Nani4481/datacenter-thermal-env

🌡️ Hyperscale AI Data Center Thermal & Power Controller

Solving the Physical Bottleneck of AI Progress

    

📖 Overview

In the race to train the next generation of Frontier Models — like Llama-4 or Gemini-3 — the absolute biggest physical bottleneck isn't silicon. It's power and cooling.

Massive GPU clusters consume megawatts of electricity. If GPUs overheat, they undergo thermal throttling, killing training performance — or worse, triggering hardware failure worth millions of dollars.

This project provides a high-fidelity OpenEnv environment that simulates a 2D grid of server racks. It challenges AI agents to act as Thermal & Power Controllers, balancing computational throughput with energy-efficient cooling while navigating hardware failures — a real problem at the frontier of AI infrastructure.

🏗️ System Architecture

The environment models a complex, stateful thermodynamic system with realistic physics. To ensure stability during concurrent automated evaluations, the API backend implements a robust Least Recently Used (LRU) Cache for session memory management.

┌─────────────────────────────────────────────────────────┐
│              DATA CENTER SIMULATION GRID                │
│                                                         │
│   ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐│
│   │  RACK R0 │  │  RACK R1 │  │  RACK R2 │  │  RACK R3 ││
│   │ 🔥 85°C  │──│ 🌡️ 72°C  │──│ ✅ 60°C  │──│ ✅ 58°C  ││
│   │  Load:90%│  │  Load:65%│  │  Load:40%│  │  Load:35%││
│   └────┬─────┘  └────┬─────┘  └────┬─────┘  └────┬─────┘│
│        │             │             │             │      │
│   ┌────▼─────────────▼─────────────▼─────────────▼──────┐ │
│   │          HVAC ZONE H0         HVAC ZONE H1          │ │
│   │          Output: 78%              Output: 45%       │ │
│   └─────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────┘

🛠️ OpenEnv Specification

1. Observation Space — DataCenterObservation

The agent receives a complete telemetry snapshot of the data center at every timestep, strictly typed via Pydantic:

class DataCenterObservation(BaseModel):
    step: int
    racks: List[RackState]          # Per-rack telemetry (temp, load, status)
    hvacs: List[HVACState]          # HVAC zone status (output %, operational status)
    current_pue: float              # Power Usage Effectiveness
    total_compute_kw: float
    total_cooling_kw: float
    load_imbalance: float           # Max-rack load − Min-rack load (%)
    thermal_warnings: int           # Racks exceeding 80°C

2. Action Space — DataCenterAction

The agent controls the physical layer through three primary vectors:

class DataCenterAction(BaseModel):
    hvac_adjustments: Dict[str, float]   # Map of hvac_id to new cooling_output_percent (0-100)
    workload_shifts: List[WorkloadShift] # Shift compute load: {from_rack, to_rack, amount_percent}
    throttles: Dict[str, float]          # Map of rack_id to throttle amount (reduce workload)

⚠️ throttles is a destructive action. Reducing workload saves hardware but sacrifices compute SLA. The agent must mathematically deduce when safe racks lack the capacity to absorb failing workloads and trigger throttles to survive.

3. Reward Function (Strictly Bounded)

The reward signal is shaped to incentivize efficiency while penalizing risk. Per hackathon rules, all rewards and final grades are strictly clamped within the bounds of [0.01, 0.99].

$$R = \text{Clamp}_{0.01}^{0.99} \left( f(\text{PUE}) - 0.1 \times |\text{Warnings}| - 0.5 \times |\text{Violations}| \right)$$

🎯 Tasks & Difficulty

🚀 Getting Started

Prerequisites

• —Docker installed and running
• —Python 3.12+
• —Hugging Face Token (HF_TOKEN): Required to authenticate via the Hugging Face inference router.

🐳 Local Development

Step 1 — Build the container:

docker build -t datacenter-env .

Step 2 — Run the FastAPI simulation server:

docker run -p 7860:7860 datacenter-env

The server implements an LRU Cache to safely handle up to 50 concurrent validation sessions without memory leaks.

Step 3 — Run the inference agent (in a new terminal):

export HF_TOKEN="hf_your_token_here"
python inference.py

📊 Baseline Results

Evaluated using Frontier-class reasoning models (Qwen-2.5-72B-Instruct / Llama-3.3-70B):

Note: High-reasoning models successfully utilize the throttles fallback action in the Hard task, proving the environment is legitimately solvable by LLMs reacting to state telemetry.

📁 Project Structure

datacenter-env/
├── Dockerfile               # Container definition & healthchecks
├── openenv.yaml             # Standardized OpenEnv task configuration
├── datacenter_env.py        # Core thermodynamics engine and strict Pydantic schemas
├── inference.py             # Agent entry point with structured prompt engineering
└── server/
    └── app.py               # FastAPI server with memory-safe LRU caching

📜 License

This project is licensed under the Apache License 2.0.

<div align="center">

Created for the Meta x Scaler OpenEnv 2026 Challenge 🏆

Pushing the physical limits of AI infrastructure, one thermal cycle at a time.

</div>