Preconditioned Krylov solvers on GPUs

被引:27
作者
Anzt, Hartwig [1 ]
Gates, Mark [1 ]
Dongarra, Jack [1 ,2 ,3 ]
Kreutzer, Moritz [4 ]
Wellein, Gerhard [4 ]
Koehler, Martin [5 ]
机构
[1] Univ Tennessee, Dept Elect Engn & Comp Sci, Innovat Comp Lab, Knoxville, TN 37996 USA
[2] Univ Manchester, Manchester, Lancs, England
[3] Oak Ridge Natl Lab, Oak Ridge, TN USA
[4] Friedrich Alexander Univ Nuermberg Erlangen, Erlangen, Germany
[5] Max Planck Inst Dynam Complex Tech Syst Magdeburg, Magdeburg, Germany
关键词
Krylov solvers; Preconditioning; GPU; ILU; Jacobi; SUBSPACE METHODS; BI-CGSTAB; ALGORITHM; VARIANT; ILU;
D O I
10.1016/j.parco.2017.05.006
中图分类号
TP301 [理论、方法];
学科分类号
081202 ;
摘要
In this paper, we study the effect of enhancing GPU-accelerated Krylov solvers with pre-conditioners. We consider the BiCGSTAB, CGS, QMR, and IDR(s) Krylov solvers. For a large set of test matrices, we assess the impact of Jacobi and incomplete factorization preconditioning on the solvers' numerical stability and time-to-solution performance. We also analyze how the use of a preconditioner impacts the choice of the fastest solver. (C) 2017 Elsevier B.V. All rights reserved.
引用
收藏
页码:32 / 44
页数:13
相关论文
共 42 条
[1]  
[Anonymous], 2016, LIGHTHOUSE FRAMEWORK
[2]  
Anzt H., 2014, 28 IEEE INT PAR DIST
[3]  
Anzt H., 2015, P 2 INT WORKSH HARDW
[4]  
Anzt H., 2016, VECP 2016 12 I UNPUB
[5]   Optimization and performance evaluation of the IDR iterative Krylov solver on GPUs [J].
Anzt, Hartwig ;
Kreutzer, Moritz ;
Ponce, Eduardo ;
Peterson, Gregory D. ;
Wellein, Gerhard ;
Dongarra, Jack .
INTERNATIONAL JOURNAL OF HIGH PERFORMANCE COMPUTING APPLICATIONS, 2018, 32 (02) :220-230
[6]   Efficiency of general Krylov methods on GPUs - An experimental study [J].
Anzt, Hartwig ;
Dongarra, Jack ;
Kreutzer, Moritz ;
Wellein, Gerhard ;
Koehler, Martin .
2016 IEEE 30TH INTERNATIONAL PARALLEL AND DISTRIBUTED PROCESSING SYMPOSIUM WORKSHOPS (IPDPSW), 2016, :683-691
[7]  
Balay S., 2016, PETSC
[8]  
Benzi M., 1999, ELECTRON T NUMER ANA, V8, P88
[9]  
Brezinski C., 1991, Numerical Algorithms, V1, P261, DOI 10.1007/BF02142326
[10]   Breakdowns in the implementation of the Lanczos method for solving linear systems [J].
Brezinski, C ;
RedivoZaglia, M ;
Sadok, H .
COMPUTERS & MATHEMATICS WITH APPLICATIONS, 1997, 33 (1-2) :31-44