Development of a finite volume library on GPUs using the adaptive mesh refinement library t8code

Kurzfassung

The objective of this end-of-studies intership is to develop a prototype numerical simulation framework usingthe adaptive mesh refinement library t8code developed by the work group scalable adaptive mesh refinement where I did this internship at the DLR (Deutsches Zentrum für Luft- und Raumfahrt). This framework should expose an API that allows any user to easily implement a finite volume solver without compromising ease of use for performance. After having done that, the goal is to assess t8code’s suitability as an AMR library used to do numerical simulation on GPUs while doing mesh adaptation on the CPU. This report is structured in multiple sections: firstly, we will detail the difference between GPU and CPU architectures to show what are the challenges when programming on GPUs and present various GPU programming models and optimization considerations. Then, we will present the different ways of doing adaptive mesh refinement and focus on t8code’s approach to AMR. Afterwards, we will present the framework that I developed as well as the design choices behind it. Next, we will show performance results and end with conclusions and perspectives by analysing the overhead of AMR and also the computational efficiency of the GPU solver.