| Abstract: |
While seismic phenomena such as crustal deformation, wave propagation, and nonlinear wave amplification vary in time and spatial scales, they can be modeled as static/dynamic, nonlinear/linear responses of domains with complex geometries. The implicit finite-element method with low-ordered unstructured elements is suitable for solving such problems; however, its cost becomes huge as the target domain is large, and many simulations must be conducted to consider uncertainty and conduct optimization. To address this problem, we have been developing high-performance finite-element analysis methods for large-scale computing systems.
In this talk, I will introduce our development on an algorithm for converting unstructured computation based on random data access to data-driven computation based on continuous data access computation for next-generation systems, where relative data access costs are expected to increase when compared with current computer systems. Here, data from past time steps are used to predict the solution of the next time step, which reduces the number of solver iterations and reduces runtime without loss in accuracy. The performance of this method applied to viscoelastic crustal deformation analysis in CPU-based Fugaku will be presented. Furthermore, we developed an algorithm such that CPUs with large memory can be utilized for the data-driven part and high-performance GPUs can be utilized for the iterative solver part at the same time to enable further improvement in time-to-solution. I will show the performance results on an NVIDIA GH200-based system with a fast CPU-GPU interconnect. Such development can be considered as a method that reduces memory access volumes or random memory accesses and converts them to high-density computations with structured data accesses, which is also expected to be useful in future systems.
|
