Overview of Linux Page Migration and Accelerating It With Multi-threading and DMA Of...
summary
Overview of Linux page migration explains that folio copying consumes up to 96.7% of migration time for large pages and that decoupling metadata and content into batch folio copy enables parallelism. Multi-threaded CPU copying reaches about 3.5x throughput with 8 threads for 2 MB folios, while DMA offload (round‑robin across channels, fallback to CPU on error) reaches 5.16x for 2 MB pages; 4 KB pages show no benefit due to metadata overheads. Remaining issues: scheduling across CCD/CCX, accounting, DMA setup cost and async offload.
AI-written summary · about
chapters (11)
- 0:00Page migration in Linux
- 1:03NUMA systems and memory
- 2:08Why page migration matters
- 5:20How Linux migrates pages
- 7:28Limitations of migration
- 9:59Batching and parallelization
- 11:30Multi-threaded CPU migration
- 12:12DMA for page migration
- 13:592MB pages performance
- 15:05Open questions and future
- 16:28Q&A
related talks
17mDSA-2LM: A CPU-Free Tiered Memory Architecture with Intel DSA
44mLarge Pages in Linux
27mLive Migration With Hardware Acceleration
45mLinux Storage Stack Explained
24mOptimizing Storage Performance for 4–5 Million IOPs
12mEvery Mapping Counts in Large Amounts: Folio Accounting
45mData Placement in Linux: Evolving Block and File I/O
42mUnderstanding the Linux Block Layer-From Single Queue to Multi-Queue Architecture