Low-Latency, High-Throughput Garbage Collection
summary
Implements a new collector called alexa (xr/rxi) on JDK 11 with mntk and claims simultaneous low tail latency and high throughput by combining reference counting with remembered sets and a block-based amix heap with recycling plus optimistic copying so evacuation is minimal; concurrent marking uses SATB folded into RC barriers. Evaluation on search and server benchmarks shows ~4% higher throughput than G1, latency consistently below G1 and Shenandoah across heap sizes (1.3x–6x), and the write barrier costs only 1.6% overhead.
AI-written summary · about
chapters (9)
- 0:00Introduction to garbage collection
- 1:20Overview of existing collectors
- 2:50Performance benchmarks and goals
- 4:16Addressing immediacy and copying
- 6:46Emix heap and optimistic copying
- 8:54Design decisions for latency
- 12:21Implementation of read barriers
- 13:51Evaluation and final results
- 16:44Q&A session
related talks
54mGarbage Collection, State of the Art
01h 03mMMTk and The Case for Modular VM Development
16mIso: Request-Private Garbage Collection
50mAn insight into Go Garbage Collection
21mGarbage Collection for Higher Inductive Types
15mUsing MMTk Garbage Collection in GHC (Lightning Talk)
34mIt's dangerous to GC alone. Take this!
01h 01mJava compilers: Why Java is one of the first options for Low latency apps