xadfs: Inverted DAX
summary
A proof-of-concept filesystem called xadfs maps mmap as the default I/O path and implements read/write by going directly to NVDIMM namespaces, bypassing the usual block layers. Benchmarks use fio-style random 70/30 tests and large-block scaling on a 40-core dual-node HP system, showing xadfs is faster for sub‑128B/sub‑block accesses but worse beyond ~1KB; measured peak is ~30 GB/s and ~10M IOPS. The implementation works and exposes where NVDIMMs benefit workloads and where traditional DAX/block paths win.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:13Overview
- 1:31Introduction: NV-DIMM
- 2:28NV-DIMM namespaces
- 3:40Introduction: App-Direct mode
- 4:55NV-DIMM: Block mode
- 6:50NV-DIMM direct access
- 9:37DAX problems
- 15:07Xadfs: bandwidth comparison
- 15:58Xadfs: IOPS comparison
- 18:33Pmem block scalability
- 20:34Conclusion
- 20:57Outlook: NV-DIMM vs fast storage
related talks
43mThe Forefront of the Development for NVDIMM on Linux Kernel
25mShared CXL 3 memory what will be required
31mDesign and implementation of Autocaching for CXLSSD
38mRethinking Writeback: Scaling Linux Filesystem and Memory Performance for the Next Decade
18mOpen vSwitch with AF_XDP as a userspace datapath
29mPlumbing challenges in Dynamic capacity device
28mLow-Latency, Deterministic Networking with Standard Linux using XDP Sockets
12mRethinking File Mapping for Persistent Memory