Apppublic

abhinavtiwary/datacenter-env

sourceHugging Facemitupdated 6mo agoView on Hugging Face

1likes

App README

🌱 Sustainable Data Center RL Environment

Built for Meta x Scaler OpenEnv Hackathon 2026

An AI agent learns to operate a large-scale data center while minimizing carbon emissions and maximizing operational efficiency.

🔴 [Live Demo →](/) | 📖 [API Docs →](/docs)

🌍 Why This Matters

Data centers consume 1-2% of global electricity and produce millions of tonnes of CO₂ annually. Meta, Google, and Microsoft have all committed to carbon-neutral data center operations.

This environment trains AI agents to make the kinds of real-time decisions that could dramatically reduce that footprint — making it directly relevant to the infrastructure teams at Meta and Hugging Face who will evaluate this submission.

🧠 Environment Design

The Agent Controls:

Parameter	Options	Effect
`cooling_level`	1-5	Higher = cooler servers, more power used
`workload_distribution`	eco/balanced/performance	Affects heat and throughput
`power_source`	solar/wind/hybrid/grid	Determines carbon emissions
`defer_non_critical`	true/false	Smart task scheduling

What the Agent Observes:

python

{
  "avg_temperature": 42.3,          # °C across all racks
  "failed_racks": 0,                # racks that have shut down
  "solar_availability": 0.85,       # 0-1, depends on weather+time
  "wind_availability": 0.62,
  "carbon_emissions_kg": 0.0,       # this step's CO2
  "total_carbon_kg": 12.4,          # episode total
  "pue": 1.28,                      # Power Usage Effectiveness
  "time_of_day": "morning",
  "weather": "sunny",
  "incoming_workload": 0.72,
  "sla_violations": 0
}

Reward Function:

Event	Reward
Temp in safe zone (<45°C)	+2.0
Using solar/wind power	+3.0
Good PUE (<1.3)	+2.0
Task completed	+0.03 each
Smart deferral (afternoon peak)	+0.8
Rack overheating (>75°C)	-3.0
Rack failure (3+ critical steps)	-8.0
Using grid when renewables available	-2.0
SLA violation	-1.5 each
Episode carbon bonus	up to +30.0

⚙️ Configuration

Difficulty Levels

Level	Workload Variance	Outside Temp Max	Weather Changes
Easy	±10%	25°C	Rare
Medium	±20%	35°C	Occasional
Hard	±30%	42°C	Frequent

Time of Day Workload

Time	Base Load
Morning	55%
Afternoon	85% (peak)
Evening	65%
Night	25%

🚀 Quick Start

Python Client

python

import asyncio
from client import DataCenterEnv
from models import DataCenterAction

async def main():
    async with DataCenterEnv(
        base_url="https://abhinavtiwary-datacenter-env.hf.space"
    ) as env:
        obs = await env.reset(difficulty="medium")

        while not obs.observation.done:
            o = obs.observation

            # Smart agent strategy
            cooling = 5 if o.avg_temperature > 60 else 3
            power   = (
                "solar" if o.solar_availability > 0.5 else
                "wind"  if o.wind_availability  > 0.5 else
                "hybrid"
            )

            obs = await env.step(DataCenterAction(
                cooling_level=cooling,
                workload_distribution="balanced",
                power_source=power,
                defer_non_critical=(o.time_of_day == "afternoon")
            ))

            print(f"Reward: {obs.reward:.2f} | Carbon: {o.total_carbon_kg:.1f}kg")

asyncio.run(main())

REST API

bash

# Start episode
curl -X POST /reset -H "Content-Type: application/json" \
  -d '{"difficulty": "hard"}'

# Take action
curl -X POST /step -H "Content-Type: application/json" \
  -d '{"cooling_level": 3, "workload_distribution": "balanced",
       "power_source": "solar", "defer_non_critical": false}'

# Get grade
curl /grade

📡 API Reference

Method	Endpoint	Description
GET	`/`	Info
GET	`/health`	Health check
POST	`/reset`	Start episode
POST	`/step`	Take action
GET	`/state`	Full state
GET	`/grade`	Performance grade
GET	`/dashboard`	Visual interface
GET	`/docs`	API documentation

📊 Baseline Performance Scores

Evaluated over 10 episodes with seed=42.

Task	Agent	Score	Carbon (kg)	Failed Racks	Grade	Pass
Easy	Random Agent	0.28	298.4	4	F	❌
Easy	Smart AI Agent	0.87	38.2	0	A	✅
Medium	Random Agent	0.24	310.7	7	F	❌
Medium	Smart AI Agent	0.79	61.4	0	B	✅
Hard	Random Agent	0.17	338.9	11	F	❌
Hard	Smart AI Agent	0.65	94.1	2	C	✅

How to reproduce

bash

# Easy
curl -X POST /reset -d '{"difficulty":"easy","seed":42}' && python inference.py

# Medium  
CLUSTER_TASK=medium python inference.py

# Hard
CLUSTER_TASK=hard python inference.py

🧪 Baseline Results

EASY difficulty → Score: 0.93 | Grade: A MEDIUM difficulty → Score: 0.85 | Grade: A HARD difficulty → Score: 0.71 | Grade: B

🏗️ Project Structure

datacenterenv/ ├── datacenterenv/ │ ├── init.py # Package exports │ └── env.py # Local testing wrapper ├── server/ │ ├── app.py # FastAPI HTTP server │ └── datacenter_environment.py # Core RL environment ├── models.py # Pydantic Action/Observation/State ├── client.py # OpenEnv WebSocket client ├── inference.py # Baseline inference script ├── dashboard.html # Interactive visual dashboard ├── openenv.yaml # OpenEnv spec manifest ├── Dockerfile # Container config └── README.md # This file ---

Built with ❤️ for the Meta x Scaler OpenEnv Hackathon 2026