VESTA: Power Modeling with Language Runtime Events
summary
VESTA builds power models from JVM runtime events (using USDT probes) and energy samples from jle/Intel RAPL, trains gradient‑boosted decision trees (XGBoost) on 37 real JVM benchmarks, and predicts energy with about 1.56% MAPE. The system tracks split‑phase event depth, keeps monitoring overhead low (~2–3% energy and runtime), outperforms linear HPC baselines on nonlinearity, and provides per‑feature explainability (SHAP) tying GC, allocation and concurrency events to power states.
AI-written summary · about
chapters (7)
- 0:00VESTA and power modeling
- 3:06Language runtime events for power
- 7:38VESTA power model design
- 9:01VESTA validation and accuracy
- 10:04VESTA explainability and features
- 11:53VESTA overhead and robustness
- 13:38Conclusion
related talks
49mThe Future is 💚: A Journey through the Energy Efficient ..
50mEnergy Efficient Design Principles of Software Engineering
05mProject Lightning Talk: Kepler Project Updates
39mCreate an Eco-Friendly World with Green Software Engineering
31mCreate an eco-friendly world with Green Software Engineering
20mMonitoring Energy Consumption and Building Energy Efficient Systems the Cloud Native Way
35mSustainability Research the Cloud Native Way
20mKepler - Optimized power usage monitoring for Kubernetes