面向高性能异构众核架构的大规模CFD并行算法与应用

项目名称： 面向高性能异构众核架构的大规模CFD并行算法与应用

项目编号： No.11502296

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 数理科学和化学

项目作者： 徐传福

作者单位： 中国人民解放军国防科技大学

项目金额： 20万元

中文摘要： 高性能异构众核架构的出现和发展给大规模CFD应用数值模拟提供了前所未有的机遇，也带来了前所未有的挑战。传统CPU平台上的CFD并行计算方法无法适应异构众核这一新型并行架构特点，迫切需要从CFD数值模拟的数值模型和算法特点出发，紧密结合异构架构特征，针对性地开展并行计算研究，使应用程序充分发掘大规模异构众核高性能计算机性能。本项目面向多区网格CFD应用选择典型显式、隐式求解方法，结合GPU/MIC等主流高性能异构众核平台体系结构和编程模型特点，研究CFD求解方法的多层次可扩展众核并行算法，发展CFD应用和众核体系结构特征相融合的性能优化方法，突破异构众核高效协同并行CFD计算一系列关键理论、技术，并将研究成果应用于自主知识产权CFD程序在国产异构众核超级计算机上的异构并行实现，以提升CFD数值模拟能力，满足CFD应用发展需求。

中文关键词： 异构并行；众核架构；负载平衡；多区网格；隐式方法

英文摘要： The emergence and development of high performance heterogeneous many-core architecture has brought enormous opportunities, and meanwhile poses an unprecedented challenge, for efficient numerical simulations of large-scale CFD applications. The algorithms and methods for parallel CFD computing on traditional CPU-enabled platforms are no longer suitable for the new parallel architecture. It is essential to develop new parallel CFD computing techniques to fully exploit the performance of large-scale heterogeneous many-core HPC systems, incorporating both the characteristics of CFD applications and many-core systems. In this proposal, based on the characteristics of architectures and programming models of popular high performance many-core accelerators/co-processors such as GPU and MIC, we target at designing and developing multi-level scalable parallel algorithms, application/architecture-aware optimization methods and other key techniques for efficient heterogeneous and collaborative computing involving typical explicit and implicit methods in multi-block CFD. Furthermore, we will apply our research works to port our in-house CFD programs onto China’s leading-edge many-core supercomputers, enhance the capability of CFD numerical simulations and meet the requirements of CFD applications.

英文关键词： heterogeneous parallel computing;many-core architecture;load balance;multi-block grids;implicit method