Multi-Gen LRU: Current Status & Next Steps
summary
Patch set 14 is upstream-ready and expected around 6.1 with eight downstream kernels and 5.15 backports for Raspberry Pi and WSL2; mGLRU already defaults on Chromebooks and aims to be the system default once free bits are relocated into lower pointer bits to fit the three new generation bits on 32-bit. The core offers page table scanning with Bloom filters to clear access bits, detect empty PTEs and THP internal fragmentation, split or collapse THPs, and reduce CPU reclaim overhead; a BPF heatmap demo with a memcached benchmark shows per-process access maps and drastically lower CPU and better eviction choices, but more reclamation tests are needed.
AI-written summary · about
chapters (15)
- 0:00Introduction to Multi-Gen LRU
- 1:04Current status and benchmarks
- 3:18Overcoming obstacles for default
- 5:52Making Multi-Gen LRU default
- 6:38Efficient page table scanning
- 8:25Producer-consumer model for scanning
- 10:16Detecting THP internal fragmentation
- 12:33Discussion on THP fragmentation
- 16:13Additional ideas for page table scanning
- 21:14Demo of BPF heat map tool
- 23:01Overview of Multi-Gen LRU benefits
- 24:25Heat map tool demonstration
- 26:40Future plans for BPF integration
- 27:45Page access info for generation placement
- 28:05Q&A and next steps
related talks
47mLinux's Leap Beyond Fixed Pages
27mBPF LSM - Updates and What next?
47mReal-Time Linux: What Is Next?
30mdrm and cgroup memory management - two great tastes
23mDemystifying Linux NPU Subsystem: From Vision to LLM at Edge
31mEnhancing Scalability of the Vmalloc Mechanism in the Linux Kernel
21mPreview: Operating Systems
44mLarge Pages in Linux