Local (arena) Memory Allocators
summary
Local arena allocators can cut runtime by up to order-of-magnitude in real benchmarks by keeping related objects contiguous and in-cache, especially in a narrow subsystem size band tied to L1 cache (the talk measures an eight‑to‑tenfold degradation from diffusion and similar gains from arenas). Concrete tools tested: a monotonic (arena) allocator (fast teardown: “wink out” memory) and a multi-pool allocator (best for many small fixed sizes). Benchmarks introduce six axes (fragmentability, density, variation, utilization, locality, contention), warn about leaking non‑memory resources when using winked destruction, and note C++17 polymorphic memory resources for integration.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:43Outline
- 2:47Benefits
- 7:42Main Memory
- 12:40Allocators
- 14:59Local Arena
- 15:39Local Alligators
- 20:25Memory Allocators
- 28:10Allocation strategies
- 35:01Buffered sequential allocate
- 36:09Normal destruction
- 42:30Multipool alligators
- 45:51Program size
- 48:01Aspects of software
- 50:28Rough indications
- 52:02Far and away
- 54:38Concurrency
- 55:22Summary
- 55:49Considerations
related talks
01h 08mLocal (Arena) Allocators Part II
58mLocal (Arena) Memory Allocators Part 1
49mMost Malleable Memory Management Method in C++
01h 25mTaming Dynamic Memory - An Introduction to Custom Allocators
01h 31mAllocator-Aware (AA) Software
01h 03mMost Malleable Memory Management Method
51mUnderstanding Allocator Impact on Runtime Performance in C++
01h 13mValue Proposition: Allocator-Aware (AA) Software