千万亿次可扩展可容错自由网格数值模拟系统

发布时间：2018-06-28 23:35

本文选题：千万亿次计算 + 无网格/粒子模拟　；参考：《计算机研究与发展》2015年04期

【摘要】：在千万亿次计算能力的驱动下,数值软件的发展进入了一个以海量并行为基本特征的历史转折期,可扩展和可容错成为大规模数值模拟的两大关键技术.petaPar模拟程序是以对传统数值技术形成优势互补的无网格类方法为切入点,面向千万亿次级计算而开发的下一代新兴通用数值模拟程序.petaPar在统一架构下实现了光滑粒子动力学(smoothed particle hydrodynamics,SPH)和物质点法(material point method,MPM)两种最为成熟和有效的无网格/粒子算法,支持多种强度、失效模型和状态方程;其中MPM支持改进的接触算法,可以处理上百万离散物体的非连续变形和相互作用计算.系统具有以下特点:1)高可扩展.实现单核单Patch极端情形下计算和通信的完全重叠,支持动态负载均衡;2)可容错.支持无人值守变进程重启动,在系统硬件出现局部热故障时可以不中止计算;3)适应硬件体系结构异构架构的变化趋势,同时支持flat MPI和MPI+Pthreads并行模型.程序在Titan千万亿次超级计算机上进行了全系统规模的可扩展性测试,结果表明该代码可线性扩展到26万个CPU核,SPH和MPM的并行效率分别为100%和96%.
[Abstract]:Driven by the power of thousands of calculations, the development of numerical software has entered a historical turning point characterized by massive parallelism. Extensible and fault-tolerant are the two key techniques of large-scale numerical simulation. PetaPar simulation program is based on meshless class method, which is complementary to the traditional numerical technique. The next generation of general-purpose numerical simulation program. PetaPar, which is developed for thousands of billions of secondary computations, implements two of the most mature and effective meshless / particle free algorithms under the unified framework: smooth particle dynamics (smoothed particle) and material point method (material point method MPM). It supports a variety of strength failure models and equations of state in which MPM supports an improved contact algorithm which can deal with discontinuous deformation and interaction calculations of millions of discrete objects. The system has the following characteristics: 1) highly extensible. In the extreme case of single core Patch, the computation and communication are completely overlapped, and dynamic load balancing is supported. It supports restarting of unattended process, and can not stop computing in case of local thermal failure of system hardware. It adapts to the trend of heterogeneous architecture of hardware architecture, and supports flat MPI and MPI Pthreads parallel model at the same time. The program is tested on Titan thousand trillion supercomputers on a system-wide scale. The results show that the code can be linearly extended to 260000 CPUs. The parallel efficiency of SPH and MPM are 100% and 96% respectively.
【作者单位】：中国科学院计算技术研究所高性能计算机研究中心;中国科学院大学;
【基金】：国家自然科学基金项目(11072241,11111140020,91130026) 橡树岭国家实验室/美国国家计算科学中心主任基金项目(MAT028)
【分类号】：TP302.8

【参考文献】