Section: New Results

Experimenting task-based runtimes on a legacy Computational Fluid Dynamics code with unstructured meshes

Advances in high performance computing hardware systems lead to higher levels of parallelism and optimizations in scientific applications and more specifically in computational fluid dynamics codes. To reduce the level of complexity that such architectures bring while attaining an acceptable amount of the parallelism offered by modern clusters, the task-based approach has gained a lot of popularity recently as it is expected to deliver portability and performance with a relatively simple programming model. In this work, we have reviewed and presented the process of adapting part of Code Saturne, a legacy code at EDF R&D into a task-based form using the PARSEC (Parallel Runtime Scheduling and Execution Control) framework. We have first shown show the adaptation of our prime algorithm to a simpler form to remove part of the complexity of our code and then present its task-based implementation. We then have compared performance of various forms of our code and discuss the perks of task-based runtimes in terms of scalability, ease of incremental deployment in a legacy CFD code, and maintainability [8].