MOARLE: Matrix operation accelerator based on run-length encoding

Abstract

Matrix computation is a key technology in various data processing tasks including data mining, machine learning, and information retrieval. Size of matrices has been increasing with the development of computational resources and dissemination of big data. Huge matrices are memory- and computational-time- consuming. Therefore, reducing the size and computational time of huge matrices is a key challenge in the data processing area. We develop MOARLE, a novel matrix computation framework that saves memory space and computational time. In contrast to conventional matrix computational methods that target to sparse matrices, MOARLE can efficiently handle both sparse matrices and dense matrices. Our experimental results show that MOARLE can reduce the memory usage to 2% of the original usage and improve the computational performance by a factor of 124x. © 2014 Springer International Publishing Switzerland.