SingularFS: A Billion-Scale Distributed File System Using a Single Metadata Server
summary
SingularFS builds a billion-scale distributed file system that uses a single metadata server and persistent memory with RDMA, replacing journaling with a KV store, 16-byte atomic CAS, hierarchical concurrency control and a hybrid inode partition. Evaluation on Optane PM vs Nova, ext4 and distributed systems shows roughly an order of magnitude higher metadata throughput, about 2× lower create/delete latency than distributed systems, and steady throughput at billion-scale, so the design works end-to-end.
AI-written summary · about
chapters (8)
- 0:00SingularFS introduction
- 1:27Single server performance
- 3:53Architecture and optimizations
- 4:37Log-free metadata
- 8:44Concurrency control protocol
- 12:37Inode partitioning
- 14:01Performance benchmarks
- 16:03Conclusion
related talks
15mInfiniFS: An Efficient Metadata Service for Large-Scale Distributed Filesystems
18mEthane: An Asymmetric File System for Disaggregated Persistent Memory
12mFacebook's Tectonic Filesystem: Efficiency from Exascale
21mNew Techniques to Improve Small I/O Workloads in Distributed File Systems
22mEveryone Loves File: File Storage Service (FSS) in Oracle Cloud Infrastructure
20mscoutfs: Large Scale POSIX Archiving
50mPerformance comparison of distributed file systems
34mModernizing Ceph Deployments at CERN: CephFS and Object Storage Across Data Centres