在多对称处理器集群体系结构下进行独立成分分析并行算法研究,在共享内存模型一级并行算法基础上,通过同步、异步迭代两种方式并行计算固定点函数,分别提出具有两级并行特性的二级同步、二级异步并行端元提取算法,并结合两者的优势,进一步提出二级分组并行算法。实验评价表明,二级同步、分组并行算法在保持原算法精度的同时,大大提高了原算法的效率,体现出良好的并行计算性能,而二级异步并行算法可在节点数较少的情况下适用。

This paper analyzes parallel ICA algorithm in symmetrical multi-processing (SMP) cluster architecture. Based on our proposed single-level share memory model parallel ICA algorithm, two-level synchronous and asynchronous parallel ICA algorithms are presented, respectively, in the manner of synchronous and asynchronous parallel iteration for computing fi xed-point function. By the use of these algorithms, two-level grouping parallel ICA algorithm is also proposed. In experiment with real hyperspectral remote sensing image, synchronous and grouping parallel ICA algorithms maintenance the endmember extraction accuracy with respect to the original algorithm. Meanwhile, they also demonstrate high parallel performance and greatly improve the performance of ICA endmember extraction. Asynchronous parallel ICA algorithm is suitable for the case of small number of nodes in cluster.