Apppublic

sassy78420/openenv-datacenter-optimisation

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

Data Center Energy Optimization — OpenEnv

An AI agent optimises a data center's energy consumption by tuning server power states, cooling settings, and workload placement — while never violating thermal limits or SLA performance targets.

Why This Environment?

Data centers consume ~1–2 % of global electricity. Operators constantly balance three competing objectives:

Objective	Metric
Reduce energy	Total kW, PUE (Power Usage Effectiveness)
Maintain performance	Workload latency ≤ SLA, no dropped jobs
Stay thermally safe	Inlet temperature ≤ 40 °C per rack

This environment models those trade-offs with a simplified but physically grounded simulation so RL / LLM agents can learn to make the same decisions a human operator would.

Observation Space

Each observation is a CallToolObservation (OpenEnv spec) containing:

Field	Description
`result`	JSON output from the last tool call
`metadata.current_score`	Normalised score `[0.0, 1.0]` reflecting progress
`metadata.initial_metrics`	PUE, power, violations at episode start

Action Space (MCP Tools)

Tool	Arguments	Description
`view_datacenter()`	—	Overview: total power, PUE, rack summaries, CRAC status
`view_rack(rack_id)`	`int`	Detailed server list for one rack
`view_workloads()`	—	All workloads with SLA and priority
`set_server_power_state(server_id, state)`	`int`, `active\	idle\	sleep\	off`	Change server power mode
`adjust_cooling(crac_id, setpoint_temp, fan_speed_pct)`	`int`, `float`, `float`	Tune CRAC unit
`migrate_workload(workload_id, target_server_id)`	`int`, `int`	Move a workload to another server
`set_power_cap(rack_id, cap_watts)`	`int`, `float`	Set rack power budget
`submit_config()`	—	Finalise and score the episode

Tasks & Difficulty

Task	Scale	Key Challenges	Grader Weights
easy	2 racks · 8 servers · 1 CRAC · 6 workloads	Idle servers wasting power, fan speed too high	Energy 60 %, Thermal 20 %, SLA 20 %
medium	4 racks · 16 servers · 2 CRACs · 12 workloads	Hotspot rack, workload imbalance, overcooling	Energy 40 %, Thermal 20 %, SLA 20 %, PUE 20 %
hard	6 racks · 24 servers · 3 CRACs · 18 workloads	Cascading thermals, tight power caps, SLA-critical migrations	Energy 30 %, Thermal 15 %, SLA 20 %, PUE 15 %, Balance 10 %, Power-cap 10 %

Baseline Scores (Mock Agent — no LLM)

Task	Score
easy	0.00
medium	0.00
hard	0.00

Mock agent submits immediately with no changes, so normalised reward from baseline is 0.0.

Reward Function

—Per-step delta reward — each step returns new_score - old_score, giving the agent immediate signal for every action.
—Partial credit across multiple dimensions (energy, thermal, SLA, PUE, balance, power-cap).
—Penalty via normalisation — the score is normalised against the initial (unoptimised) state, so doing nothing yields 0.
—Step cap at 20 to prevent infinite loops.

Setup & Usage

Install

bash

pip install -e .

Run Inference

bash

export HF_TOKEN="your_token"
export API_BASE_URL="https://router.huggingface.co/v1"
export MODEL_NAME="Qwen/Qwen3.5-35B-A3B"
python inference.py

Validate OpenEnv Spec

bash

openenv validate

Docker

bash

cd server
docker build -t dc-energy-env .
docker run -p 8000:8000 dc-energy-env

Environment Variables

Variable	Required	Default	Description
`HF_TOKEN`	Yes (for real LLM)	—	Hugging Face API key
`API_BASE_URL`	No	`https://router.huggingface.co/v1`	LLM endpoint
`MODEL_NAME`	No	`Qwen/Qwen2.5-72B-Instruct`	Model identifier
`DC_TASK`	No	`easy`	Task override for the server

Project Structure

├── datacenter_generator.py   # Scenario generation (baseline + golden)
├── graders.py                # Multi-dimensional scoring per task
├── inference.py              # Baseline LLM inference script
├── openenv.yaml              # OpenEnv metadata
├── pyproject.toml            # Python project configuration
├── client.py                 # MCPToolClient wrapper
├── __init__.py               # Package exports
├── server/
│   ├── datacenter_environment.py  # MCPEnvironment implementation
│   ├── app.py                     # FastAPI app factory
│   └── Dockerfile                 # Container build
└── README.md

License

BSD-style license. See LICENSE file.