Let us discuss parallelization of first (D + Lo) DU1 Lower Triangular equation as Upper triangular case is trivially the same by reversing order of indices
|
RNR-93-007 distinguishs Wavefront (2D) and hyperplane(3D) methods and concludes hyperplane best.
-
We agree but will draw all pictures in two dimensions -- they have obvious three dimensional extensions
|
Key observation is that:
-
After Processor 1 has finished case i=j=k=1
-
Processor 1 can proceed with case i=2 j=k=1 while
-
Processor 2 can compute i=1 k=2 and j=1
-
This (only taking k=1 for Processor 1 and k=2 for Processor 2) gives a pipeline with Processor 2 following one grid point behind Processor 1.
|