×

Performance analysis of MG preconditioning on Intel Xeon Phi: towards scalability for extreme scale problems with fractional Laplacians. (English) Zbl 1476.65035

Lirkov, Ivan (ed.) et al., Large-scale scientific computing. 11th international conference, LSSC 2017, Sozopol, Bulgaria, June 5–9, 2017. Revised selected papers. Cham: Springer. Lect. Notes Comput. Sci. 10665, 304-312 (2018).
Summary: The Intel Xeon Phi architecture is currently a popular choice for supercomputers, with many entries of the Top 500 list, using it either as main processors or as accelerators/coprocessors. In this paper, we explore the performance and scalability of the Intel Xeon Phi chips in the context of large sparse linear systems, commonly arising from the discretization of PDEs. At the first step, the PCG [O. Axelsson, Iterative solution methods. Cambridge: Cambridge Univ. Press (1996; Zbl 0845.65011)] is applied as a basic iterative solution method in the case of sparse SPD problems. The parallel multigrid (MG) implementation from Trilinos ML package is utilized as a preconditioner. A matrix free algebraic multilevel solver is used to reduce the memory requirements, thus allowing the cores to be more efficiently utilized. The second part of the paper is devoted to the fractional Laplacian, that is, we consider the equation \(-\varDelta^\alpha\mathbf{u} = \mathbf{f}\), \(0<\alpha < 1\), \(\varOmega\subset\mathbb{R}^d\). The related elliptic boundary value problem describes anomalous diffusion phenomena also referred to as super-diffusion. The implemented method approximates the solution of the nonlocal problem by a series of local elliptic problems. The currently available numerical methods for fractional diffusion Laplacian have computational complexity, comparable e.g., to the complexity of solving local elliptic problem in \(\tilde{\varOmega}\subset\mathbb{R}^{d+1}\). The presented parallel results are for \(\varOmega =(0,1)^3\), including meshes of very large scale. The numerical experiments are run on the Avitohol computer at the Institute of Information and Communication Technologies, IICT-BAS. The presented results show very good scalability when the CPU-cores and MIC work together for a certain number of compute nodes.
For the entire collection see [Zbl 1435.65014].

MSC:

65F08 Preconditioners for iterative methods
65N55 Multigrid methods; domain decomposition for boundary value problems involving PDEs
65Y10 Numerical algorithms for specific classes of architectures

Citations:

Zbl 0845.65011

Software:

Trilinos
PDFBibTeX XMLCite
Full Text: DOI

References:

[1] Axelsson, O.: Iterative Solution Methods. Cambridge University Press, New York (1996) · Zbl 0845.65011
[2] Bonito, A., Pasciak, J.: Numerical approximation of fractional powers of elliptic operators. Math. Comput. 84, 2083-2110 (2015) · Zbl 1331.65159 · doi:10.1090/S0025-5718-2015-02937-8
[3] Ciegis, R., Starikovicius, V., Margenov, S., Kriauzienė, R.: Parallel solvers for fractional power diffusion problems. Concurrency Comput. Pract. Exper. 29, e4216 (2017). https://doi.org/10.1002/cpe.4216 · doi:10.1002/cpe.4216
[4] Harizanov, S., Margenov, S.: Positive Approximations of the Inverse of Fractional Powers of SPD M-Matrices, arXiv:1706.07620, June 2017, submitted · Zbl 06962749
[5] Harizanov, S., Lazarov, R., Margenov, S., Marinov, P., Vutov, Y.: Optimal Solvers for Linear Systems with Fractional Powers of Sparse SPD Matrices, arXiv:1612.04846v1, December 2016, submitted · Zbl 1513.65132
[6] Chen, L., Nochetto, R., Enrique, O., Salgado, A.J.: Multilevel methods for nonuniformly elliptic operators and fractional diffusion. Math. Comput. 85, 2583-2607 (2016) · Zbl 1344.65117 · doi:10.1090/mcom/3089
[7] Vabishchevich, P.N.: Numerically solving an equation for fractional powers of elliptic operators. J. Comput. Phys. 282, 289-302 (2015) · Zbl 1352.65557 · doi:10.1016/j.jcp.2014.11.022
This reference list is based on information provided by the publisher or from digital mathematics libraries. Its items are heuristically matched to zbMATH identifiers and may contain data conversion errors. In some cases that data have been complemented/enhanced by data from zbMATH Open. This attempts to reflect the references listed in the original paper as accurately as possible without claiming completeness or a perfect matching.