Work

Beaver: A High-Performance and Crash-Consistent File System Cache via PM-DRAM Collaborative Memory Tiering

Year: 2025

Type: article

Abstract: The in-memory cache layer is crucial in building a file system. Crash-consistency is highly desirable for applications running on the file system, ensuring that data is written in an all-or-none fashion during unexpected system failures or crashes. However, existing works fail to achieve both high read and write performance in constructing a crash-consistent cache layer. In this paper, we propose Beaver, a new in-memory file system cache that achieves both crash-consistency and high read/write performance. Beaver exploits a read/write distinguishable memory hierarchy involving both PM and DRAM. Beaver uses the PM to serve write operations for crash-consistency and high write performance, and uses the DRAM to serve read operations for high read performance. We also introduce a kprefetcher to asynchronously move data from PM to DRAM off the critical path of write operations. We extensively evaluated Beaver with the latest file system cache using real-world applications and popular benchmarks. Application evaluation results show that Beaver outperformed other file system caches by up to 16.1% and 11.0% on RocksDB and MinIO, respectively. Beaver's source code is available at https://github.com/PanQL/Beaver.

Source: ACM SIGMETRICS Performance Evaluation Review

DOI: 10.1145/3744970.3727273

Authors: Qinglin Pan, Ji Qi, Jiatai He, Heng Zhang, Jiageng Yu, Yanjun Wu

Institutions: Chinese Academy of Sciences, Institute of Software, University of Chinese Academy of Sciences, University of Hong Kong

Language: English

Topic: Advanced Data Storage Technologies

Subfield: Computer Networks and Communications

Field: Computer Science

Domain: Physical Sciences

Open Access status: bronze

Funder: Universitas Brawijaya

Award: G1246308732