DAMON-based Pages Migration for {C,G,X}PU [un]attached NUMA nodes
summary
DAMON monitors page recency, frequency and stability and DAMOS automates promotion/demotion; SK Hynix reports up to 94% Alama CP CPP benchmark speed up for a CXL tiering proof, a TPP DAMON implementation reports about 42% TowBench improvement, and Micron shows up to 25% gains. The implementation autotunes hot/cold thresholds, limits demotion to ~200 MB per second and aims for 0.5% headroom on upper tiers, uses 1,000-region default (~64 KiB) overhead, and adds a damon_record_access API to attribute accesses by CPU/GPU/CXL.
AI-written summary · about
chapters (11)
- 0:00Introduction to NUMA and DAMON
- 1:26DAMON for memory tiering
- 3:37TPP implementation details
- 5:49Extending to holistic NUMA
- 8:00DAMON ops layer and accessor tracking
- 9:50DAMON record access API
- 11:37Challenges and future work
- 13:04Memory overhead comparison
- 14:07DAMON vs PG hot tracking
- 17:15Collaboration with PG hot
- 18:21Q&A and closing