Migrating Applications to the Cloud in a Cost Effective Manner
summary
Automated machine‑learning chooses instance types and JVM/OS knobs, runs 100–200 tests per instance using a Latin hypercube plus hill‑climbing and regression, and predicts cost/performance tradeoffs—examples show about a 25% cost saving by moving to a smaller instance after tuning. Empirical findings: 5–6 parameters typically each contribute ~10–15% uplift; tests run ~1 hour so billing granularity and a placement scheduler matter; occasional bug discoveries yield orders‑of‑magnitude gains, so the system both models and exploits real, sometimes counterintuitive, behavior.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:52The underlying problem
- 2:45Agile and lean
- 3:34Machine learning
- 7:59Developers
- 9:50Amazon
- 11:29The 8 Queens Problem
- 12:11Performance Testing
- 13:03No Silver Bullet
- 14:01counterintuitive results
- 16:42how we run the test
- 20:48latency graph
- 22:50time wasted
- 24:40scheduling
- 26:46random parameters
- 28:25optimization
- 31:03predictions
related talks
41mA cloud cost savings spectacular
33mThe elephant in the cloud: Strategies of cloud cost optimisation
20mAtomics Habits in cost optimizations in the cloud
24mOptimizing Cloud Costs with Machine Learning - A CloudVane Story
42mHow We Saved Cloud Budget by Using Autoscaling
01h 03mCloud Cost Optimization for Java Developers
20mHow the cloud changed the way I do analytics and research
25mMigrate Your Enterprise Applications to the Cloud