Abstract:
The rapid increase in data center size and number, driven by escalating internet and cloud
computing demands, has led to high energy consumption and public concern. Densely
packed high-powered systems within data centers generate significant radiant heat,
necessitating effective and Reliable cooling solutions for maintaining uptime. This
research presents a novel method to evaluate Computer Room Air Conditioning (CRAC)
system performance and efficiency. Firstly, a rack-level heat transfer probabilistic
constraint is introduced, integrating environmental conditions such as ambient
temperature, humidity, and airflow patterns, which significantly impact heat transfer
processes and are accurately incorporated to reflect real-world scenarios. Additionally, the
model accounts for specific configurations and thermal properties of data server racks,
enabling precise simulation of heat generation and dissipation patterns. The probabilistic
variables undergone training including the layout of servers, types of cooling mechanisms
employed, and the material properties of the racks. Secondly, modelling the CRAC
system’s heat transfer rate as random distribution facilitates effective thermal load
management and balances computational demands with accuracy. Based on the output
from two probabilistic performance functions, a multi-response Gaussian process
(AMRGP) model is developed using an adaptive sampling technique, enhancing predictive
accuracy and efficiency by training the predicted responses with a learning U-function to
calculate the probability of failure and reliability of the model. The proposed method also
improves risk assessment by predicting the likelihood of failure events, aiding in the
development of a powerful tool for designing and evaluating CRAC system reliability in
complex and uncertain environments. This research thus represents a significant
advancement in the field of data center engineering, providing a robust framework for
future development in thermal management and reliability assessment.
