The Woodruff School

SUMMARY

Thermal management of data centers remains a challenge because of their ever-increasing power densities and decreasing server footprints. Current lack of dynamic control over global provisioning and local distribution of cooling resources often result in wasteful overcooling. These trends motivate this thesis research, which focuses on the development of a reliable and energy-efficient framework for allocating cooling resources to meet thermal management requirements, while minimizing energy consumption and adverse environmental impacts. A key component of energy-efficient thermal management is real-time accurate prediction of temperature distribution in data centers. This first section of this dissertation focuses on development and comparison of four Data Driven Modeling (DDM) methods, namely Artificial Neural Networks (ANN), Support Vector Regression (SVR), Gaussian Process Regression (GPR) and Proper Orthogonal Decomposition (POD). These DDM methods were trained on datasets generated from offline Computational Fluid Dynamics/Heat Transfer (CFD/HT) simulations for real-time prediction of temperature and airflow distributions in a data center. Using CFD simulation results to train DDMs transfers computational complexity from model execution (in CFD) to model setup and development. To generate the training data, a physics-based and experimentally validated room-level CFD/HT model was developed using the commercial software Future Facilities 6Sigma Room. Another key component of the overall framework is a model to estimate the cooling power consumed by a data center. This research developed a model based on thermodynamic analyses of data center cooling equipment, as described here. Finally, development and implementation of a Genetic Algorithm (GA) based optimization framework in a data center lab is presented. The optimization framework employs an ANN based model to predict rack inlet air temperatures and a thermodynamic model to optimize cooling energy consumption. Results from a test run of 7.5 hours in the Data Center Laboratory indicate that implementing this optimization framework for dynamic provisioning of cooling resources reduces cooling power consumption by 20% compared with baseline operation without this optimization.

SUBJECT:	Ph.D. Dissertation Defense

BY:	Jayati Athavale

TIME:	Friday, December 7, 2018, 1:30 p.m.

PLACE:	MRDC Building, 4211

TITLE:	Artificial Neural Network Based Prediction and Cooling Energy Optimization of Data Centers

COMMITTEE:	Dr. Yogendra Joshi, Co-Chair (ME) Dr. Minami Yoda, Co-Chair (ME) Dr. Satish Kumar (ME) Dr. Godfried Augenbroe (DESIGN) Dr. Ada Gavrilovska (CS)

SUMMARY Thermal management of data centers remains a challenge because of their ever-increasing power densities and decreasing server footprints. Current lack of dynamic control over global provisioning and local distribution of cooling resources often result in wasteful overcooling. These trends motivate this thesis research, which focuses on the development of a reliable and energy-efficient framework for allocating cooling resources to meet thermal management requirements, while minimizing energy consumption and adverse environmental impacts. A key component of energy-efficient thermal management is real-time accurate prediction of temperature distribution in data centers. This first section of this dissertation focuses on development and comparison of four Data Driven Modeling (DDM) methods, namely Artificial Neural Networks (ANN), Support Vector Regression (SVR), Gaussian Process Regression (GPR) and Proper Orthogonal Decomposition (POD). These DDM methods were trained on datasets generated from offline Computational Fluid Dynamics/Heat Transfer (CFD/HT) simulations for real-time prediction of temperature and airflow distributions in a data center. Using CFD simulation results to train DDMs transfers computational complexity from model execution (in CFD) to model setup and development. To generate the training data, a physics-based and experimentally validated room-level CFD/HT model was developed using the commercial software Future Facilities 6Sigma Room. Another key component of the overall framework is a model to estimate the cooling power consumed by a data center. This research developed a model based on thermodynamic analyses of data center cooling equipment, as described here. Finally, development and implementation of a Genetic Algorithm (GA) based optimization framework in a data center lab is presented. The optimization framework employs an ANN based model to predict rack inlet air temperatures and a thermodynamic model to optimize cooling energy consumption. Results from a test run of 7.5 hours in the Data Center Laboratory indicate that implementing this optimization framework for dynamic provisioning of cooling resources reduces cooling power consumption by 20% compared with baseline operation without this optimization.