The reason is that OpenMP is a mature language that allows to efficiently exploit the huge computational capabilities of parallel embedded architectures. Moreover, OpenMP allows to express parallelism on top of the current technologies used in embedded designs (e.g., C/C++ applications). At a lower level, OpenMP provides a powerful task-centric model that allows to define very sophisticated types of regular and irregular parallelism. While OpenMP provides relevant features for embedded systems, both the programming interface and the execution model are completely agnostic to the timing requirements of real-time systems. This thesis evaluates the use of OpenMP to develop future critical real-time embedded systems. The first contribution analyzes the OpenMP specification from a timing perspective. It proposes new features to be incorporated in the OpenMP standard and a set of guidelines to implement critical real-time systems with OpenMP. The second contribution develops new methods to analyze and predict the timing behavior of parallel applications, so that the notion of parallelism can be safely incorporated into critical real-time systems. Finally, the proposed techniques are evaluated with both synthetic applications and real use cases parallelized with OpenMP. With the above contributions, this thesis pushes the limits of the use of task-based parallel programming models in general, and OpenMP in particular, in critical real-time embedded domains.