Highly scalable implementation of an implicit matrix-free solver for gas dynamics on GPU-accelerated clusters

详细信息    查看全文

• 作者：Igor Menshov ; Pavel Pavlukhin
• 关键词：CFD ; LU ; SGS ; Parallel algorithms ; CUDA ; MPI
• 刊名：The Journal of Supercomputing
• 出版年：2017
• 出版时间：February 2017
• 年：2017
• 卷：73
• 期：2
• 页码：631-638
• 全文大小：
• 刊物类别：Computer Science
• 刊物主题：Programming Languages, Compilers, Interpreters; Processor Architectures; Computer Science, general;
• 出版者：Springer US
• ISSN：1573-0484
• 卷排序：73

文摘

A numerical approach for solving gas dynamics on Cartesian grids is considered which employs an implicit time marching scheme with the matrix-free Lower-Upper Symmetric Gauss–Seidel (LU-SGS) method for solving discrete equations. Boundary conditions are treated with an embedded-boundary method. The method has two attractive features—(1) algorithmic uniformity of calculations and (2) structured memory accesses that well fit massively parallel architectures with GPU accelerators. We propose a novel CUDA+MPI computational algorithm scalable up to hundreds of GPUs and give in-depth analysis of its implementation (interoperability issues, libraries tuning).