This paper describes the application of high performance asynchronous stencil computations for 3D acoustic modeling on a synthetic land survey. Using the Finite-Difference Time-Domain (FDTD) method, a parallel Multicore Wavefront Diamond-tiling (MWD) stencil kernel (Malas et al. 2015, Malas et al. 2017) drives the high performance execution using temporal blocking to maximize data locality, while reducing the expensive horizontal data movement. As absorbing boundary conditions, we use Convolutional Perfectly Matched Layer (CPML), which have to be redesigned to not interrupt the asynchronous execution flow engendered by the MWD stencil kernel for the inner-domain points. The main idea consists in weakening the data dependencies by moving the CPML computations into the inner-computational loop of the MWD stencil kernel (Akbudak et al. 2019). In addition to handling the absorbing boundary conditions, applying the asynchronous MWD with CPML kernels to a realistic land survey requires the extraction of the wavefield value at each receiver position. We revisit the default extraction process and make it also compliant with the overall asynchrony of the 3D acoustic modeling. We report performance improvement up to 24% against the standard spatial blocking algorithm on Intel multicore chips using the synthetic land survey, which is representative of an area of interest in Saudi Arabia. While these results concur with previous performance campaign assessment, we can actually produce and assess the resulting 3D shot gather accuracy. To our knowledge, this is the first time the effectiveness of asynchronous MWD stencil kernel with CPML absorbing boundary conditions is demonstrated in an industrial seismic application.
|Original language||English (US)|
|Title of host publication||SPE Middle East Oil and Gas Show and Conference|
|Publisher||Society of Petroleum Engineers (SPE)|
|State||Published - Mar 15 2019|
Bibliographical noteKAUST Repository Item: Exported on 2020-10-01
Acknowledgements: Computations were done on KAUST's Shaheen II supercomputer. We acknowledge the support of the KSL supercomputing lab.