| 1 |
We decompose the three dimensional grid over the Nproc nodes of a parallel machines
|
| 2 |
There are the usual choices for this depending on the shape of regions stored in each processor
|
| 3 |
Example 1: Distribute over j(h) with all values of i and k (x and z) for each j stored in a node
-
Then one has communication at sweep 2 but none at sweep 1 or 3.
|
| 4 |
Example 2: Distribute as cubic blocks with all three indices i j and k distributed.
|
| 5 |
This type of distribution is normally optimal for nearest neighbor problems as it minimizes surface over volume effects.
-
This is not quite clear for ADI, as each sweep only in one dimension
|
