Hello all, The way PETSc MatCreateMPIAIJ distributes an N*N square matrix is row wise. And internally every processor stores its local matrix into two sub matrices one Diagonal and other part is off-Diagonal. (more here --> http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/Mat/MatCreateMPIAIJ.html ).
So when PETSc MatMul (v = A.x) is called it basically tries to hide communication of vector (x) by overlapping computation. As for as I understand every process first initiates a kind of non-Blocking Bcast of vector (x) and continues to do MVM of diagonal submatrix and then waits till communication is done and finally does MVM for the off-Diagonal matrix. My question is this (since I am new) what was the historical reason PETSc opted for this Diagonal and off-Diagonal storage? And what if I want to change the overlapping strategy of MVM by lets say introducing a ring based communication of vector (x), then I have to partition the local matrix into not two sub-matrices but P sub-matrices (here P = number of processors). Does PETSc provide this facility or one has to go from scratch to implement different techniques to store local matrix? Thanks Bibrak -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120220/5fc853c8/attachment.htm>
