AFPT: Accelerating Read Performance of In-Memory File System Through Adaptive File Page Table

Emerging non-volatile memory (NVM) technologies are expected to revolutionize storage systems by providing cheap, persistent and fast data accesses through memory bus interface. In oder to fully exploit NVM, many in-memory file systems are proposed to achieve excellent performance and strong consistency. Besides, to mitigate the read-write asymmetric problem of NVM, many optimization strategies are designed to hide the long write latency to NVM in critical path of file operations, such as path resolution. However, we find that the index structure of state-of-the-art in-memory file systems cannot provide fast read performance in various use scenarios. In this paper, we propose Adaptive File Page Table (AFPT), a novel index scheme that combines software search and MMU mapping to provide excellent read performance for different workloads. For small requests, software search routines are used to locate data pages by traversing the file index structure. For large requests, we allocate a continuous address space and build file page table to utilize hardware MMU for address translation. A Cost Model is proposed to determine when to build page table for a file. This model is 1.38-competitive against optimal solution. We implement AFPT in PMFS and NOVA and evaluate the performance with micro-benchmarks and application workloads. The experimental results show that AFPT improves file system performance by up to 55.62% and 41.78% for NOVA and PMFS, respectively.