Abstract:Distributed file system (DFS), capable of efficiently managing storage resources in data centers, has become the cornerstone of data-intensive applications. To reduce the overhead of path resolution during file system operation, distributed file systems adopt client-side metadata cache, while ensuring the consistency of metadata cache by synchronizing metadata modifications. However, the state-of-the-art metadata consistency mechanisms introduce significant performance overhead and still fail to correctly synchronize all metadata modifications. To address these limitations, a lightweight metadata consistency mechanism is proposed. By employing a namespace modification method based on concurrent broadcast and a directory deletion method based on lazy broadcast and tombstones, the proposed mechanism overcomes the performance bottleneck and resolves the correctness issue. The evaluation results show that the proposed design reduces the namespace modification latency by 65.8%–66.9% and achieves throughput speedups ranging from 2.94–4.53 times. For real-world applications, the job submission latency of Spark workloads is reduced by 43.6%.