Scaling Implicit Solvers to Millions of Cores
Math Library Design
TimeWednesday, June 19th1:45pm - 2:15pm CEST
DescriptionThe rapid development of supercomputers and the growing needs of extreme-scale applications pose grand challenges in the design of highly efficient and scalable parallel algorithms. It is of crucial importance to adapt the parallel algorithm with the architectural features so that the potential of the supercomputers can be fully unleashed at the extreme scale. In this talk, I will summarize our recent efforts on designing implicit solvers that aim to take full advantage of state-of-the-art many-core based heterogeneous supercomputers. A series of examples on both Tianhe-2 and Sunway TaihuLight are presented to show how high performance can be achieved in emergning benchmarks such as HPCG and HPGMG and challenging applications such as nonhydrostatic atmospheric dynamics and gaseous wave detonations.