Using blocks correlations to improve the I/O performance of large network storage system

Abstract

In the large network storage system, the operation of continuously reading discrete small blocks severely impacts the I/O performance. To solve this problem, this paper designs and implements a system prototype, which implements precise prefetch and regulates the data distribution according the small blocks correlations, mined by a novel heuristic algorithm between the file system and block device. The system performance can be improved evenly and continuously without interruption and sudden state transitions. Furthermore, compared with other algorithms, this heuristic algorithm thinks about both the locality and the globality of the correlations. Through the experiments, it has been proved that the prototype and the algorithm are effective and the system I/O performance can be enhanced distinctly. Furthermore, the prototype can be used universally by not modifying the file system and the storage devices © Springer-Verlag Berlin Heidelberg 2005.