项目名称: 面向众核计算的数值方法协同设计--一种高效且高精度广义有限元方法研究
项目编号: No.11472274
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 数理科学和化学
项目作者: 田荣
作者单位: 北京应用物理与计算数学研究所
项目金额: 98万元
中文摘要: 超级计算机浮点运算能力的迅速提高和数据移动能力的缓慢增长已经形成了非常明显的浮点效率鸿沟。浮点效率鸿沟反映了传统数值算法(应用数学学科)与新兴硬件结构(计算机学科)之间发展的不平衡和不匹配。在目前的新兴众核处理器上,一个高效的数值算法应该使单位访存所完成的浮点运算次数最大化,从而尽可能地享受由这一轮技术变革带来的新的免费午餐--超强浮点运算能力。这极可能将导致数值算法设计在思路上、甚至原则上的根本性转变。本项目首次尝试面向新兴计算机体系结构,从充分释放众核处理器冗余计算能力的角度出发,发展一种新的高效且高精度(无额外自由度)的广义有限元方法,并针对无网格/粒子/广义有限元法等一类计算较为密集的数值算法,深入研究这种硬件效能和数值精度双赢的新型计算模式的普适性。
中文关键词: 计算力学;无网格法;广义有限元;并行计算;扩展有限元
英文摘要: The work is motivated by the challenge in fully utilizing high floating point performance of the emerging manycore architecture. A fundamental trend of computer architecture evolving towards exaflop/s (10^18 floating point operations (flops) per second) is the fast increasing floating point performance (so-called free flops) accompanied by much slowly increasing the capability of data movement. Numerical simulation is facing a challenge posed by the unbalanced increase in the compute power and the bandwidth. The tomorrow's numerical simulation is likely to be subject to a bandwidth constraint instead of a performance constraint. A numerical method of PDEs should be designed to request more flops per unit data movement so that the free flops are not sitting idle and wasted. A meshfree method is able to increase arithmetic intensity for the same number of dofs by increasing the influence radius or by increasing the accuracy of approximation and therefore it may bear a potential in utilizing the redundant floating point capability of the emerging manycore processor. This idea will be tried out on a GFEM without extra dof in this project.
英文关键词: Computational Mechanics;Meshfree;GFEM;Parallel Computing;XFEM