GPU Accelerated Computing Towards a Fast and Scalable Seismic Wave Modelling in SEISCOPE SEM46 Code

Jian Cao (1) , Romain Brossier (1) , Eduardo Cabrera (2) , Josep de la Puente (2) , Ludovic Métivier (1, 3) , Alizia Tarayoun (1)

1 ISTerre - Institut des Sciences de la Terre
2 BSC-CNS - Barcelona Supercomputing Center - Centro Nacional de Supercomputacion
3 EDP - Equations aux Dérivées Partielles

Main objectives This study aims at the development of GPU-accelerated code for a fast and scalable seismic wave modelling using the spectral element method (SEM) within the framework of our full waveform modelling and inversion code SEM46. Overall, it contains the single GPU algorithm investigation to explore the computational efficiency that stems from the application of Cartesian-based structured meshes and the multi-GPU implementation based on the domain-decomposition strategy. New aspects covered (1) Three types of parallel prototypes of SEM-based CUDA kernel are presented and compared in terms of modelling accuracy and computational efficiency. (2) To benefit from the Cartesian-based structured mesh, an element-wise parallelization with odd-even mesh coloring is proposed which achieves a significant speedup over the equivalent serial CPU reference code. (3) With the help of CUDA-AWARE MPI, an excellent scaling is obtained in the domain-decomposition-based multi-GPU implementation, which boosts its applicability on large-scale realistic problems.

Domaines: Géophysique [physics.geo-ph]; Modélisation et simulation

Jian Cao, Romain Brossier, Eduardo Cabrera, Josep de la Puente, Ludovic Métivier, et al.. GPU Accelerated Computing Towards a Fast and Scalable Seismic Wave Modelling in SEISCOPE SEM46 Code. Sixth EAGE High Performance Computing Workshop, Sep 2022, Milan, Italy. pp.1-5, ⟨10.3997/2214-4609.2022615010⟩. ⟨hal-03852696⟩