Page-level and Fleet-wide Data Access Monitoring for Meta
summary
DAMON now exposes page-level monitoring (filters for huge pages, anon/file-backed/active lists) and a kernel agent damon_stat so a node can export idle time percentiles for fleet aggregation; page insight answers whether a 2 MB huge page backs an 816 KiB hot region. An intervals autotuning feedback loop chooses sampling (example: tuned ~500 ms sample / 8 s aggregate) and converges; real deployments collect thousands of machines, show usable patterns and run under <0.1% single-CPU overhead.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 7:19Page-level monitoring at Meta
- 13:44Page-level monitoring challenges
- 14:05Fleet-wide monitoring motivation
- 20:36Autotuning monitoring intervals
- 22:26Aggregating fleet data formats
- 26:26Introducing diamond_stat
- 29:00Deployment and usage
- 31:50Q&A insights and discussion
related talks
45mSelf-Driving DAMON/S: Controlled and Automated Access-aware Efficient Systems
37mPerformance Diagnostics with Open Distro for Elasticsearch
58mBetter MySQL With Percona PMM
30mWhy we recommend PMM to our clients
01h 03mMonitoring Performance on Massive Games.
52mNot on Our Watch: an Introduction to Application Performance Monitoring
40mMastering small chunks of data in high frequency ad-tech systems
38mAdvanced PMM Usage