Can File Systems Survive in Data-centric World?
summary
File systems introduce measurable overhead as byte‑addressable persistent memory, kernel bypass and hardware accelerators converge; the question is whether to evolve or replace them. The talk explains work on SSDFS, kernel clients and legacy HFS/HFS+ support, cites 60% data movement as a core inefficiency and contrasts human brain (20 W) with AI energy demands, and surveys alternatives: move metadata into device firmware, push logic to user space, collapse to VFS‑only on persistent memory, or adopt CXL and computation‑in‑storage. Legacy data and permissions keep some filesystem semantics necessary.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:04File system core concepts
- 4:16Big data rise and limits
- 6:45Persistent memory and copying
- 9:15File systems future survival
- 13:54File system building blocks
- 16:46Why file systems overhead
- 18:11File systems mediator role
- 21:47Future file concepts
- 23:16Active objects and data management
- 25:27Permission models for distributed systems
- 26:53Programming models near data
- 29:44Future data processing vision
- 32:56Solutions and paradigm shifts
- 34:43Closing remarks
- 35:04Q&A
related talks
56mThe evolution of storage on Linux
14mRe-think Data Management Software Design Upon the Arrival of Storage Hardware with
19mFlatFS: Flatten Hierarchical File System Namespace on Non-volatile Memories
42mProduction FS: Adapt or Die
30mHybrid XFS: Idea to Exabytes in 12 Months
31mScaling Autonomous-Driving R&D: A 100 PB, 100 Billion-File Open-Source Elastic File...
31mFear not, for the filesystem is with you
33mThe Evolution and Innovation of CubeFS