Iterative Algorithm’s Parallelism And Otimization Based On Data Partitioning

Iterative algorithm is a classical algorithm, has been widely used in real life, such as scientific computing, data mining algorithms. With the advent of the era of large data, the data quantum that can be obtained increased greatly and also the strictly time requirement. How to devise an machanism that can not only ensure the accuracy re-quirment but also meet the time requirment is what many scholars concern, in this paper we complete two algorithms optilization and parallelism of this kind.The first application, FOR3D optimization and parallelism that we complete will be introduced at the very beginning. Three-dimensional acousic propagation an impor-tant research topic in the marine survey and the military. As sound is a kind of wave, acousic propagation satisfies the wave equation, which is a high order partial differ-ential equation. It’s almost impossible to give an accurate solution to such kind of equation.As an numerical solution to such problem, FOR3D plays an important role in solving such equations.However, with the increasingly strict demand of result pre-cision,computation time becomes a significant performance bottleneck. Based on the original serial program FOR3D, we designed corresponding parallel FOR3D version. By calculating the propagation loss, FOR3D parallel program results show that, within the accuracy requirement, the parallel version decreases the calculation time with the number of processors increasing. And we almost get a linear acceleration.Then the second application, K-means optimization and parallelism. K-means al-gorithm, one of the most important member of the clustering algorithms, has a wide range of applications for its simple and elegant design mechanism. In this paper, we proposed a kind of new initial points selection machanism, which combine K-means++algorithm with point density. In the new machanism, we also use the Mean Shift algorithm to reduce the iteration time. The new machanism is validated on UCI data sets, and the results show that the mechanisms we design in most cases, can at least gain the same performances as K-means++do. Finally, a parallel version of the latter zlgorithm is proposed. The parallel version algorithm results show that in large amount of data situation, the experiment can achieve great speedup, for the data transformation time can be ignored.