A machine learning system that predicts server hotspots and optimizes workload distribution across data center racks. Built in 4 hours for the Claude NJIT Hackathon (1st Place).
DataCool combines predictive modeling with constraint-based optimization to prevent thermal overload in data centers. The system achieves 94% prediction accuracy and reduces critical overheating incidents by 75% through intelligent workload redistribution.
Machine Learning Pipeline
- Histogram Gradient Boosting classifier for hotspot prediction
- SHAP integration for model interpretability
- Real-time probability scoring per rack
Optimization Engine
- Constraint satisfaction algorithm balancing thermal physics, migration costs, and capacity limits
- Thermal coupling simulation between adjacent racks
- Zone-based cooling efficiency modeling
Visualization Dashboard
- Interactive 3D/2D temperature heatmaps
- Before/after optimization comparison
- Detailed migration plan analysis
Before → During → After
cd DataCool
pip install -r requirements.txt
streamlit run datacenter_ai_enhanced.pyDashboard opens at http://localhost:8501
Simulates 20-30 racks with correlated features:
- CPU load, temperature, network usage, power consumption
- Thermal zones (hot/cold aisle placement)
- Spatial positioning for adjacency calculations
- Training data: synthetic racks with realistic thermal correlations
- Hotspot criteria: CPU > 80% AND temperature > 70°C
- Outputs: risk probabilities with confidence intervals
- Identify high-risk racks (sorted by prediction probability)
- Find candidate receivers (low CPU, low temperature)
- Calculate workload transfers minimizing cost × distance
- Simulate thermal redistribution with physics model
- Validate improvements against constraints
- Base: 0.5 units per % CPU transferred
- Distance penalty: 1.0 + 0.2 × Manhattan distance
- Budget tracking with efficiency metrics
Typical optimization run:
- Hotspots reduced: 4 → 1 (75% reduction)
- Max temperature: 78.3°C → 71.2°C
- Power consumption: 125.4 kW → 123.8 kW
- Migration cost: 47.3 units (within budget)
| Feature | Traditional Monitoring | DataCool |
|---|---|---|
| Prediction | Reactive | Proactive (94% accuracy) |
| Optimization | Manual | Automated constraint-based |
| Physics Model | Static thresholds | Dynamic thermal coupling |
| Visualization | Basic graphs | Interactive 3D heatmaps |
| Explainability | Black box | SHAP integration |
Adjust thermal physics in datacenter_ai_enhanced.py:
TEMP_CPU_COEFF = 0.45 # Temperature per CPU %
TEMP_ADJACENCY_COEFF = 0.08 # Thermal coupling
MIGRATION_COST_PER_PCT = 0.5 # Cost factorSwap ML models:
- XGBoost:
xgboost.XGBClassifier - LightGBM:
lightgbm.LGBMClassifier - Neural networks:
sklearn.neural_network.MLPClassifier
- scikit-learn (HistGradientBoosting)
- SHAP (explainability)
- scipy.optimize
- Streamlit (dashboard)
- Three.js, Plotly (visualization)
- pandas, numpy
DataCool/
├── datacenter_ai_enhanced.py # Main application
├── sim.py # Original prototype
├── requirements.txt
└── README.md
- Time-series forecasting with LSTM/GRU
- Real-time monitoring via WebSocket integration
- Multi-objective optimization (temperature, power, latency)
- Network topology constraints
- Container/VM-aware scheduling
SHAP computation slow: Disable in sidebar or reduce rack count
No hotspots generated: Adjust random seed or lower CPU threshold
Minimal optimization gains: Increase migration budget or check for clustered hotspots
Built for the Claude NJIT Hackathon 2025 by Yahil, Guru, Jossue, and Sergio