Speed is awesome, but low latency is sublime
summary
Argues that reducing VM pause times matters more than raw throughput and gives concrete fixes: implements incremental marking to split long mark phases, adds parallel and concurrent sweeping to make sweeping invisible, and uses global pretenuring plus concurrent compilation to avoid expensive copies and long compile pauses. Targets pauses under ~5 ms (ideally <1 ms), reports a doubled RMS score on the splay Octane test, but still sees outliers up to 46 ms and a 249 ms frame on a Nexus 10 Android 4.3.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:54What is latency
- 1:44High and low latency
- 2:00Why low latency
- 3:48Overview
- 4:27Garbage collectors
- 6:11Incremental marking
- 10:15Global preteen ring
- 12:22Worstcase application
- 12:47Compiler
- 13:54Concurrent compilation
- 17:02Benchmarks
- 18:15Results
- 20:45Conclusion
