Abstract
Storage devices use write buffers to improve performance, where multiple write requests are processed in parallel and completed in a random order. This may result in data loss in the event of a sudden failure. Therefore, Linux filesystems provide the fsync() system call to prevent data loss and ensure write order. However, the fsync() system call in F2FS, one of the most popular filesystems, is inefficient and insufficient for guaranteeing data consistency. We propose a new technique called Order-Preserving Recovery with Write pointer (OPRW) to ensure data consistency during the filesystem recovery. OPRW utilizes the write pointer in Zoned NameSpace (ZNS) SSDs to efficiently determine the persistence of data without the need for I/O operations. This approach allows OPRW to provide a higher level of consistency and performance improvement for fsync(), while minimizing the impact on recovery time. As a result, our solution improves performance by up to 1.2x on realistic workloads.