CDOS: Block Correlations Directed I/O Performance Optimizing System

Data I/O has become the bottleneck of distributed network servers. This paper proposes a block Correlations Directed Optimizing System (CDOS), which provides high ability of mining block correlations and optimizes the I/O performance based on the mining results, to improve the I/O performance of distributed servers. CDOS proposes a novel block correlations mining method based on noise filter and uses block correlations to optimize the policy of prefetching, cache replacement and data layout on the I/O path of the distributed systems. For most real system workloads, CDOS can reduce the overhead of mining by one or two orders compared with traditional approach and decrease the average of system response time by 50%~79%.