Fast and Parallelized Crash Consistency with Opportunistic Order Elimination

File systems rely on enforcing storage order for crash recovery. However, this ordering requirement limits I/O parallelism, especially in persistent memory (PM) file systems where ordered I/Os are synchronized for direct access, preventing file systems from fully exploiting PM I/O parallelism. This paper proposes a new I/O mechanism, called opportunistic order elimination, to achieve fast and parallelized crash consistency in PM file systems. The key idea is to eliminate the order between user data and file system metadata via fast checksums, and overlap checksum computation and metadata I/O with parallel data writes, thereby fully exploiting the parallelism of PM. An array of techniques are further introduced to control, dispatch, and recover parallelized I/O for efficiency, concurrency, and reliability. Based on this, we build ChimeraFS atop Linux kernel. A range of benchmarks and applications suggest that ChimeraFS achieves reliable crash recovery, significantly outperforms state-of-the-art PM file systems (e.g., OdinFS), and potentially approaches the ideal parallel write bandwidth of PM.