大规模并行粒子模拟中基于主从架构的负载平衡策略研究与实现

Design and implementation of a load balancing method in large-scale particle simulation based on master/slave model

粒子方法作为一种直观而且普适的模拟方法,在化工、材料、生物等领域得到了广泛的应用。本文采用主从(Master／Slave)并行模型,基于正交递归对分法(recursive coordinate bisection,RCB)和规则网格,设计出一种适用于具有多种复杂粒子的粒子系统并行模拟的负载平衡方法。应用该方法,在集群系统上采用离散元方法模拟了某钢铁公司滚筒系统。模拟采用4个节点,根据是否利用负载平衡模块分别进行测试,计算结果分析表明采用本文的负载平衡策略,计算效率提高了约14％。本文的算法具有较好的通用性,可以方便地应用于其它粒子方法的大规模并行计算问题。

As an intuitionistic and generic approach, particle method is widely used in many fields, such as chemical engineering, material, and biology. In this paper, a parallel algorithm with dynamic load balancing （DLB） suitable for large scale simulation of systems composed of different kinds of particles is presented. The algorithm adopts Master/Slave parallel model and is based on recursive coordinate bisection （RCB） and regular cell. It is applied to the parallel simulation of the tumbling barrel system in a steel corporation on cluster of workstation with discrete element method （DEM）. Compared with the simulation without DLB, the simulation can improve the parallel efficiency by about 14 percent. The algorithm also has good generality and can be expediently applied to the parallel simulation of other particle methods.