|
Do normal block decomposition
|
Parallel Phase I: Update all Red Points
-
Communicate black points at k-1 to halo in each processor
-
Compute red points to iterate k
|
Parallel Phase II: Update all Black Points
-
Communicate red points at k to halo in each processor
-
Compute black points to iterate k
|
This has similar efficiency analysis to Jacobi except a little more sensitive to latency
-
Same amount of communication but twice as many messages
|
In electrical power system and similar simulations, one gets irregular sparse matrices and no way to get such a clean parallel algorithm
-
In fact not clear if there is a sequential Gauss Seidel that converges
|
