On the Performance of Parallel Algebraic Multigrid

Naegel, Arne ; Falgout, Robert

Lawrence Livermore National Laboratory

关键词: Parallel Processing;    B Codes;    Algorithms;    Performance;    99 General And Miscellaneous//Mathematics, Computing, And Information Science;   

DOI  :  10.2172/15003254
RP-ID  :  UCRL-ID-151859
RP-ID  :  W-7405-ENG-48
RP-ID  :  15003254

美国|英语

来源: UNT Digital Library

【 摘 要 】

As algebraic multigrid (AMG) can be applied to a wide variety of problems on extremely large, unstructured grids on the one hand and massively parallel computer systems are available on the other hand, there has been a need for parallel implementations of AMG. This report analyses the performance of BoomerAMG[1], a parallel algorithm developed at Lawrence Livermore National Laboratory. Abstracting the implementation we describe the basic components of the costly setup phase and analyze their behavior in a (massively) parallel distributed memory environment. We present numerical results, compare them to the developed theory and finally aim at possible improvements in the future.