A Scalable Parallel SPH Implementation for Distributed Computing

彬彬

1 Introduction

The Smooth Particle Hydrodynamics (SPH) method is a Lagrangian

approach. The fluid is approximated by particles that are

free to move around with physical properties (see Fig. 1, left).

A major limitation with distributed computing is the bottleneck associated

with cross machine communication. Lowering the communication

latency is evident for a good performance of parallel

systems. However, the SPH scheme unfortunately introduces considerable

communication among the nodes. This motivates our investigation

of lowering the communication latency with the strategy

of overlapping computation and communication.

In this paper, we separate the particle exchanging process into two

parts. Nodes first exchange the necessary small amount of particles

to perform the majority of inter-particle calculations. During these

calculations, the rest of particles which should be exchanged for the

next simulation step are exchanged among nodes. Therefore, the

communication latency can be almost hidden by overlapping dataexchanging

with computation. The simulation involving millions of

fluid particles were run with good performance (see Fig. 1, right).

2 Parallel Simulation

For parallel simulations, we employ a manager-worker model and

the ORB decomposition. The manager node maintains load balancing.

In ORB method, the basic idea of load balancing is to shift

sub-domain boundaries. Shifting boundaries causes particles to be

transferred to neighboring domains and thus leads to a distribution

of workload. The worker nodes perform the calculation of particle

interaction over the assigned areas. The nodes communicate each

other through MPI.

彬彬

奇

不错不错，收藏了

晃晃

凡系斑竹滴话要听；凡系朋友滴帖要顶!

奇

好铁多多发，感激分享

C.R.CAN

提醒猪猪，千万不能让你看见

奇

既来之，则看之！

C.R.CAN

真不错,全存下来了.

奇

都闪开，介个帖子，偶来顶

菜刀吻电线

呵呵,真得不错哦!!