NVIDIA has invited SNL to present results of a GPU performant shock hydrodynamics code at their Super Computing (SC17) booth.

NVIDIA has invited SNL to present results of a GPU performant coupled hydrodynamics, low Magnetic Reynolds number  (low Rm) code at their Super Computing 17 (SC17) booth. Researchers at Sandia are developing a new shock hydrodynamics capability, based on adaptive Lagrangian techniques targeted at next generation architectures. The code simulates shock hydrodynamics on GPU  architectures using the Kokkos library to provide portability across architectures.  Mesh and field data management, as well as adaptive Lagrangian operations are being developed to run exclusively on the GPU.  New algorithms using tetrahedral elements and a predictor-corrector time integrator have been implemented. Low Rm physics is solved using NVIDIA’s AmgX GPU-aware, algebraic multigrid solver. Using an exemplar problem provided by our NW partners we have demonstrated good scaling and performance on next generation architectures.  Notably, the exemplar problem demonstrates the advantages of a device-centric design philosophy, where the hydrodynamics physics solve, including adaptivity and remapping, are hosted on the coprocessor with exceptional performance on the GPU relative to traditional multi-core architectures. Additionally, solve times for the low Rm physics with the AmgX software demonstrate sub-second solve times for million degree of freedom problems. Next steps include full-scale testing on Trinity (on both the Haswell and KNL partitions) as well as Sierra as it becomes available, the addition of robust treatment for material/material interactions and the inclusion of more comprehensive MHD physics.

Contact
Glen Hansen, gahanse@sandia.gov

February 1, 2018