Jmvx: Fast Multi-threaded Multi-Version eXecution and Record-Replay for Managed Languages

on YouTube ↗18m

" summary

Jmvx instruments Java bytecode to intercept native and nondeterministic library calls (found via strace signals), implements a lock‑free shared mmap ring buffer between JVMs, and enforces monitor ordering with a Lamport clock monitor ordering so multi-threaded MVX and RR tolerate JVM GC/JIT divergence. Evaluation shows MVX overhead averaging 5% when multithreading support is disabled and 47% when enabled, recording costs 8% (25%) and replay 13% (73%), and it beats RR’s 4.33x/4.7x costs with 1.25x/1.73x for record/replay.

AI-written summary · about

" chapters (19)

1. 0:00JMVX introduction and goals
2. 1:04Record replay for debugging
3. 2:06From replay to multi-version execution
4. 3:09JMVX design for JVM complexities
5. 4:12Nondeterministic Java APIs
6. 4:55Instrumenting Java methods
7. 5:56Class loading and concurrency handling
8. 6:37Thread ordering with Lamport clocks
9. 7:41Bytecode for synchronized methods
10. 8:01Fast multi-threaded execution
11. 8:23Divergences in program output
12. 9:06JMVX instrumentation performance
13. 10:08Record replay vs RR
14. 10:50Chronicler multi-threaded comparison
15. 11:55JMVX poor Lucene performance
16. 12:38Ring buffer size impact
17. 12:59Scaling with threads and benchmarks
18. 13:41Key takeaways and capabilities
19. 14:04Q&A