Research Of Dynamic Load Balancing Model On The Parallel Computing Platform For Simulation Of Ion Transmission

With the rapid development of information society, a large amount of data need to be processed in a short time. It is hard for single machine series mode in traditional way, but parallel computing technology is one of the effective methods to solve such problems. Cluster is a kind of Parallel computing technology, which has some advantages, such as low investment, easy to install, quick to gain effect, etc. Because of these advantages, Cluster will be the master in the parallel computing area.This thesis stemed from the theoretical and experimental research in the space charge effect and the stability of high-current accelerators, which belonged to the Institute of Modern Physics, Chinese Academy of Sciences. In this paper, parallel cluster technology was used to accomplish the large numerical calculation of ion transmission by PC cluster nodes which cooperated with each other. We also proposed a dynamic load balancing model for heterogeneous cluster, which was based on the feedback of the load state changing, at the same time in realizing the parallel program of ion motion simulation. The main contributions of this paper were as follows:Firstly, considering the processes of ion transport, the writer designed a model in response to the simulation problem of the ion transmission, using Divide-and-Conquer algorithm and a method named PCAM(Partitioning, Communication, Agglomeration, Mapping) for designing parallel software. The writer designed and implemented a parallel program by MPI interface and C programming language according to the model.Secondly, the writer presented a DLBMSF(Dynamic Load Balancing Model based on Status Feedback)model based on the status feedback, and realized the model by taking advantage of the MPI message-passing mechanism. In this model, an integrated weight of available resource was the indicator for task dispatching, and an improved weighted round-robin scheduling algorithm called IWRRS were used for task distribution. The model adopted dynamic load balancing algorithm triggered by slave node's load status changing, which improved the parallel efficiency of cluster greatly. In order to guarantee the stability and reliability of cluster systems, the model also added periodic heartbeat detection mechanism.Lastly, how the ratio of computing time to communication time impacts on Parallel speedup is also talked abou in this paper. The writer combined the ion simulation parallel program with the DLBMSF model, experiment also tested that the DLBMSF can bring a great load balancing efficiency, comparing with the Round-Robin Scheduling provided by MPICH.