Redesigning Data and Metadata Updates in PM File Systems with Persistent CPU Caches

Abstract: The recent advent of persistent memory has revolutionized file system design by enabling efficient, fast, and durable data updates. However, to cope with mismatched access granularities and volatile CPU caches, existing file systems resort to costly data and metadata update approaches, leading to high write-back latency and excessive PM bandwidth consumption. Recent cache persistence techniques, such as Intel’s eADR, address this issue by automatically flushing data from CPU caches to PM during power failures, making CPU caches effectively persistent. Persistent CPU caches provide both cacheline access granularity and data durability, acting as a new storage medium that can greatly reduce the latency and throughput overhead of data persistence. We present FusionFS, a file system that leverages persistent CPU caches to redesign data and metadata update approaches. FusionFS employs an adaptive data update approach that chooses the most effective mechanism based on file access patterns, minimizing PM bandwidth consumption and update latency. FusionFS also adopts an aggregated metadata update approach that consolidates small entries in persistent log buffers before appending them to PM logs, minimizing small random writes to PM. Experimental results show that FusionFS outperforms existing PM file systems in terms of latency and throughput in various scenarios.