Domain Dependencies in an Adaptive Mesh Refinement Non-Hydrostatic Dycore

We present a non-hydrostatic dynamical core ("dycore") that has refinement as well as higher-order accuracy in both time and space. The complex dependencies of both the multi-level time integration scheme and the spatial discretizations have been expressed as a data dependency graph, so as to expose more parallelism in the domain decomposition. We present tradeoffs between the scaling, order of accuracy, and memory constraints, based on a simplified model of the application data dependencies. The approach is demonstrated for both the Xeon Ivy Bridge and Phi, providing insight into the scalability of next-generation dycores on next-generation architectures. Implications for global atmospheric simulations at < 10km will be discussed, as well as numerical and computational research opportunities for both improved performance and throughput.