Workflow models for heterogeneous distributed systems

The role of data in modern scientific workflows becomes more and more crucial. The unprecedented amount of data available in the digital era, combined with the recent advancements in Machine Learning and High-Performance Computing (HPC), let computers surpass human performances in a wide range of fields, such as Computer Vision, Natural Language Processing and Bioinformatics. However, a solid data management strategy becomes crucial for key aspects like performance optimisation, privacy preservation and security. Most modern programming paradigms for Big Data analysis adhere to the principle of data locality: moving computation closer to the data to remove transfer-related overheads and risks. Still, there are scenarios in which it is worth, or even unavoidable, to transfer data between different steps of a complex workflow. The contribution of this dissertation is twofold. First, it defines a novel methodology for distributed modular applications, allowing topology-aware scheduling and data management while separating business logic, data dependencies, parallel patterns and execution environments. In addition, it introduces computational notebooks as a high-level and user-friendly interface to this new kind of workflow, aiming to flatten the learning curve and improve the adoption of such methodology. Each of these contributions is accompanied by a full-fledged, Open Source implementation, which has been used for evaluation purposes and allows the interested reader to experience the related methodology first-hand. The validity of the proposed approaches has been demonstrated on a total of five real scientific applications in the domains of Deep Learning, Bioinformatics and Molecular Dynamics Simulation, executing them on large-scale mixed cloud-High-Performance Computing (HPC) infrastructures.