Enhancing prediction accuracy in PCM-based file prefetch by constained pattern replacement algorithm

Abstract

In modern file systems, I/O latency is still major bottleneck of performance and predictive file prefetching is one of promising approaches that can enhance I/O performance of file system. To utilize predictive file prefetching to file system, there should be a file access pattern prediction model that can predict future file access. Partitioned Context Model(PCM) [2] is known as one of the most accurate file access pattern prediction models[3]. In order to predict longer sequence, the order of PCM must be increased. However, the prediction accuracy of PCM decreases when PCM is in high order. Careful analysis reveals that the pattern replacement algorithm in the PCM is the major cause in decay of the prediction accuracy. The pattern replacement algorithm destroys file access patterns without successful training of newly occurred file access patterns. We proposed the constrained pattern replacement algorithm to overcome this adverse effect by revising replacement condition. The simulation results using the DFSTrace system trace[13] show that the proposed algorithm improves prediction accuracy without any extra cost by 3.5% compared to traditional pattern replacement algorithm of PCM(about 40% of the accuracy bound of 7%).