Core Concepts
DCRL-Green, a flexible and configurable reinforcement learning environment, enables the design and optimization of data centers to significantly reduce their carbon footprint.
Abstract
The paper presents DCRL-Green, an OpenAI Gym-based framework for reinforcement learning in data centers (DCs). DCRL-Green offers customizable DC configurations and targets various sustainability goals, empowering ML researchers to mitigate the climate change effects of rising DC workloads.
Key highlights:
Simulation Framework: DCRL-Green models multi-zone IT rooms, allowing users to provide custom IT cabinet geometry, server power specifications, and an HVAC system comprising chillers, pumps, and cooling towers. It also includes models for grid carbon intensity-aware load shifting and battery supply.
Configurability: Users can extensively tailor data center designs, adjusting elements from server specifics to HVAC details via a JSON object file. Parameters such as workload profiles and weather data can also be modified, enabling rapid prototyping of different data center designs.
Interface: DCRL-Green provides interfaces for applying reinforcement learning-based control, using the scalable RLLib to facilitate single as well as holistic multi-agent optimization of data center carbon footprint.
The authors demonstrate the effectiveness of DCRL-Green by optimizing HVAC setpoint control using classical single-agent reinforcement learning, leading to a 7% carbon emission reduction, and multi-agent reinforcement learning, leading to a 13% carbon emission reduction.
As future work, the authors suggest incorporating CFD neural surrogates to automate parameter generation for custom data center configurations. By lowering energy usage and shifting consumption to periods with more available green energy, DCRL-Green can significantly impact the reduction of data centers' carbon footprint and contribute to the fight against climate change.
Stats
The paper reports the following key metrics:
Using single-agent reinforcement learning (PPO and A2C) for HVAC setpoint control, the carbon footprint was reduced by 7-13% compared to the ASHRAE Guideline 36 control.
Using a multi-agent reinforcement learning approach (MADDPG), the carbon footprint was reduced by 13% compared to the ASHRAE Guideline 36 control.
Quotes
"DCRL-Green fosters sustainable data center operations, promoting collaborative green computing research within the ML community."
"Lowering the energy usage of data centers and shifting the energy consumption to periods when more green energy is available on the power grid can significantly impact the reduction of the carbon footprint of the data centers and help fight climate change."