The Design and Implementation of an Efficient Data Consistency Mechanism for In-Memory File Systems

Abstract

The Non-Volatile Memory (NVM) based in-memory file systems show great potential in supporting real-time data processing for their extremely high performance. The reliability of file systems is ensured by data consistency mechanisms. The existing data consistency mechanisms, however, largely degrades the performance of the in-memory file system without fully exploiting the characteristics of NVM. In this paper, we propose an efficient data consistency mechanism, Amphibian Update Strategy (AUS), taking advantages of the virtual address space of NVM. In the proposed AUS technique, the backup spaces of file data are organized and accessed by the contiguous virtual address space of the kernel. We present the Direct-Copy and Exchanging approaches to efficiently update the primary file data for the requests with different sizes. We implemented different data consistency mechanisms in a real in-memory file system, SIMFS. Extensive experiments are conducted. The experimental results show that AUS achieves 2.4 times, 1.8 times, and 1.7 times faster than the legacy journaling, short-circuit shadow paging, and the state-of-the-art technique adaptive logging, respectively.