基于MapReduce的增量矩阵乘法设计  

Design of Incremental Matrix Multiplication Based on MapReduce

作　　者：张沛然[1] 徐云[2] 

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230027 [2]中国科学技术大学安徽省高性能计算重点实验室,合肥230027

出　　处：《计算机系统应用》2016年第9期176-182,共7页Computer Systems & Applications

基　　金：国家自然科学基金(61033009)

摘　　要：矩阵乘法是许多应用中的核心计算,在这些应用中只是少量矩阵元素发生改变,如果全量重新计算则工作量很大,因此增量计算是解决该问题的有效手段.本文提出了一种基于Map Reudce模型的增量矩阵乘法计算方法,以及计算矩阵中变化元素的高效识别方法,通过利用矩阵元素的摘要信息快速计算出变化元素,然后将矩阵乘法计算过程转换为一系列等价的连接问题,实现了一种有效的矩阵乘法增量计算.对于矩阵元素变化率较小的情形,计算实验表明提出的方法计算时间上明显优于全量重新计算方法.Matrix Multiplication is a core computation in many applications, it has been observed that only a very small fraction of the matrix elements gradually change in these applications, and the full recomputation costs a lot, thus, incremental processing is a promising approach to handle this problem. This paper proposes a design of incremental matrix multiplication based on MapReduce with a fast method to capture the changes of the matrix, by using the abstract informationof matrix elements to calculate the changes of the matrix effectively and transforms this problem to series of equivalent join problems, which realizes a efficient matrix multiplication in an incremental way. Experimental results show significant performance improvements of our method compared to full recomputation when the change ratio is small.

关 键 词：矩阵乘法 增量计算 MAPREDUCE 

分 类 号：O151.21[理学—数学]