Massively Parallel Scalable Atmosphere Model

Sandia has partnered with the National Center for Atmospheric Research (NCAR) to help develop and improve the Spectral Element Atmosphere Model (SEAM), which is the dynamical core of the High-Order Multiscale Modeling Environment (HOMME). A dynamical core is defined as that part of a weather or climate model that solves the primitive equations, which are the stratified Eulerian equations of fluid motion in spherical coordinates and constitute the most computationally intensive calculations in an atmosphere or ocean model. Currently, the world's fastest dynamical core runs on the Japanese Earth Simulator, which recently achieved over 26 teraflops on 5120 vector processors, at an equatorial resolution of roughly 10 km. Our objective is to compete with this, in both absolute and cost performance measures on the next generation of proposed DOE architectures.The Japanese dynamical core uses the spectral transform method (STM), a common production-level dynamical core that has known scalability issues on non-vector machines such as those typically employed at DOE. SEAM is the most highly scalable dynamical core among research codes today, and won an honorable mention in the 2001 Gordon Bell competition. We are working to add certain Sandia capabilities to SEAM, such as the Trilinos solver framework and the Zoltan load balancing library, to further improve its performance, especially at high resolution. We define high resolution as anything more resolved than current production capabilities of roughly 150km, with a target of 10km.

Element-based approaches permit new capabilities impossible for the STM, including both static mesh refinement around geographic features and dynamic refinement around moving fronts. This capability will make SEAM particularly viable for emergency response modeling of radiological, chemical or biological dispersions. This will be made possible by tools developed at Sandia such as AztecOO, Chaco and Zoltan. Research will consist of improving the time-stepping capability of SEAM; of h-, p-, and hp-refinement strategy studies applied to standard test cases; and application of dynamic load balancing algorithms. (Contacts: Mark Taylor and Bill Spotz)


May 6, 2008