高可扩展可容错的无网格/粒子程序petaPar及其测试  

A Scalable and Fault Tolerant Particle Simulation Code

作　　者：黎雷生[1,2] 田荣[1] 

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院大学,北京100049

出　　处：《科研信息化技术与应用》2013年第5期3-9,共7页E-science Technology & Application

摘　　要：petaPar粒子模拟程序面向千万亿次级计算,在统一框架下实现两种广受关注的粒子模拟算法:光滑粒子流体动力学(Smoothed Particle Hydrodynamics,SPH)和物质点法(Material Point Method,MPM)。代码支持多种材料模型、强度模型和失效模型,适合模拟大变形、高应变率和流固耦合问题。支持纯MPI和MPI+X混合两种并行模型。系统具有可容错性,支持无人值守变进程重启。在Titan上测试表明,petaPar可线性扩展到26万CPU核,SPH和MPM算法并行效率相对8 192核分别为87%和90%。Powered by petaflop supercomputers, numerical simulation steps into a completely new era, a new generation of simulation code is expected to explore the parallelism of hundreds of thousands of processor cores, petaPar is targeted at petascale particle simulation on petaflop systems. It unifies two most popular and powerful particle methods, the Smoothed Particle Hydrodynamics （SPH） and the Material Point Method （MPM）. The code supports a number of material models, strength models and failure models, and is suitable for large deformation, high strain rates and fluid-solid interaction. Parallel implementations support both flat MPI and MPI＋X hybrid parallel models. The code is highly fault tolerant in the sense that it can support unattended process restart from any time step. Scalability tests on Titan shows that the code is linearly scaled up to 260K CPUcores and delivers 87% and 90% parallel efficiency relative to 8 192 CPU cores for MPM and SPH respectively.

关 键 词：千万亿次计算 粒子模拟 SPH MPM 可扩展性 可容错性 

分 类 号：TP333.4[自动化与计算机技术—计算机系统结构]