【摘要】 针对潜在语义分析中词汇-文本矩阵奇异值分解的特点,设计并实现了一种基于单边Jacobi的矩阵奇异值分解的并行算法。并行算法采用了一种新的扫描策略和任务划分策略,该策略在一次扫描中能产生n(n1-)/2个不同的列向量对,同时能够对矩阵的列向量按模排序,使奇异值按从大到小的顺序排列。通过在自强3000高性能计算机上的实验表明,并行算法大大缩短了奇异值分解的计算时间,而且随着矩阵规模逐渐变大,加速比趋于稳定。更多还原

【Abstract】 According to singular value decomposition(SVD) character of term-document matrix in latent semantic analysis(LSA),a parallel algorithm of SVD computation based on one-side Jacobi was designed and implemented.The new parallel scanning strategy and task partition strategy were adopted in the parallel algorithm,these strategies can generate n(n-1)/2 different column vector pairs once in a scanning operation and sort the column vectors in matrix on norms to queue the singular values in a sequence from big to small at the same time.Some experimental results attained on Ziqiang 3000 hi-performance computer show that this parallel algorithm greatly reduces the computation time of SVD,and its speedup ratio becomes stable along with the increase of matrix dimension.更多还原