Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Computational Science, Engineering & Technology Series
ADVANCES IN PARALLEL, DISTRIBUTED, GRID AND CLOUD COMPUTING FOR ENGINEERING
Edited by: P. Iványi, B.H.V. Topping and G. Várady
A Multi-GPU Framework for Structural Optimization under Uncertainty
D. Herrero-Pérez and J. Martínez-Frutos
Computational Mechanics & Scientific Computing Group, Department of Structures and Construction, Technical University of Cartagena, Murcia, Spain
D. Herrero-Pérez, J. Martínez-Frutos, "A Multi-GPU Framework for Structural Optimization under Uncertainty", in P. Iványi, B.H.V. Topping and G. Várady, (Editors), "Advances in Parallel, Distributed, Grid and Cloud Computing for Engineering", Saxe-Coburg Publications, Stirlingshire, UK, Chapter 2, pp 9-27, 2017. doi:10.4203/csets.40.2
Keywords: GPU computing, multilevel parallelism, software frameworks, uncertainty, topology optimization.
This paper proposes a software framework to facilitate the development of algorithms for structural optimization under uncertainty on systems with multiple Graphics Processing Units (GPUs). The computational challenge of this problem requires the use of high performance computing techniques both to be able to address it and to achieve results in a reasonable amount of time. The use of a software framework that facilitates the implementation in a flexible and reusable way and permits the load balancing becomes crucial to take advantage of parallelism in off-the-shelf accelerator hardware, such as multicore and manycore accelerators. The modular design of the software framework permits the flexible design of the flowchart and the implementation of different functionalities for diverse hardware architectures. All these features facilitate the proper exploitation of multilevel parallelism provided by multi-GPU systems, which is of paramount importance to obtain reasonable performance. An instance of the proposal is presented in the numerical experiments, where task level parallelism is used to concurrently evaluate, through the nodes of a GPU cluster, the independent finite element simulations arising from non-intrusive uncertainty propagation methods. Such finite element models are then solved on the available GPUs exploiting the data level parallelism. The numerical experiments show how significant improvements in computational efficiency and scalability are achieved.
purchase the full-text of this chapter (price £25)