Memory tiering
summary
Memory tiering treats cold pages by clearing the access bit and rechecking after about a minute, which reliably finds long-unused pages, while hot-page detection requires sampling or PMU data that vendors currently don’t expose broadly (hardware counters are often limited to ~4,000 entries). Policy experiments show user-space policies (Google) adapt faster than kernel heuristics, DMA engines can accelerate raw moves but metadata and unmapping dominate cost, and prior hardware attempts like Intel’s two-level memory reduced software work but failed to satisfy most workloads. access bit works for cold pages hardware counters limited (~4k)
AI-written summary · about
chapters (8)
- 0:00Introduction to memory tiering
- 1:03Memory hierarchy challenges
- 3:24Explicit vs. implicit placement
- 6:26Page detection strategies
- 8:40Cold page detection mechanisms
- 12:52Hot page detection challenges
- 17:34Kernel and hardware limitations
- 24:14Page migration and policy
related talks
48mPersistent Memory Usage within Linux Environment
19mP2CACHE: Exploring Tiered Memory for In-Kernel File Systems Caching
11mAdvanced Parallel Memory Virtualization
22mEMD: Fair and Efficient Dynamic Memory De-bloating of Transparent Huge Pages
01h 11mPerformance engineering - being friendly to your hardware
15mNomad: Non-Exclusive Memory Tiering via Transactional Page Migration
45mUsing hardware hints for optimal page placement
12mRethinking File Mapping for Persistent Memory