Load Balancing Algorithm In The Large-scale CFD Applications With Multi-zone Structured Grids

Computational Fluid Dynamics(CFD) have been widely used in every fields nowadays, such as automobile manufacturing, meteorology, aerospace, rapid transportation, chemical industry and other fields. Development of modern high-performance computers provides a platform for large-scale CFD simulation. How to take advantage of the supercomputing power of high-performance computers to improve the parallelism and scalability of large-scale CFD programs and to complete numerical simulation in a shorter time is a serious challenge. For example, how to reach the best load balancing performance in large-scale CFD parallel applications, especially how to repartition the multi-block structured grids and distribute them across processors is still a difficult problem that need to be solved immediately.This paper aimed at the load balancing of parallel CFD simulations with structured grids on heterogeneous HPC platforms. The main work and contributions are as follows:1. As to the case that the number of grid zone larger than the number of processes in load balancing of parallel CFD simulations with multi-block structured grids, we take advantage of the good search performance of genetic algorithm, a work distribution strategy of multi-zone structural grids based on genetic algorithm is proposed. We abandoned the traditional method that measures communication overhead between different processes by the communication traffic between them, and propose a new communication overhead model instead, which can adapted to the new heterogeneous high performance computation platform. Experiment results show that when the number of grid zone is much larger than the number of processes, the new algorithm has a better performance in communication overhead as to the traditional greedy load balancing algorithm.2. Aimed to the grid repartitioning in load balancing of parallel CFD simulations with multi-block structured grids, we developed both a regular multidimensional repartitioning algorithm and a smart multidimensional repartitioning algorithms. The resulting grid blocks have shapes closer to “cube” by using the new algorithms, which can lead to lower ratio of communication to computation. This means they can reduce the amount of inter-process communication effectively. Smart multidimensional repartitioning algorithm can generate less connection boundaries than the traditional algorithm does, which reduced the number of inter-process communication.3. Aimed to the problem of parallel CFD application with multi-block structured grids that has arbitrarily distribution of grid block volume, we proposed a new large-scale parallel task assignment strategy. It includes three steps: repartitioning the original grids, combining the resulting grid blocks based on genetic algorithm, and minor adjusting. The test result of parallel numerical simulations for large scale CFD applications show that our algorithm can adapted to the general cases with any distribution of sizes of grid blocks. It can be smoothly used for the heterogeneous platforms, and has a better performance in load balancing compared to the traditional load balancing algorithm.4. We implement the above algorithm and developed a software tool for the load balancing of multi-zone structural grid CFD parallel computation. Among others, easy-to-use is one of its advantages.