Slibfs: Towards Small Write Optimization in Per-File Virtualization NVM-based Filesystems

Non-volatile memory (NVM) technologies can provide both data persistence and high performance at memory level. However, prevalent kernel-based and userspace filesystems alike fall short in adequately addressing the challenges posed by small writes, despite the dominance of small write operations as revealed by analyses of modern filesystem workloads. These previous filesystems either compromise on write consistency or resort to Copy-on-Write methodologies, which incur substantial write amplification. In this paper, we propose a small write optimization method for NVM and implemented a prototype Slibfs. Leveraging an evolved per-file virtualization design, Slibfs innovates with write transaction isolation strategies, separating transactions by type to handle varying write sizes optimally. It further incorporates a Journal-Log decoupling structure to segregate write payload from transaction metadata, and establishes a dedicated pipeline for small write transactions (S Write Tx), ensuring rapid and consistent write operations. Extensive evaluations confirm Slibfs’s effectiveness, exhibiting a 2.25x remarkable throughput improvement over MadFS, specifically for small write scenarios. Hence, Slibfs marks a pivotal step towards realizing high-performance, reliable NVM filesystems that adeptly cater to the unique challenges of small write workloads.