Log Structured Filesystem Revisited: In the era of Flash Based Storage

Abstract

The segment based append-only update nature of the log-structured filesystem extremely well addresses the filesystem requirement for NAND Flash storage. Despite its promising nature, however, few users actually use the log-structured filesystem to manage their storage in real-settings. This work provides effective solutions to four key technical issues that existing filesystems failed to address. We develop log-structured direct IO to support the direct IO mode of the filesystem write operation, and suspend-aware segment cleaning to seamlessly integrate the background segment cleaning with suspend mode of the mobile device. We find that the most log-structured filesystems are subject to segment thrashing, while F2FS handles fsync() efficiently. According to our extensive experiment, with 10% filesystem free space, the segment cleaning overhead becomes negligible. We develop Disaggregate Mapping which is specifically tailored for log-structured filesystem that not only reduces the mapping table size to 1/100 compared to page mapping, but also manage the same lvel of write amplification. We perform comprehensive performance experiment of the developed schemes in server platform (single SSD, RAID 0 and RAID 5) as well as smartphone (Motorolla Moto G). The log-structured direct IO brings as much as × 6 performance improvement in random write performance in RAID 5 against the EXT4. Suspend-aware segment cleaning reclaims × 6 more segment than the stock segment cleaning in smartphone.