IPLFS

Abstract

In this work, we develop the log-structured filesystem that is free from garbage collection. There are two key technical ingredients: IPLFS, a log-structured filesystem for infinite partition, and Interval Mapping, a space-efficient LBA-to-PBA mapping for infinite filesystem partition. In IPLFS, we separate the filesystem partition size from the physical storage size and set the size of the logical partition large enough so that there is no lack of free segments in the logical partition during SSD’s lifespan. This allows the filesystem to write the updates in append-only fashion without reclaiming the invalid filesystem blocks. We revise the metadata structure of the baseline filesystem, F2FS, so that it can efficiently handle the storage partition with 2^64 sectors. We develop Interval Mapping to minimize the memory requirement for the LBA-to-PBA translation in FTL. Interval Mapping is a three level mapping tree. It maintains mapping only for actively used filesystem region. With Interval Mapping, the FTL can maintain the mapping for the 264 sector range with almost identical memory requirement with the page mapping whose LBA range is limited by the size of the storage capacity. We implement the IPLFS on Linux kernel 5.11.0 and prototype the Interval Mapping in OpenSSD. By eliminating the filesystem level garbage collection, IPLFS outperforms F2FS by up to 12.8× (FIO) and 3.73× (MySQL YCSB A), respectively.