Automating Performance Tuning with Machine Learning
summary
Claims an automated reinforcement learning optimizer applies configurations via Kubernetes APIs runs Locust load tests and collects Prometheus and Istio metrics to score results and iterate. On Google Online Boutique it finds a configuration that improves cost‑efficiency by 77% cost-efficiency in ~24 hours while meeting SLOs (90th latency ≤500 ms, errors <2%). It also shows a Kubernetes + JVM use case: JVM memory +250%, switch G1→Parallel GC, fewer GC threads, yielding +28% throughput (74→95 TPS), +19% overall throughput and −60% latency peak.
AI-written summary · about
chapters (19)
- 0:00Intro
- 1:19SREs care about efficiency and performan
- 2:48Tuning system configuration matters...
- 3:57but it is getting harder and harder
- 5:23Key requirements for a new approach
- 6:56ML techniques for smart exploration
- 9:29ML enables automated performance tuning
- 11:13and a new performance tuning process
- 11:52The target system: Online Boutique
- 12:14Use Case: optimizing cost of K8s microservices while ensuring reliability
- 13:07The reference architecture
- 13:44The optimization goals & constraints
- 15:10Best configuration found by ML in 24H improves cost efficiency by 77%
- 16:21Best config: optimal resources assigned to microservices
- 17:21Best config: higher performance & efficiency for the overall service Baseline vs Best Service throughout Baseline vs Best Service po response time
- 17:48Use Case: maximizing service performance & efficiency with JVM tuning
- 19:06Best config: +28% throughput, and meeting SLOS
- 20:03Best config: optimal JVM options 8
- 21:03Key takeaways
related talks
18mLet the machines optimize the machines
27mLeverage ML to automatically tune K8s apps
14mPerformance Optimization in High-Volume Transaction Systems
41mAutomated Performance Tuning with Bayesian Optimization
01h 01mContinuous Optimization of MIcroservices using Machine Learning
22mHow to make Dev and Ops folks look good as an SRE
25mIntelligent Kubernetes Workload Optimization
56mPerformance tuning Twitter services with Graal and Machine Learning