LibGeoDecomp (Library for Geometric Decomposition) is a generic library for stencil codes, specifically targeted at grid systems. The library eases development of such tightly coupled codes by taking over the complex task of parallelization and parameter tuning, thus enabling scientists to focus on their simulation code, rather than the technical details of the compute system.
Stencil codes are time and space discrete simulations on a regular grid, where each cell has a fixed neighborhood. Examples include cellular automata (e.g. Conway's Game of Life) and the Jacobi Method. Since cells require the state of their neighbors for each update, these problems are considered to be tightly coupled, causing significant challenges for the implementer. Grid systems aggravate this situation by their heterogeneity in respect to both, networks and CPUs.
LibGeoDecomp's main features are:

• Parallelization via MPI (and OpenMP coming up)

• Accelerator offloading (currently Nvidia CUDA, Cell in the works)

• Dynamic load balancing

• Customizable domain decomposition techniques

• Latency hiding via wide ghostzones and overlapping communication and computation (multi halo support in the works)

• Parallel I/O

• Customizable output formats
  

Written mostly as C++ class templates, the library itself inflicts little computational overhead.
For further reference see the LibGeoDecomp homepage: http://www.libgeodecomp.org