摘要
移动粒子半隐式算法(MPS)是一种有效的处理自由表面流动的粒子法,计算量大是该方法的主要缺点,并行化是解决此问题的有效途径,对于MPS方法,压力泊松方程的求解是并行算法的重点和难点,因为这是MPS算法中最耗时的部分,且串行算法中采用的不完全乔莱斯基共轭梯度(ICCG)算法难以并行,本文引入了一种高效且易于并行的对称兰乔斯算法(SLA)求解压力泊松方程,在多核服务器节点上采用共享内存的OpenMP模型、在多个节点之间采用消息传递的MPI模型完成了MPS算法的并行,溃坝问题的并行结果表明,在100核的服务器上,本文算法获得了37.5倍的加速比.
The Moving Particle Semi-implicit Method （MPS） has been proven useful in simulation of free-surface flows. One of its main drawbacks is the high computational load. The solution for the Pressure Poisson Equation （PPE） is the critical issue in parallelizing the MPS algorithm because it is the most time-consuming part and the original ICCG solver is difficult to parallelize. The Symmetric Lanczos Algorithm （SLA）, which is easy to parallelize and as effective as ICCG, is introduced to solve PPE. After decomposing computational domain into subdomains which are one by one assigned to different nodes of a cluster, the MPI model is used to exchange the information of particles between the adjacent subdomains and the OpenMP model is used to accelerate the simulation in each node. It is shown that MPS can be accelerated by 37.5 times on a cluster with 100 CPU cores by the hybrid OpenMP/MPI model.
出处
《工程热物理学报》
EI
CAS
CSCD
北大核心
2013年第11期2077-2080,共4页
Journal of Engineering Thermophysics
基金
国家自然科学基金项目(No.50976087)
高等学校博士学科点专项科研基金资助项目(No.20090201110001)
