Understanding and Finding Crash-Consistency Bugs in Parallel File Systems
Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14·
summary
Study finds 34 vulnerabilities across parallel file systems, with 15 and 19 in BeeGFS and OrangeFS and 5 in XD4, including two persistence-order bugs that cause data loss for an atomic replace via rename workflow. PFSCheck implements trace permutation plus pruning, multi-level IO and network tracing, crash-state emulation, legal-state replay, and bug classification, automatically generating crash states, exercising post-crash recovery tools, and scaling to large HPC clusters while claiming low overhead and systematic detection of cross-node consistency failures.
AI-written summary · about
chapters (8)
- 0:00Start
- 0:40PFS failures are frequent and expensive
- 1:27Introduction to parallel file systems
- 2:43A PFS failure example
- 3:25A study of crash vulnerabilities on PFSes
- 4:19PFS crash vulnerabilities
- 5:14Crash vulnerability example
- 11:12Conclusion
related talks
20mScalable and Accurate Application-Level Crash-Consistency Testing(…)
13mUpFuzz: Detecting Data Format Incompatibility Bugs during Distributed Storage System...
52mFantastic C++ Bugs and How to Test Them
18mPreview: Verification (Correctness & Debugging)
28mExtreme OS Kernel Testing
19mCAFault: Enhance Fault Injection Technique in Practical Distributed Systems via...
18mEfficient Exposure of Partial Failure Bugs in Distributed Systems with Inferred...
29mScaleCheck: A Single-Machine Approach for Discovering Scalability Bugs...
