Reconsidering "Reconsidering Custom Memory Allocation"

on YouTube ↗25m

" summary

Re-evaluates Berger et al. 2002 on modern hardware and allocators (glibc malloc, jemalloc and mimalloc) using added real-world benchmarks (Clang, Blender). Finds per-class allocators deliver only ~2% typical speedup (upper bound ~10% on microbenchmarks), while region allocators still give measurable wins up to ~14–15% (down from 44% in 2002). Introduces adversary allocation (pre-fills heap with garbage) and shows naive allocators can suffer up to ~70% slowdown on a parser while region allocators remain stable.

AI-written summary · about

" chapters (14)

1. 0:00Introduction
2. 1:542002 influential paper overview
3. 3:40Custom allocator strategies
4. 5:24Region allocators in practice
5. 7:31Hardware and allocator evolution
6. 9:14Per-class allocator performance
7. 10:37Region allocators update
8. 11:19Memory placement costs
9. 13:03Adversarial allocation method
10. 14:46Region allocators fragmentation
11. 15:48Findings and trade-offs
12. 16:49Acknowledgments
13. 17:11Q&A
14. 24:52Final thanks