Revisiting Partial Tracing for Safe, Efficient, and Concurrent Garbage Collection in(…)
summary
Implements a library-quality concurrent GC for unmanaged languages that keeps precise reference-counted roots and traces the heap, using phase consensus and a hazard-pointer phase barrier to avoid stop-the-world pauses and most atomic RC updates. Microbenchmarks on lock-free maps show P99.99 tail latency falls ~6× versus reference counting and ~32× versus conservative tracing while throughput matches manual RCU; memory use rises modestly but stays below other automatic collectors. Replacing Moka’s epoch reclamation removes ~120 lines with negligible throughput (±0.25%) and P99 (<1%) change.
AI-written summary · about
chapters (7)
- 0:00Partial tracing GC intro
- 3:10Manual memory limits
- 6:23Auto GC for unmanaged langs
- 8:10Concurrent deferred design
- 14:36Performance benchmarks
- 16:48Rust cache library use
- 17:30Q&A
related talks
33mDigging through the garbage
01h 03mMost Malleable Memory Management Method
20mIntroduction to Epoch-Based Memory Reclamation
49mMost Malleable Memory Management Method in C++
55mAbusing Your Memory Model for Fun and Profit
01h 39m“Leak-Freedom in C++... By Default.”
27m“From security to performance to GPU programming..."
45mFil-C: memory safety with fanatical C/C++ compatibility