“Local ('Arena') Memory Allocators (part 2 of 2)”

on YouTube ↗01h 02m

" summary

Four targeted benchmarks compare the platform global allocator, tcmalloc/jemalloc, and local strategies (monotonic, multipool, composite) using real containers and node sizes; results show global allocators lose consistently, monotonic allocator often wins, and local allocators prevent massive slowdown from memory reordering: random diffusion (not fragmentation) causes 10–16× access-time degradation, local allocators cut that to about 2.5× in worst cases, and multipool overflows (e.g. >2KB) fall back to the global allocator and become slow. monotonic allocator and multipool strategies outperform for long-running and high-utilization scenarios.

AI-written summary · about

" chapters (10)

1. 0:00Introduction and roadmap
2. 4:21Short running build up tear down
3. 18:31Global allocators inferiority
4. 19:14Long running access locality
5. 39:43Diffusion and shuffling effects
6. 47:01Takeaways for long running systems
7. 50:17Utilization and multipool allocators
8. 53:59Contention and allocator choice
9. 57:31Fragmentability concept
10. 1:00:46Final conclusions and summary