Load Balancing for Multi-Physics Simulations
Chunsen Chen, Kenli Li, Xu Liu, Keqin Li · 2019
Multi-physics simulations are usually essential to simplify researches on complex physical phenomena. In this paper, we extend the rectangular partitioning from single-physics simulations to multi-physics simulations. This work presents the Single-rectangle Tightly-coupled algorithm (ST) and Multirectangle Tightly-coupled algorithm (MT) based on rectangular partitioning. The MT algorithm is more flexible and achieves higher load balancing efficiency than the ST algorithm by greedily assigning the remaining unallocated most loaded rectangle to the least loaded processor and assigning k rectangles to each of the processors. To study the effect of machine characteristics on load balancing, the experiments use two types of machine parameters based on Tianhe-2 supercomputer. Inspired by the roofline model, this paper analyzes the relationship among the choice of the load balancing algorithms, the load characteristics of multi-physics application and the features of high performance parallel computer through the comparison experiment of looselycoupled algorithm and tightly-coupled algorithm. In the case of LARED-P, MT algorithm can further reduce the total simulation time compared with the ST algorithm. Through experiments, we show that whether the load balancing algorithm is effective depends on both the load characteristics of multi-physics simulation and the features of high performance parallel computer.