Scalable Parallel Algebraic Multigrid Solvers

Bank, R ; Lu, S ; Tong, C ; Vassilevski, P

Lawrence Livermore National Laboratory

关键词: Communications;    Velocity;    Convergence;    Algorithms;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;   

DOI  :  10.2172/15015127
RP-ID  :  UCRL-TR-210788
RP-ID  :  W-7405-ENG-48
RP-ID  :  15015127

美国|英语

来源: UNT Digital Library

【 摘 要 】

The authors propose a parallel algebraic multilevel algorithm (AMG), which has the novel feature that the subproblem residing in each processor is defined over the entire partition domain, although the vast majority of unknowns for each subproblem are associated with the partition owned by the corresponding processor. This feature ensures that a global coarse description of the problem is contained within each of the subproblems. The advantages of this approach are that interprocessor communication is minimized in the solution process while an optimal order of convergence rate is preserved; and the speed of local subproblem solvers can be maximized using the best existing sequential algebraic solvers.