High-fidelity nonlinear low-order unstructured implicit finite-element seismic simulation of important structures by accelerated element-by-element method

摘要

We enable large-scale high-fidelity finite-element seismic response simulations of important structures, that are expected to contribute towards improvement in seismic design verification, by reducing cost of the nonlinear dynamic unstructured low-order implicit finite-element method. Most of the computational cost of this method is involved in the element-by-element (EBE) method, which is a typical example of a "low computation/(data load or store)" kernel that appears in many applications that is not straightforward to attain performance on current computer systems. Therefore, special care based on computer science is required to make use of the potential of computer architecture and achieve fast analysis. In this study, we developed a kernel algorithm and implementation suitable for the target Arm v8.2-A scalable vector extension (SVE) CPU-based supercomputer Fugaku. 5.11and 8.69-fold speedup was attained by using the developed EBE kernel in a standard preconditioned conjugate gradient solver and a state-of-the-art SC14 massively parallel solver algorithm, respectively. Furthermore, by using the developed EBE kernel in the state-of-the-art solver, a 49 billion degrees-of-freedom high-fidelity seismic response analysis can be conducted in practical speed corresponding to 60,000 time-steps in half a day using Fugaku. The obtained insights are expected to be useful for accelerating other scientific computing methods with "low computation/(data load or store)" kernels.