Open AccessPrediction of parallel NIKE3D performance on the KSR1 system
Author(s) -
Peter Su,
T. Zacharia,
Robert E. Fulton
Publication year - 1995
Language(s) - English
Resource type - Reports
DOI - 10.2172/86297
Subject(s) - cholesky decomposition , parallel computing , multiprocessing , computer science , computation , supercomputer , oak ridge national laboratory , parallel algorithm , distributed memory , finite element method , computational science , shared memory , algorithm , engineering , structural engineering , eigenvalues and eigenvectors , physics , quantum mechanics , nuclear physics
Finite element method is one of the bases for numerical solutions to engineering problems. Complex engineering problems using finite element analysis typically imply excessively large computational time. Parallel supercomputers have the potential for significantly increasing calculation speeds in order to meet these computational requirements. This paper predicts parallel NIKE3D performance on the Kendall Square Research (KSR1) system. The first part of the prediction is based on the implementation of parallel Cholesky (U{sup T}DU) matrix decomposition algorithm through actual computations on the KSRI multiprocessor system, with 64 processors, at Oak Ridge National Laboratory. The other predictions are based on actual computations for parallel element matrix generation, parallel global stiffness matrix assembly, and parallel forward/backward substitution on the BBN TC2000 multiprocessor system at Lawrence Livermore National Laboratory. The preliminary results indicate that parallel NIKE3D performance can be attractive under local/shared-memory multiprocessor system environments