PDExa

Optimized Software Methods and Technologies for Partial Differential Equations

The PDExa project develops highly scalable numerical solvers for partial differential equations, particularly in the field of fluid mechanics. From the diverse field of partial differential equations, the focus is on incompressible flow phenomena with coupled transport processes. The project PDExa develops novel software technologies for the simulation of partial differential equations on exascale supercomputers. Our contributions are added to the widely used open-source projects, splitting the efforts into the development of mathematical concepts for finite-element methods, solvers for the application in the context of computational fluid dynamics (CFD), and linear algebra tools. 

The focus of the sub-project at the Chair of Computer Architecture and Parallel Systems (CAPS-TUM) is to analyze the performance portability of the essential algorithms to all common GPUs (AMD, NVIDIA, Intel) and CPUs (Intel, AMD, ARM). Cross-platform abstractions for GPUs or vectorization are used for this purpose. In contrast to the US ECP project with general portability layers, such as Kokkos, Raja or Alpaka, problem-specific approaches with a software interface at the level of mathematical calculation of finite elements are to be investigated and implemented with simultaneously higher performance. Furthermore, at CAPS-TUM, we aim to provide end-to-end performance modeling and analysis to guide application developers in identifying key resource bottlenecks and estimating performance. To this end, we focus on both node-level and inter-node analyses of application performance. The key focus is to optimize applications for achieving high performance using distributed GPU setups in large-scale scenarios.