TL;DR
This paper proposes a novel approach to reuse waste heat from distributed data centers by installing servers in residential homes to provide heating. They developed a thermal-aware workload scheduling system that uses both thermodynamic and machine learning models to optimally allocate tasks, ensuring homes are heated while meeting data center performance demands.
What Problem Does This Paper Address?
Data centers are massive energy consumers, with a significant portion of that energy being converted into heat that is typically wasted. This excess heat contributes to the environmental impact of data centers and represents a missed opportunity for energy recovery. While some efforts have been made to reuse data center heat for district heating, these often face challenges related to location, heat quality, and infrastructure costs. This paper tackles this by proposing a distributed data center model where servers are placed directly within residential buildings, acting as both computational resources and primary heat sources for the homes. The core challenge is to intelligently schedule computational workloads to generate the right amount of heat at the right time to maintain a comfortable room temperature, without compromising the performance of the data center tasks.
Methods
The authors propose a two-pronged approach:
- Thermal-Aware Workload Scheduling: This is framed as a constraint satisfaction problem. The goal is to determine an optimal scheduling matrix that assigns tasks to servers. This scheduling must satisfy:
- Service Level Agreements (SLAs): Ensuring tasks meet their deadlines and resource requirements.
- Heat Demand: Generating enough heat to meet the desired room temperature setpoint for the residents.
- Server Resource Constraints: Not exceeding the CPU, RAM, and HDD capacity of each server.
- Heat Demand Estimation Models: To inform the scheduling, two models are developed to estimate the heat demand:
- Thermodynamic Model: This model uses physical laws to predict how much heat a server needs to dissipate to change the room temperature from its current state to the desired setpoint. It considers factors like server power consumption, heat capacity, airflow, and room characteristics.
- Machine Learning (ML) Model: This approach uses monitored data (room temperature, server temperature, power consumption) to learn the relationship between workload execution and heat generation. They experimented with several ML algorithms, including Linear Regression, Polynomial Regression, Gradient Boosting Regressor, Random Forest Regressor, Support Vector Regression, and K-Nearest Neighbors Regression, with Gradient Boosting Regressor showing the best performance.
Datasets
The authors utilized data from an operational distributed data center, specifically from "QRad" server heaters. This data included ambient room temperature, server power consumption, and server temperature, recorded at a 10-second granularity over several months. This dataset was pre-processed through filtering, smoothing, and relevant sample extraction before being used for model training and validation.
Key Insights and Ideas
- Decentralized Heat Reuse: Placing servers directly in homes fundamentally changes the heat reuse paradigm, eliminating the need for complex district heating infrastructure and directly addressing local heating needs.
- Dual-Purpose Hardware: The concept of using IT equipment as a dual-purpose "computing heater" is a compelling way to improve energy efficiency and reduce the carbon footprint of both data centers and residential heating.
- Balancing Act: The core challenge lies in balancing the competing demands of computational performance (SLAs) and thermal comfort (heating). The proposed scheduling framework directly addresses this trade-off.
- Model Synergy: Combining thermodynamic principles with machine learning provides a robust approach. The thermodynamic model offers a physics-based understanding, while ML captures complex real-world dynamics and can adapt to specific environments.
Key Results
- Correlation Accuracy: The thermodynamic model achieved a correlation accuracy of 11.98% between server heat/power demand and actual heat demand. The best-performing ML model, Gradient Boosting Regressor, achieved a correlation accuracy of 4.74%.
- Reasonable Time to Setpoint: The proposed workload scheduling solution successfully distributed workloads to meet temperature setpoints within a reasonable timeframe.
- Accurate Power Following: The server power demand accurately followed the estimated heat demand.
- Model Performance: The Gradient Boosting Regressor was identified as the most suitable ML model for predicting heater power demand, showing less than 5% error (under 10W).
- Simulation Validation: The simulation results demonstrated that the system could effectively manage server power to meet predicted heat demand, leading to room temperature increases that matched setpoints.
Limitations and Future Work
- Model Complexity: The paper acknowledges that the thermodynamic function ($f_{\theta}$) can be complex, making direct optimization challenging. This led to the subproblem approach.
- Generalizability: While validated on an operational distributed DC, the models' performance might vary with different server types, room characteristics, and environmental conditions.
- Scalability: The paper focuses on a specific setup; scaling this to a large number of distributed servers and homes would require further investigation.
- Future Work: The authors suggest that further research could explore more complex thermodynamic models, investigate different ML architectures, and conduct real-world pilot deployments to validate the system's long-term performance and user acceptance. The paper also mentions the need to consider the impact of task migration on SLAs and the potential for optimizing workload distribution for energy efficiency beyond just heat generation.
Summary updated