Speaker
Description
ExaGRyPE is a suite of solvers for numerical relativity based on ExaHyPE 2, our second-generation Exascale Hyperbolic PDE Engine. This solver tackles the Einstein equations/relativistic hydrodynamics equations under a 3+1 foliation, with a focus on compact object spacetimes. The implementation utilizes a block-structured Cartesian grid with higher-order Finite Difference schemes and adaptive mesh refinement while enabling massive parallelism through message passing, domain decomposition, and task parallelism.
Our approach formalizes simulation creation as a sequence of lowering operations, where abstract logical tasks are broken down into progressively finer tasks until reaching a C++ executable level. The program logic is specified through a domain-specific Python interface, which maps to numerical tasks, then to technical tasks for parallelization, and finally to task graphs containing PDE evaluations, initial conditions, and boundary conditions. This architecture creates a rigorous separation of concerns, shielding users from technical details and simplifying the development of novel physical models. This approach enables researchers to efficiently develop and deploy novel physical models while leveraging the full computational potential of exascale systems.
In this talk, we are going to present the ExaGRyPE architecture and demonstrate its capabilities through comprehensive benchmarks. We will also highlight our unique solver-coupling feature, which allows us to seamlessly integrate finite volume scheme for relativistic hydrodynamics within our high-order finite-difference framework for spacetime evolution.
