MRG Interactive Developments

The Global Environmental Multiscale (GEM) model has been (and continues to be) under development at the Meterological Serivce of Canada's Numerical Research Laboratory (RPN). For more information about the model, consult either the local documentation or the RPN Community Modelling site.

Parallelization in the GEM model is primarily accomplished using a standard message passing interface (MPI). The parallelization strategy is based on decomposition of the full (called "global", even when the GEM model is running in limited-area mode) domain into a set of tiles or subdomains. Computational tasks on each of these subdomains is farmed out to a separate processor, and communication is minimized by passing only inteface (called "halo") regions between topologically-adjacent processors. This approach provides effective scaling on both machines that leverage either shared or distributed main memory, or any combination of the two.

The GEM model solver uses matrix transformations as part of its Fourier transform-based methodology. These transformations impose a set of strict criteria on the extent to which the domain can be decomposed into tiles in each dimension. The most restrictive domain length is generally in the vertical direction, where the dimension length is generally on the order of 60, depending on the resolution of the model grid. Since this "nk" dimension is transposed with the horizontal local grid dimensions, the comparison of "nk" and "npx" (number of processors in the X-dimension) and "npy" (number of processors in the Y-dimension) imposes non-trivial limitations. These limitations are explicitly computed in the RPN library subprogram comm/mpi/RPN_COMM_limit.f at runtime; however, for large grids, queue and preprocessing times can be prohibitive for a trial-and-error approach. Computation of the valid domain splitting is most easily accomplished using the calculator provided here.

Questions or Comments? Contact our WebMaster.