A unified object-oriented framework for CPU+GPU explicit hyperbolic solvers. (October 2020)

• Record Type: Journal Article
• Authors: Conde, Daniel A.S.; Canelas, Ricardo B.; Ferreira, Rui M.L.
• Is Part Of: Advances in engineering software. Volume 148(2020)
• Publication Date: 2020-10
• Subjects: Hyperbolic conservation laws -- Parallel computing -- Contextual object splitting -- OMP -- CUDA; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2020.102802
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: Development of CPU+GPU hyperbolic solvers is unified under a simple object-oriented framework. Benchmarks show how data-structure layouts have significant impacts in code scalability. Coherent memory-space ordering and thread-interleaving techniques are used to improve constant and sparse workload performance. CPU parallel performance reveals supra-linear speedups on hyper-threading enabled processors. Speedup on GPUs is around 40x relatively to sequential CPU performance for the tested setup. Abstract: A unified design solution for heterogeneous explicit hyperbolic solvers is herein introduced. The proposed design is entirely cross-compatible between CPUs and GPUs, through an intuitive object-oriented approach. The advantages of a unified CPU+GPU development approach are discussed and exemplified, and a complete description of the data and code structures are provided and benchmarked. The benefits of different object-oriented designs are quantified under static and dynamic loads in terms of parallel performance and scalability. A fair comparison with graphics processors provides a realistic measure of achievable GPU implementation benefits. Both automatically and manually tuned GPU executions are compared and shown to also have a significant impact on the obtained performance. Overall, the proposed design combines a good sequential performance with a supra-linear scalability on modern CPUs. On GPUs, execution is shown to be up to 40 times faster than its single-threaded counterpart, opening a wider range of applicable model scales and resolutions.