Amik StCyr - Shell

Abstract

‘High-Order Methods for Seismic Imaging on Many-core Architectures’

Reverse Time Migration (RTM) is a widely used wave-equation based seismic imaging algorithm. In this work, three different discretization strategies for RTM are presented. The first method considered is a very high-order finite-difference method. The latter employs stacked blocks of varying lateral resolution and variable grid spacing in the depth direction. The main novelty of this algorithm resides on the ability to avoid disk IO entirely. This is achieved by saving boundaries during the forward propagation and on the usage of compression to lower the overall memory requirements. The second discretization proposed is a high-order discontinuous Galerkin method on tetrahedral meshes. To help with the stringent CFL condition, a multirate time-integrator is used. The approach also avoids disk IO by saving face flux contributions at every time-step during the forward propagation pass. We also propose an improved imaging condition. Finally, a recent joiner in the high-order crowd is Radial-Basis-Function based finite-differences. The latter outperforms all of our wave equations solvers on a simple benchmark problem.

All RTM flavors were developed with the ability to handle multiple hardware platforms (Clusters of Intel Xeon or Nvidia GPUs), by making use of various parallel programming APIs (MPI, CUDA, OpenCL, OpenMP) under a MPI+X strategy. The finite-difference code employs static polymorphism techniques while the discontinuous Galerkin and RBF ones are using a Just In Time (JIT) compilation approach.

Short biography:

Amik St-Cyr worked in Shell for the last 4 years in the Computation & Modelling group based in Houston. He recently convinced the Seismic Applications Team in Rijswijk to take him under its wing to participate in the Exascale SIPMAP project. Before joining Shell, he co-developed amongst other things, for the US government, a next generation climate model currently scaling to O(100000) cores. Earlier to this, he was a postdoc in high-performance computational fluid dynamics at McGill University (in Canada) where he developed a library to MPI-ize CFD codes. He studied applied mathematics and his PhD was on central-schemes for 3D shock capturing methods on unstructured finite-volume meshes (now one of the default numerical methods in OpenFOAM). He studied mathematical-physics before discovering his passion for scientific computing.