In file systems, a single write system call can make multiple modifications to data and metadata, but such changes are not flushed in an atomic way. To retain the consistency of file systems, conventional approaches guarantee crash consistency in exchange for sacrificing system performance. To mitigate the performance penalty, non-volatile memory (NVM) technologies are believed to be a good candidate for this purpose owing to their low latency and byte-addressibility. However, none of the prior proposals that exploit NVM provide both scalability and strict ordering of modifications. In this article, we propose montage, a crash consistency framework for file systems that consists of two parts. First, montage splits NVM space into multiple staging areas and synchronizes the flushing of modifications these to the storage device. Second, montage uses the pipelined architecture to speed up data flushing to storage. We apply montage to two journaling file systems (ext4 and JFS) and evaluate them on a multicore server with high-performance storage. The evaluation results demonstrate that our design exhibits better performance by a wide margin than recent NVM-based journaling file systems.

在文件系统中，单个 write 系统调用可以对数据和元数据进行多次修改，但此类更改不会以原子方式刷新。为了保持文件系统的一致性，传统方法以牺牲系统性能为代价来保证崩溃一致性。为了减轻性能损失，非易失性存储器(NVM) 技术因其低延迟和字节可寻址性而被认为是用于此目的的良好候选者。然而，之前利用 NVM 的提议都没有提供可扩展性和严格的修改顺序。在本文中，我们提出了montage，这是一个由两部分组成的文件系统崩溃一致性框架。一、蒙太奇将 NVM 空间分成多个暂存区，并将这些修改同步刷新到存储设备。其次，蒙太奇使用流水线架构来加速数据刷新到存储。我们将蒙太奇应用于两个日志文件系统（ext4 和 JFS），并在具有高性能存储的多核服务器上对其进行评估。评估结果表明，我们的设计比最近的基于 NVM 的日志文件系统表现出更好的性能。