Beyond RSS: Towards Intelligent Dynamic Memory Management (Work in Progress)
summary
This work models allocator–workload interaction as 2D rectangle packing (virtual address vs time), implements a trace-based simulation that records only virtual-memory events, and defines a fragmentation metric as colored-area ratio (external and internal). Experiments across seven modern allocators and single- and multi-threaded workloads compute peak RSS and run two‑tailed hypothesis tests (≥99% confidence), finding a monotonic relationship with the metric in 94% of cases, two workloads show no correlation and a few show surprising negative correlation.
AI-written summary · about
chapters (12)
- 0:00Dynamic memory management intro
- 1:11Dynamic allocation basics
- 2:58malloc challenges
- 3:40Fragmentation problem
- 5:48Fragmentation research question
- 6:52History of fragmentation
- 8:16Trace-based fragmentation methods
- 9:22Workload sensitivity to fragmentation
- 10:04Rectangle packing model
- 11:06Fragmentation experimental results
- 12:09Implications and future work
- 12:31Conclusion & Q&A
related talks
39mMemory Efficient Java
51mUnderstanding Allocator Impact on Runtime Performance in C++
58mLocal (Arena) Memory Allocators Part 1
01h 41mLocal (arena) Memory Allocators
01h 08mLocal (Arena) Allocators Part II
49mMaking Allocators Work, Part I
26mAllocating Less: Really Thin Rust Cloud Apps
27m“From security to performance to GPU programming..."