Preconditioned Conjugate Gradient Method for Solution of Large Finite Element Problems on CPU and GPU

• Autorzy: Fialko S. Y. , Zeglen F.
• Języki publikacji: EN

Abstrakty

EN

In this article the preconditioned conjugate gradient (PCG) method, realized on GPU and intended to solution of large finite element problems of structural mechanics, is considered. The mathematical formulation of problem results in solution of linear equation sets with sparse symmetrical positive definite matrices. The authors use incomplete Cholesky factorization by value approach, based on technique of sparse matrices, for creation of efficient preconditioning, which ensures a stable convergence for weakly conditioned problems mentioned above. The research focuses on realization of PCG solver on GPU with using of CUBLAS and CUSPARSE libraries. Taking into account a restricted amount of GPU core memory, the efficiency and reliability of GPU PCG solver are checked and these factors are compared with data obtained with using of CPU version of this solver, working on large amount of RAM. The real-life large problems, taken from SCAD Soft collection, are considered for such a comparison.

Słowa kluczowe

EN

conjugate gradient; incomplete Cholesky factorization; iterative solvers; NVIDIA CUDA; preconditioned conjugate gradient

• Wydawca: Instytut Łączności - Państwowy Instytut Badawczy
• Czasopismo: Journal of Telecommunications and Information Technology
• Rocznik: 2016
• Tom: nr 2
• Strony: 26--33
• Opis fizyczny: Bibliogr. 20 poz., rys., tab.

Twórcy

autor

Fialko S. Y.

• Institute of Computer Science, Faculty of Physics, Mathematics and Computer Science, Tadeusz Kościuszko Cracow University of Technology, Warszawska st 24, 31-155 Cracow, Poland

autor

Zeglen F.

• Institute of Computer Science, Faculty of Physics, Mathematics and Computer Science, Tadeusz Kościuszko Cracow University of Technology, Warszawska st 24, 31-155 Cracow, Poland

Bibliografia

• [1] J. Zhang and L. Zhang, “Efficient CUDA polynomial preconditioned conjugate gradient solver for finite element computation of elasticity problems”, Mathem. Problems in Engin., article ID 398438, pp. 1–12, 2013.
• [2] M. Verschoor and A. C. Jalba, “Analysis and performance estimation of the Conjugate Gradient method on multiple GPUs”, Parallel Comput., vol. 38, no. 10–11, pp. 552–575, 2012.
• [3] S. Georgescu, P. Chow, and H. Okuda, “GPU Acceleration for FEMBased Structural Analysis”, Arch. Comput. Methods Engin., vol. 20, no. 2, pp. 111–121, 2013.
• [4] C. Yao, Z. Yonghua, Z. Wei, Z. Lian, “GPU-accelerated incomplete Cholesky factorization preconditioned conjugate gradient method”, J. of Comp. Res. & Develop., vol. 52, no. 4, pp. 843–850, 2015.
• [5] S. Y. Fialko, “Iterative methods for solving large-scale problems of structural mechanics using multi-core computers”, Archiv. of Civil and Mechan. Engin., vol. 14, no. 1, pp. 190–203, 2014.
• [6] M. Suarjana and K. H. Law, “A robust incomplete factorization based on value and space constraints”, Int. J. for Numerical Methods in Engin., vol. 38, pp. 1703–1719, 1995.
• [7] K. Malkowski, I. Lee, P. Raghavan, and M. J. Irwin, “Conjugate gradient sparse solver: Performance-power characteristics”, in Proc. 20th IEEE Int. Parallel & Distrib. Process. Symp. IPDPS 2006, Rhodes Island, Greece, 2006.
• [8] M. Papadrakakis, “Solving Large-Scale Problems in Mechanics”. Wiley, 1993.
• [9] S. Y. Fialko, “Parallel direct solver for solving systems of linear equations resulting from finite element method on multi-core desktops and workstations”, Comp. & Mathem. with Appl., vol. 70, pp. 2968–2987, 2015.
• [10] A. Jennings, “Development of an ICCG algorithm for large sparse systems”, in Preconditioned Methods. Theory and Applications, D. J. Evans, Ed. Gordon and Breach Publishers, 1983, pp. 425–438.
• [11] Nvidia Corporation, “cuBLAS API” [Online]. Available: http://docs.nvidia.com/cuda/cublas/
• [12] Nvidia Corporation, “cuSPARSE API” [Online]. Available: http://docs.nvidia.com/cuda/cusparse/
• [13] Nvidia Corporation, “csrsvsolve()” [Online]. Available: http://docs.nvidia.com/cuda/cusparse/#cusparse-lt-t-gt-csrsvsolve
• [14] Nvidia Corporation, “csrmv()” [Online]. Available: http://docs.nvidia.com/cuda/cusparse/#cusparse-lt-t-gt-csrmv
• [15] Nvidia Corporation, “cuSPARSE Indexing and Data Formats” [Online]. Available: http://docs.nvidia.com/cuda/cusparse/#cusparse-indexing-and-data-formats
• [16] Nvidia Corporation, “CUDA Library – conjugateGradientPrecond – Preconditioned Conjugate Gradient” [Online]. Available: http://docs.nvidia.com/cuda/cuda-samples/#preconditionedconjugate-gradient/
• [17] B.-Y. Su and K. Keutzer, “clSpMV: A Cross-Platform OpenCL SpMV Framework on GPUs” [Online]. Available: http://parlab.eecs.berkeley.edu/sites/all/parlab/files/clspMV-%20Keutzer.pdf
• [18] M. Naumov, “Parallel Solution of Sparse Triangular Linear Systems in the Preconditioned Iterative Methods on the GPU” [Online]. Available: https://research.nvidia.com/publication/parallelsolution-sparse-triangular-linear-systems-preconditioned-iterativemethods-gpu
• [19] S. Fialko and F. Zeglen, “ Block preconditioned conjugate gradient method for extraction of natural vibration frequencies in structural analysis”, in Proc. Feder. Conf. Comp. Sci. & Inform. Syst. FedCSIS 2015, Łódź, Poland, 2015, vol. 3, pp. 655–66.
• [20] V. S. Karpilovskii, E. Z. Kriksunov, A. A. Malyarenko, M. A. Mikitarenko, A. V. Perelmuter, and M. A. Perelmuter, “SCAD computational complex”, ASV, 2004 (in Russian).