Secrets of the Shenandoah Garbage Collector
summary
Shenandoah uses concurrent marking with per-worker task queues and a work-stealing policy that in JDK 11 picks two random queues, causing up to a 25% guaranteed fail rate when many queues are inactive. Added fine-grained metrics (slow pop attempts, per-target steal breakdown) and a biased steal change that prefers the last-successful victim, raising steal success from ~66–75% to ~90% (3.5M attempts on SPECjvm, some GCs 99%+). Observes queue/worker duality and suggests restricting steals to active queues to push success further.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:49Shenandoah GC overview
- 5:31Goal
- 5:43How does work-stealing work?
- 9:03What else do we need?
- 10:14What we might expect...
- 12:33and what we got
- 13:18What's going wrong/on in a GC?
- 15:35How many workers should we use?
- 17:35Work-stealing in JDK 11
- 21:01Biases...not always bad?
- 23:39Can we do even better?
related talks
56mShenandoah: The Garbage Collector That Could
49mConcurrent Garbage Collectors: ZGC & Shenandoah
47mConcurrent Garbage Collectors: ZGC & Shenandoah
01h 01mOpenJDK: In the new age of Concurrent Garbage Collectors
01h 07mShenandoah – ultra-low Pause Time Garbage Collector #underthehood
44mGarbage Collector: How to Tame the Beast and Make It Your Best Ally
29mShenandoah 2 0: Now That We’ve Gotten the GC Pause Times Under Control, What’s Next?
52mGarbage collection for the humans