基于面向对象的粒子类模拟并行计算研究

针对经典分子动力学和PIC方法等粒子类模拟方法具有粒子动态移动、粒子计算局部性好等共性,首先,提出了粒子量数据片对象.该对象是单网格片上的一团粒子,其中网格片是包含多个网格单元的矩形区域.然后,设计了并行算法,包括对象之间的粒子迁移和数据交换以及动态负载平衡.最后,在JASMIN框架上具体实现,进而开发了并行经典分子动力学程序和并行PIC程序.在64个处理器上实测表明,并行PIC程序模拟包含3百万个网格、2千万个粒子的复杂物理模型时,获得了80%的并行效率.

Parallel Computation for Particle Simulations Based on Object-Oriented Design

Particle simulations including classical molecular dynamics(MD) and particle-in-cell(PIC) have some common characteristics,which include dynamic movement of particles,spatial locality of particle computation,etc.In consideration of these characteristics,firstly,particle-patch-data object is presented.The object is a group of particles in a single patch.The patch is a rectangular block including many cells.Secondly,a corresponding parallel algorithm is designed.The algorithm deals with dynamical load balancing,particles migration and data exchange among these objects.Finally,the algorithm is implemented on JASMIN infrastructure.Then,parallel classical molecular dynamics program and parallel PIC program are developed.Numerical experimental results of PIC program show that the parallel efficiency with 80% is demonstrated by simulating a complex physical model including 3 million cells and 20 million particles on 64 processors of one MPP.