Efficient SIMD and Shared-Memory Parallelization of 3D Acoustic Wave Propagation Simulation

• Chahinèze Ztoti Université Paris-Saclay
• Claude Tadonki Mines Paris - PSL University
• Roblex Nana Tchakoute Mines Paris - PSL University
• Hervé Chauris Mines Paris - PSL University

Resumo

We are interested in the implementation of a mesh-based modeling for a numerical simulation of acoustic wave propagation. Considering a 2D validated version, we derive a 3D design by extending the spatial domain with a third dimension. The dimension n here refers to the spatial domain, which is modeled as a n-dimensional mesh grid augmented by a time axis. With this 4D model, considering large-scale scenarios goes with a significant demand of memory space, which can therefore justify delving into the domain decomposition strategy in the perspective of running on a distributed memory parallel machine or a cluster of standard processors. We investigate the required volume of memory and make it minimal considering the specific rules of the simulation. Next, beside ordinary code refactoring and optimization, we explore all levers of computing power at the level of a single processor, more precisely multi-thread parallelism and SIMD. Experimental results with various scenarios illustrate the possibility to handle bigger instances and an impressive global speedup (×20 on a specific execution scenario) compared to the original version of the code. A natural step further for this work is the design and implementation of distributed memory parallel version through domain decomposition, and GPU acceleration.

Palavras-chave: Three-dimensional displays, Codes, Single instruction multiple data, Memory management, Parallel machines, Numerical simulation, Numerical models, Standards, Optimization, Acoustic waves, propagation, simulation, multi-core, OpenMP, AVX