Mingkai Dong and Haibo Chen, Institute of Parallel and Distributed Systems, Shanghai Jiao Tong University
Fast, byte-addressable NVM promises near cache latency and near memory bus throughput for file system operations. However, unanticipated cache line eviction may lead to disordered metadata update and thus existing NVM file systems (NVMFS) use synchronous cache flushes to ensure consistency, which extends critical path latency.
In this paper, we revisit soft updates, an intriguing idea that eliminates most synchronous metadata updates through delayed writes and dependency tracking, in the context of NVMFS. We show that on one hand byte-addressability of NVM significantly simplifies dependency tracking and enforcement by allowing better directory organization and closely matching the per-pointer dependency tracking of soft updates. On the other hand, per-cache-line failure atomicity of NVM cannot ensure the correctness of soft updates, which relies on block write atomicity; page cache, which is necessary for dual views in soft updates, becomes inefficient due to double writes and duplicated metadata. To guarantee the correctness and consistency without synchronous cache flushes and page cache, we propose pointer-based dual views, which shares most data structures but uses different pointers in different views, to allow delayed persistency and eliminate file system checking after a crash. In this way, our system, namely SoupFS, significantly shortens the critical path latency by delaying almost all synchronous cache flushes.We have implemented SoupFS as a POSIX-compliant file system for Linux and evaluated it against state-of-the-art NVMFS like PMFS and NOVA. Performance results show that SoupFS can have notably lower latency and modestly higher throughput compared to existing NVMFS.
Open Access Media
USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.