A GPU solver for symmetric positive-definite matrices vs. traditional codes. (1st November 2019)

• Record Type: Journal Article
• Authors: Bohacek, Jan; Kharicha, Abdellah; Ludwig, Andreas; Wu, Menghuai; Holzmann, Tobias; Karimi-Sibaki, Ebrahim
• Is Part Of: Computers & mathematics with applications. Volume 78:issue 9(2019)
• Page Start: 2933
• Page End: 2943
• Publication Date: 2019-11-01
• Subjects: Linear solver -- Inverse task -- Heat transfer -- GPU -- CUDA -- OpenFOAM; Electronic data processing -- Periodicals; Mathematics -- Data processing -- Periodicals; 510.28541
• DOI: 10.1016/j.camwa.2019.02.034
• Languages: English
• ISSNs: 0898-1221

Abstract

Abstract: In Heat Transfer and Fluid Flow Laboratory in Brno, the inverse heat conduction problem (IHCP) has been extensively used to reconstruct thermal boundary conditions at hot surfaces of solid materials cooled by spraying nozzles. More than three decades of experience and cooperation with industries has proven our experimental/numerical technique to be reliable and very accurate. However, a typical calculation requires relatively long calculation time. The transient heat diffusion in a multi-material sample is the most computationally costly ingredient of the algorithm. In the present paper, the potential for speeding up our calculations is manifested by firstly benchmarking it against traditional CFD codes such as OpenFOAM (FDIC) and ANSYS Fluent (AMG). Secondly, we also unveil a unique comparison between the performance of three in-house GPU codes each written by a different PhD student/postdoc. Chronologically listed, one student pushed his luck with a fully explicit scheme, while the other two, including us, bet on implicit methods namely the line-by-line method in OpenCL and the conjugate gradient method with the deflated truncated Neumann series preconditioner in CUDA C.