HTMFS: Strong Consistency Comes for Free with Hardware Transactional Memory...
summary
Claims that hardware transactional memory plus a small runtime gives strong per-request crash atomicity while staying fast. The system implements HOP: split file-system ops, wrap only visible writes in HTM transactions, use sequence counts, an atomic allocator with per-thread free lists, per-bucket seqs for directories and a path check for rename, and cooperative locks as fallbacks to handle HTM capacity limits. It builds HTMFS in user space on XFS internals and evaluates with FXMark, TPC‑C and LevelDB showing up to 8.4x speedups versus state-of-the-art PM file systems.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:23Performance vs Consistency
- 1:41Strong consistency
- 3:03Persistent Memory
- 3:49Hardware Transactional Memory
- 5:33Crash Consistency Mechanism Compari
- 5:53Challenges
- 7:13HOP Hardware-assisted Optimistic Persisten
- 8:35Use HOP to build HTMFS
- 9:29Data Accesses
- 9:58Atomic Memory Allocator
- 11:09Improve Dentry Scalability
- 11:31Prevent Rename Cycle
- 12:23Outline
- 12:36Evaluation: FxMark Data Read/Writ
- 13:21Evaluation: FxMark Rename
- 13:45Evaluation: Real-world Applications
- 14:23Conclusion
related talks
01h 01mUnderstanding Hardware Transactional Memory
01h 33mWhy You Should Use Software Transactional Memory
12mSPHT: Scalable Persistent Hardware Transactions
01h 00mWhat did C++ do for Transactional Memory?
01h 03m“Transactional Memory in Practice"
14mArchTM: Architecture-Aware, High Performance Transaction for Persistent Memory
01h 13mTransactional memory
55m “High Performance C++ Concurrent Transactional Data Structures"