Comparison of Several Scalable Programming Models

摘要

The NRL Layered Ocean Model (NLOM) is written in the tiled data parallel programming style, and uses an application specific programming interface to isolate operations that require communication. This allows different scalable programming models to be 'plugged' into NLOM with relatively little effort. NLOM is similar to other OGCM's, except that it uses a direct Helmholtz's equation solver as part of its semi-implicit time scheme and typically runs with a very large horizontal extent and very few layers in the vertical. There are now several Fortran-based SPMD programming models to chose from on machines with a hardware global memory: (a) MPI-1 message passing, (b) MPI-2 put/get, (c) BSP, (d) SHMEM, (e) F--, (f) OpenMP, and (g) HPF. These models are compared and contrasted based on actual experience with NLOM and related kernel benchmarks.