RFUSE: Modernizing Userspace Filesystem Framework through Scalable Kernel-Userspace...
summary
RFUSE implements per-core ring buffers and a hybrid busy-poll to replace FUSE’s single pending queue, preserves FUSE APIs so existing filesystems relink without code changes, and runs on an 80-core PCIe Gen4 NVMe machine. Microbenchmarks show 53% lower latency for a trivial create path and IO tests report 2.27x throughput at 32 threads versus FUSE while matching ext4 on sequential workloads. Source code and evaluation scripts are available on the project GitHub.
AI-written summary · about
chapters (11)
- 0:00RFUSE introduction
- 2:46FUSE performance overhead
- 5:56FUSE scalability limits
- 6:39RFUSE design goals
- 9:51Ring channel mechanism
- 11:54Hybrid polling for wake-ups
- 12:58Async requests handling
- 14:02FUSE compatibility
- 14:24Evaluation setup
- 17:14Metadata scalability
- 18:19Conclusion
related talks
36mMUSE: MTD in Userspace, or How to Extend FUSE for Fun and Profit
03mFuseSoC in Three Minutes
44mSpeeding up Kernel Testing and Debugging with virt-meng
22mBridging the Hardware Gap With Code Harnesses on the Hugging Face Kernels Hub
27mInsights on the Re-Architecting to a Single Codebase, Running on Different Hardware...
36mA Harbor-side Chat on CephFS
51mRage Against the Ecosystem
16mMemory Allocation Profiling upcoming features