EMD: Fair and Efficient Dynamic Memory De-bloating of Transparent Huge Pages
summary
EMD is a kernel-level system for dynamic de-bloating of transparent huge pages that breaks huge pages into base pages, deduplicates zero-filled bases, and prioritizes regions using a bloat map of buckets so only the most wasteful regions are split first. Evaluation shows up to 69% faster performance versus HawkEye, breaks 6–41% fewer huge pages depending on workload, achieves balanced reclamation across instances (e.g., 27/27/24/23% at 40% debloat) with just 0.2% scanning overhead and no hardware changes.
AI-written summary · about
chapters (13)
- 0:00EMD introduction
- 0:21Virtual memory overview
- 3:32Memory bloat with THP
- 6:01Existing systems handling bloat
- 7:28Hawkeye’s unfairness
- 9:57EMD design
- 12:51EMD multi-tenant strategy
- 13:12Evaluation setup
- 14:17Performance impact
- 15:40Hawkeye vs EMD comparison
- 16:43Fairness in multi-tenant
- 17:46EMD key takeaways
- 18:10Q&A
related talks
30mMemory tiering
43mHybrid THP Mechanism. Selective Use of Huge Pages by Hot Applications
47mLinux's Leap Beyond Fixed Pages
14mDecoupling Large Folios from Transparent Huge Pages
48mPersistent Memory Usage within Linux Environment
11mAdvanced Parallel Memory Virtualization
46mHighmem deprecation planning
38mMemory Page-pooling with Auto Refill Feature for Improved Memory Allocati...