Optimization of Reading Data via Classified Block Access Patterns in File Systems

Abstract

This paper proposes a novel mechanism to model a time series of block accesses for profiling block access patterns, to intentionally direct block data prefetching. The basic idea behind this scheme is that the block accesses within a certain offset domain may have some correlations that may contribute to classify an access pattern on the file system level. Moreover, the technique of adjacency matrix is employed to represent an access pattern for accelerating pattern matching, and then benefits I/O optimization eventually. Through a series of emulation experiments based on several realistic block traces on the disk, the experimental results show that the newly proposed prefetching mechanism outperforms other comparisons. Specifically, it can reduce average I/O response time by 14.6%-17.9% in contrast to the commonly used sequential prefetching scheme, and 4.1%-10.5% compared with frequent sequence mining-based prefetching but with less space and time overhead.