Understanding the efficiency of parallel incomplete Cholesky preconditioners on the performance of ICCG solvers for multi-core and GPU systems
Dept. of Electr. & Comput. Eng., McGill Univ., Montreal, QC, Canada
14th Biennial IEEE Conference on Electromagnetic Field Computation (CEFC), 2010
This paper aims at understanding the effect of parallelizing incomplete Cholesky (IC) factorization on the overall performance of the incomplete Cholesky conjugate gradient (ICCG) solver method, optimized on multi-core and GPU based Systems. Parallel IC preconditioners, which are based on graph reordering and arbitrary levels of allowed fill-in, are tested on structured and unstructured matrices to study the tradeoffs between the level of fill-in, degree of IC parallelism and convergence rate of the ICCG solver.
May 14, 2011 by hgpu
