P22 - Efficient Data Managment in Fully Spectral Dynamo Simulations on Heterogeneous Nodes
DescriptionOur CFD framework QuICC, based on a fully spectral method, has been successfully used for various dynamo simulations in spherical and Cartesian geometries. It runs efficiently on a few thousands of cores using a 2D data distribution based on a distributed memory paradigm (MPI). In order to better harness the computing power of current and upcoming HPC systems, we present our work on refactoring the framework to introduce a hybrid distributed and shared memory parallelization (MPI + X). Our fully spectral method in a spherical geometry leads to 3D sparse tensors with a well defined block structure. Our strategy is based on the principle of separation of concerns which is applied on multiple levels. The operators API map to mathematical operations on tensors, without knowledge of the data layout or back-end. The tensors are represented by a type that we call "View" which encodes sparsity and memory layout. The refactorization of the new API and data layout results in a code base that has a lower memory footprint, it is more composable thus easier to maintain and extend to cover different back-ends. The API and a performance comparison for different operators and back-ends (CPU and GPU) will be presented.
TimeTuesday, June 2719:30 - 21:30 CEST