Escolha do ladrilhamento para um simulador de ondas acústicas em gpus por meio de aprendizado de máquina

Abstract

The simulation of acoustic wave propagation is crucial in fields such as geophysics and seismic imaging, being modeled by numerical methods such as finite difference methods (FDM). These simulations are resource-intensive, especially in large-scale problems with 3D grids and multiple time steps. The use of GPUs has shown promise due to their parallel processing power, but one challenge is the memory access overhead. Tiling, which divides the grid into smaller blocks, improves data locality, optimizing memory access and increasing performance. However, selecting the optimal tile size for a given computation is not a trivial task. The optimal tile size depends on a variety of factors, including the specific architecture of the GPU, the size of the problem being solved, and the characteristics of the data being processed. In practice, the optimal tile size can vary significantly depending on the GPU’s memory hierarchy, the bandwidth between the processor and memory, and the computational intensity of the kernel. Moreover, the choice of tile size can also affect the parallelism and load balancing of the computation, making it a complex trade-off that requires careful tuning. In this study, we used machine learning to predict optimized tile sizes for acoustic wave simulations. We evaluated six algorithms (KNN, Decision Tree, Random Forest, XGBoost, LightGBM, and J48), and the results showed significant improvement, with the best model achieving improvement coefficients of 1.17 for the Turing GPU (RTX2080) and 1.11 for the Volta GPU (V100), as well as a success rate of over 75% for both GPUs.