Bridging the Memory Hotness Gap in Edge Systems with Hotness-Segregated Object Allocation
summary
HotMalloc reallocates objects by hotness instead of size, using profile-guided optimization to learn access counts offline and an offset-encoded call-site context to pick a hotness sub-heap at allocation time with zero per-access overhead. It exposes sub-heap hotness to the kernel and targets swapping and deduplication; evaluations report page-fault reductions up to 90%, preserved huge-page stability ~93%, throughput boosts up to 7.6× in microbenchmarks, deduplication gains up to 40.6×, and overall kernel improvements of 4.6%–42.1%.
AI-written summary · about
chapters (14)
- 0:00Hotness-Segregated Object Allocation intro
- 1:06Memory hotness challenges
- 4:01Root cause of page hotness skew
- 5:27HotMelo allocator design
- 6:52Offline profiling and classification
- 9:23Identifying object hotness
- 10:12HotMelo memory strategies
- 11:38Kernel interface for hotness
- 12:20Evaluation setup and benchmarks
- 13:01Reducing page hotness skew
- 14:06Memory swapping improvements
- 15:53Memory duplication with code-first
- 16:59Conclusion and key takeaways
- 18:05Q&A
related talks
20mGray-in-Young: A Generational Garbage Collection for Processing-in-Memory
14mFreedom from memory interference
38mDynamic Data Structures and Memory Management on GPUs
22mvNV-Heap: An Ownership-Based Virtually Non-volatile Heap for Embedded Systems
13mHW/SW design recommendations for 16kb devices
13mEnabling DDR segments on demand during memory pressure for DDR power reduction
01h 03mEmbedded: Customizing Dynamic Memory Management in C++
15mA Managed Memory System for micro controllers with NOR flash memory