Commentary

A few points regarding distribution can be raised:

• if the ONTO clause is absent then objects are distributed onto the default grid (which is often supplied at run-time as a command line argument).
• BLOCK distribution is useful for neighbourhood calculations. Many algorithms are of this form.
• CYCLIC gives a good load balance. Each processor will have the same number of elements (give or take 1 element) so will perform the same number of calculations. As each processor receives a `random scattering' of elements then the intensity of calculation should also be random meaning that every processor should have to do more of less the same amount of work.

  Load balancing is important because at many points throughout the program there will be synchronisation points. Processors must wait until the last processor has finished its calculation before the next calculation can be started.

• collapsing a dimension is useful when a whole column or row is used in a single computation.
• It can generally be assumed that all scalar variables will be replicated on every processor. This means that the value of a particular scalar will be the same on every processor. The compiler will be responsible for keeping these values coherent except when control returns from an extrinsic. It is up to the user to ensure coherency in this case.

Return to corresponding overview page