Abstract: |
|
This article introduces EDGE, a solver package for fused seismic simulations. Fused seismic simulations are a novel technique addressing one of the grand challenges of computational seismology: large ensemble runs of geometrically similar forward simulations. Application fields include, but are not limited to: uncertainty quantification in the context of seismic hazard analysis or the accurate derivation of velocity models through tomographic inversion. For efficient and accurate handling of complex model geometries (topography, fault geometries, material heterogeneities), EDGE utilizes the Discontinuous Galerkin (DG) method for spatial and Arbitrary high order DERivatives (ADER) for time discretization, implemented for unstructured tetrahedral meshes. This scheme requires sparse and dense matrix multiplications at the kernel level. By choosing a sufficient memory layout and relying on runtime code generation and specialization, both sparse and dense operations can be vectorized efficiently on wide-SIMD machines such as the Intel Xeon Phi processor. We present a convergence study of single and fused seismic simulations, code validation in an established benchmark, as well as a detailed performance assessment for different discretization orders. As target architecture we select the recently released Intel Xeon Phi processor which powers the Theta and Cori-II supercomputers. For a single sixth order seismic forward simulation we achieved 10.4 PFLOPS of hardware performance and 5.0 PFLOPS for fused simulations in fourth order both occupying 9,000 nodes of Cori-II. From a throughput perspective, fused seismic simulations can outperform single forward simulation by 1.8X to 4.6X, depending on the chosen order of the method. |
|