Self-Driving DAMON/S: Controlled and Automated Access-aware Efficient Systems
summary
Kernel subsystem DAMON provides controlled data‑access monitoring with dynamic monitoring intervals autotuning, dynamic space granularity (merge/split regions) and time granularity (per‑region age counters), plus DAMOS for automated memory actions with DAMOS quotas and filters. Defaults start at 5 ms/100 ms; autotuning lets users target ~4% of max events with safeguards (min 5 ms, max 10 s) and converges (example: ~370 ms). Prototype shows 2× speedup in an artificial test, negligible CPU overhead in production, and scale testing up to 1 TB scale tested.
AI-written summary · about
chapters (14)
- 0:00Introduction to Self-Driving DAMON/S
- 2:29What is DAMON
- 8:11Dynamic granularity in DAMON
- 12:31DAMOS automated memory ops
- 17:28DAMOS quota and autotuning
- 21:29Why monitoring intervals matter
- 27:15Autotuning monitoring intervals
- 32:36DAMONstat simplified monitoring
- 35:28DAMON resources and community
- 36:53Summary and future of DAMON
- 37:39Q&A: Memory overhead limits
- 40:22Q&A: Distinguishing metadata access
- 41:52Q&A: Reclamation and system pressure
- 44:03Q&A: Proactive reclamation benefits
related talks
34mPage-level and Fleet-wide Data Access Monitoring for Meta
46mNew Innovations in Power and Performance Management
46mTaming the 9s
36mMitigating Noisy Neighbours: Advanced Container Resource Management
41mThe Importance of Continuous Regression for HW & SW Development
23mAdobe + Aerospike: Real-time Streaming Segmentation
22mAccelerate the Image Distribution for DADI Using Intel DSA
26mProcessing one million data points daily with Dapr