Fast journaling made simple with NVM

초록

Journaling file systems (JFS), which base their architecture on traditional database logging, write journals sequentially through a form of centralized logging, primarily to restore file systems in the face of an unexpected failure. Due to the extra write operations required for journaling, the performance of the JFS is negatively impacted, thus suffering performance degradation. The fast non-volatile memory (NVM) was expected to easily address such performance issues by providing NVM space as storage for journals. However, we observe that replacing disk space with NVM for journaling naively would inevitably face a scalability problem inherent in the transaction management of journaling file systems, even if we are able to use byte-addressable NVM. To address this problem, we propose a scalable file system that manages file system transactions using concurrent data structures combined with NVM-based decentralized journaling for scalable performance. In addition, we enable multiple I/O requests to be issued to high-performance storage devices supporting multiple I/O channels. We evaluate our file system on a multicore server with high-performance storage. Evaluation results demonstrate that our file system exhibits better performance by a wide margin than the vanilla ext4 file system and a recent NVM-based journaling file system.

키워드