Towards Exascale Hyperbolic PDE Engines – Are We Skating to Where the Puck Will Be?

Exascale supercomputers are “ante portas”. They pose novel challenges to simulation software, such as how to cope with heterogeneous performance of compute resources, increased failure rates and respective demands on resiliency, or with scalability on multiple levels of parallelism, which are at the same time suffering from unpredictable performance. CPU/GPU power will continue to outperform memory and communication hardware, which asks for algorithms that feature high arithmetic intensity and minimize data movement across the entire memory hierarchy. Can we successfully develop simulation software that will run efficiently on the supercomputers of 2030? This talk will present on recent work to better prepare simulation software for exascale, focusing on two concrete packages: ExaHyPE and SeisSol. ExaHyPE is an engine to solve hyperbolic PDE systems. While it provides flexibility with respect to the tackled PDEs, it focuses on high-order discontinuous Galerkin (with a-posteriori Finite-Volume-based limiting) on tree-structure Cartesian meshes as the underlying numerical scheme. I will outline ExaHyPE’s code generation approach to tailor the engine to different needs of application, algorithm and code optimisation experts, and will highlight a fine-grain task-offloading strategy which can respond to performance fluctuations in hardware and combines with an MPI rank replication approach. Similar, I will present recent work on the earthquake simulation package SeisSol. SeisSol allows several modelling variants for seismic wave propagation and earthquake sourcing. It relies on code generation for its performance critical element-local small tensor operations, to cope with the complexity of realising several model variants on different hardware.     «

Exascale supercomputers are “ante portas”. They pose novel challenges to simulation software, such as how to cope with heterogeneous performance of compute resources, increased failure rates and respective demands on resiliency, or with scalability on multiple levels of parallelism, which are at the same time suffering from unpredictable performance. CPU/GPU power will continue to outperform memory and communication hardware, which asks for algorithms that feature high arithmetic intensity and m...     »