| 1 |
Warning: Your compiler may do things differently!
|
| 2 |
Computation is static, homogeneous, and over full array (with respect to the edges)
|
| 3 |
Accesses to node array are ³nearest neighbor² in the mesh
-
This is not reflected in the index order!
-
\ This does not favor either BLOCK or CYCLIC
|
| 4 |
To minimize communication, edge and node distributions must fit the mesh topology
-
HPF¹s regular distributions are not ideal for this
-
HPF 2.0 indirect distributions may be better, but require careful construction
|
| 5 |
The bottom line, part II
-
No silver bullet
-
Order the nodes and edges to bring ³close² entities together, then use BLOCK
|
