A Configurable Pythonic Data Center Model for Sustainable Cooling and ML Integration

Avisek Naug, Antonio Guillen, Ricardo Luna Gutierrez, Vineet Gundecha, Sahand Ghorbanpour, Sajad Mousavi, Ashwin Ramesh Babu, Soumyendu SarkararXiv 2024 · arXiv

TL;DR

The authors introduce PyDCM, a native Python library and Gymnasium environment designed for rapid prototyping and reinforcement learning (RL) control of data center cooling systems. By eliminating the heavy communication overhead of traditional tools like EnergyPlus, PyDCM accelerates simulations by up to 40x, enabling researchers to easily train carbon-aware controllers that minimize energy consumption and mitigate thermal hotspots.

What Problem Does This Paper Address?

Data centers are notorious energy hogs, generating a massive carbon footprint due to continuous CPU workloads and the complex HVAC systems required to cool them. To make data centers more sustainable, researchers are increasingly turning to intelligent control algorithms, particularly Deep Reinforcement Learning (DRL).

However, standard building simulation tools like EnergyPlus or Modelica are not native to Python—the lingua franca of modern ML. Interfacing these tools with Python RL frameworks (like Ray or RLlib) requires complex wrapper software, introducing significant cross-process communication overhead. Furthermore, these legacy tools can be slow and rigid, making it difficult to rapidly prototype custom, heterogeneous server room layouts or quickly evaluate real-time sustainability metrics like grid carbon intensity (CI).

Methods

PyDCM addresses these bottlenecks through a native Python, Object-Oriented Design (OOD) framework:

  • Vectorized Thermal Calculations: Instead of relying on slow iterative loops, PyDCM vectorizes thermodynamic equations across all simulated CPUs and server racks, drastically speeding up compute times.
  • CFD Integration: The model incorporates precomputed Computational Fluid Dynamics (CFD) data to accurately capture spatial temperature gradients (inlet/outlet airflows) without the need to run computationally prohibitive live fluid dynamics simulations.
  • End-to-End Cooling Chain Modeling: It models the entire thermal path: from individual CPU power curves and cabinet fans to Computer Room Air Conditioning (CRAC) units, chillers, and cooling towers.
  • Gymnasium Interface: The entire framework is wrapped as a standard Farama Foundation Gymnasium environment, making it plug-and-play with standard RL libraries.
Figure 1: PyDCM Data Center Modeling Framework.

Datasets

PyDCM is a simulation framework rather than a traditional dataset-driven ML model. However, to run simulations, it ingests:

  • Data center configuration specifications (geometry, CPU power curves, fan specs).
  • Precomputed CFD thermal distribution matrices.
  • Real-world weather data and local grid carbon intensity (CI) data ($gCO_2/kWh$).

Key Insights and Ideas

  • The Cost of Language Barriers: A massive chunk of simulation time in RL research is wasted on data serialization between Python controllers and C++ simulation engines. Keeping everything native to Python yields immediate, compounding dividends.
  • EnergyPlus Simplifications: The authors observed that EnergyPlus exhibits a flat simulation time regardless of the number of CPUs. This strongly implies that EnergyPlus calculates the thermal properties of a single CPU and simply scales it linearly. While efficient, this prevents researchers from modeling realistic, heterogeneous data centers where workloads vary wildly across different racks.
  • Carbon-Aware Feedback Loops: By integrating real-time grid carbon intensity data directly into the environment, RL agents can learn to "pre-cool" data centers or shift cooling loads to times when renewable energy is abundant on the grid.

Key Results

The authors benchmarked PyDCM against an EnergyPlus data center implementation across several key metrics:

  • Sub-Millisecond Step Times: PyDCM achieved a 99.99% reduction in environment reset time (dropping from 2.67 seconds to 0.03 milliseconds) and a 71.33% reduction in step time.
  • Scalability at Hyper-Scale: For standard configurations, PyDCM operates 40x faster than EnergyPlus. When scaling up to a "Hyper-Scale Data Center" with more than 10,000 CPUs, PyDCM still maintains a 16x speedup.
Figure 2: Simulation speed up relative to the current implementation of EnergyPlus (E+). * Hyper- scale data center consists of more than 10,000 CPUs. (6)
  • Lower Resource Footprint: During RL training, PyDCM required less RAM (16.84 GB vs. 18.20 GB) and lower CPU utilization (18.21% vs. 20.64%) than EnergyPlus.
  • Proven Sustainability Gains: Using PyDCM, the authors trained DRL agents (PPO and A2C) to optimize HVAC setpoints. Compared to a standard ASHRAE Guideline 36 baseline controller, the RL agent achieved a 7.36% reduction in energy consumption and a 7.23% reduction in carbon footprint.
Figure 3: Applications of PyDCM

Limitations and Future Work

Not discussed in the paper.

Summary updated

Abstract

There have been growing discussions on estimating and subsequently reducing the operational carbon footprint of enterprise data centers. The design and intelligent control for data centers have an important impact on data center carbon footprint. In this paper, we showcase PyDCM, a Python library that enables extremely fast prototyping of data center design and applies reinforcement learning-enabled control with the purpose of evaluating key sustainability metrics including carbon footprint, energy consumption, and observing temperature hotspots. We demonstrate these capabilities of PyDCM and compare them to existing works in EnergyPlus for modeling data centers. PyDCM can also be used as a standalone Gymnasium environment for demonstrating sustainability-focused data center control.
BibTeX
@misc{naug2024configurable,
  author = {Avisek Naug and Antonio Guillen and Ricardo Luna Gutierrez and Vineet Gundecha and Sahand Ghorbanpour and Sajad Mousavi and Ashwin Ramesh Babu and Soumyendu Sarkar},
  title = {{A Configurable Pythonic Data Center Model for Sustainable Cooling and ML Integration}},
  year = {2024},
  eprint = {2404.12498},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.12498},
}

Loading related content...

Figure