Parallelization strategy

Mesh-spectral computations are readily parallelized using the following approach:

• The overall structure of the computation is based on the SPMD (single program, multiple data) model; that is, it consists of some number N of processes all executing the same program, each on its own data.
• Each 3-dimensional data grid is distributed over the processes based on a 3-dimensional process grid of some or all of the N processes. It is usually the case that at any point in the computation all data grids will be distributed over the same process grid, but this is not required. Also, in the course of a computation, a data grid can be redistributed --- i.e., the process grid over which it is distributed can be changed. § 1.2.3 gives an example of such a computation.
• Non-grid variables --- global constants and reduction variables, e.g. --- are duplicated in each process.

• Distributing the data
• Neighbor operations
• Row or column operations
• Performing I/O