An RNN Based Mechanism for File Prefetching
Hui Chen, Enqiang Zhou, Jie Liu, Zhicheng Zhang · 2019
Many applications spend a large proportion of the execution time to access files. To narrow the increasing gap between computing and I/O performance, several optimization techniques were adopted, such as data prefetching and data layout optimization. However, the effectiveness of these optimization processes heavily depends on the understanding of the I/O behavior. Traditionally, spatial locality and temporal locality are mainly considered for data prefetching and scheduling policy. Whereas for most real-world workloads, the file access pattern is hard to capture. For the goal of deeply and intelligently understanding the I/O access pattern of modern applications, and efficiently optimizing the performance of current file systems, we propose a new mechanism to embed file names to vectors and train a gated recurrent neural network to provide policies for file prefetching and cache replacing.