Much Ado About Offset Vectors

on YouTube ↗24m

" summary

Offset vector compaction gets a concrete implementation and evaluation: a family of OVCs is implemented in MMTk and OpenJDK 21 and benchmarked with DaCapo on a 16‑core, 32‑thread Zen4 machine. The paper introduces a branch-free algorithm for computing forwarding info, compares offset vector compaction variants (CLmall, CCL, vanilla) and a one-pass compactor, and finds branch-free variants win, CCL slightly outperforms CLmall, OVCs are competitive with semi-space/Immix and run the tightest heaps, and one-pass fails to speed up parallel runs due to synchronization overheads; artifacts are open source.

AI-written summary · about

" chapters (14)

1. 0:00Offset vector compaction intro
2. 1:29Lisp 2 compaction
3. 3:16Limitations of Lisp 2
4. 4:21Offset vector compaction
5. 6:29History of offset compactors
6. 7:13First uses of offset compaction
7. 9:40Why offset vector compaction
8. 11:07Evaluating in MMTk
9. 12:30Mark bit map handling
10. 12:51Parallel performance vs canonical
11. 14:37One-pass compactor analysis
12. 17:23Why investigate offset compactors
13. 18:26Where used today
14. 19:08Q&A