FlatFS: Flatten Hierarchical File System Namespace on Non-volatile Memories
summary
FlatFS flattens a POSIX hierarchical namespace into a single indexed namespace and implements a Linux pmfs-based prototype that actually works on Intel Optane DCPMM. It replaces component-by-component lookup with a coordinated whole-path lookup that cuts path-resolve cost by up to 28×, uses a range-optimized br3 tree for fast directory-range operations (reported ~20× lower latency), and a write-optimized compressed index key layout to reduce small random writes. Benchmarks versus ext4/xfs/btrfs and others show ~4× improvement in a real app energy/throughput and 18–37% latency reductions.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:45Non-Volatile Memory
- 1:22Hierarchical File System Namespace
- 2:01#1: Slow, Non-scalable Path Walk
- 3:07#2: Expensive Namespace Tree Traversal
- 4:14FlatFS Overview and Design Goals
- 5:32FlatFS System Architecture
- 6:25Coordinated Path Walk
- 10:42Range-optimized Index Tree
- 12:43Write-optimized Compressed Key
- 14:19Namespace Crash Consistency
- 15:19FlatFS Implementation and Evaluation
- 15:42Path Walk Evaluation
- 16:58Directory Range Operation Evaluation
- 17:29Filebench Benchmark
- 17:51Git Application
related talks
42mProduction FS: Adapt or Die
42mCan File Systems Survive in Data-centric World?
14mFalconFS: Distributed File System for Large-Scale Deep Learning Pipeline
33mThe Evolution and Innovation of CubeFS
16mexF2FS: Transaction Support in Log-Structured Filesystem
29mCubeFS in Action: Empowering Users Through Case Studies
33mSSDFS: Flash-friendly File System with Highly Minimized GC Activity, Diff-on-write, & Deduplication
30mHybrid XFS: Idea to Exabytes in 12 Months