SenseFS: Enhancing Metadata Efficiency via Adaptive Placement in Disaggregated Persistent Memory
Zhitao Chen, Yuxiao Han · 2025
Distributed File Systems (DFSs) form the backbone of modern data centers' storage infrastructures, with metadata services critical to their performance. The recent emergence of the Disaggregated Persistent Memory (DPM) architecture, which decouples compute and persistent memory resources to enhance hardware utilization, presents an opportunity to innovate meta-data services. However, building a metadata service on the DPM architecture poses two major challenges: (1) numerous small remote reads saturate the lOPS capacity of RDMA Network Interface Cards (RNICs), and (2) frequent cache misses incur high latency penalties. Existing solutions fail to address the above challenges, making it impossible to fully utilize the DPM architecture to improve metadata service performance. This paper presents SenseFS, a metadata service tailored for the DPM architecture, It addresses these challenges with three key designs: (1) organizing metadata into fixed-size locality-preserving metadata buckets (MDBs) to enable efficient access with fewer remote reads, (2) mining the metadata correlation to optimize metadata placement, and (3) adopting a learning-based approach to allocate MDB space based on access patterns to improve space utilization. Evaluation results demonstrate that SenseFS significantly enhances metadata efficiency, achieving higher throughput and reduced latency compared to state-of-the-art solutions.