| 1 |
Communication is needed when one sweeps in a given index, say i, and that index location changes
|
| 2 |
Note Algorithms for sequential tridiagonal solvers are given in the appendix of volume 1 of Hirsch
|
| 3 |
For a scalar tridiagonal matrix with N rows, the solution takes time 5N
|
| 4 |
For a block tridiagonal matrix, one increases this by a factor between 25 and 125 for 5 by 5 blocks.
|
| 5 |
Note communication proportional to "5" i.e. size of vector being calculated as one changes block size
|
| 6 |
However calculation proportional to 52 to 53 as one changes block size i.e. it increases faster with block size than communication
-
This illustrates that extra CFD components (1 goes to 5) increases efficiency of parallel CFD codes compared to comparable scalar equations
|
