Internet of Things and Neural Network Based Energy Optimization and Predictive Maintenance Techniques in Heterogeneous Data Centers

Abstract

Rapid growth of cloud-based systems is accelerating growth of data centers. Private and public cloud service providers are increasingly deploying data centers all around the world. The need for edge locations by cloud computing providers has created large demand for leasing space and power from midsize data centers in smaller cities. Midsize data centers are typically modular and heterogeneous demanding 100% availability along with high service level agreements. Data centers are recognized as an increasingly troublesome percentage of electricity consumption. Growing energy costs and environmental responsibility have placed the data center industry, particularly midsize data centers under increasing pressure to improve its operational efficiency. The power consumption is mainly due to servers and networking devices on computing side and cooling systems on the facility side. The facility side systems have complex interactions with each other. The static control logic and high number of configuration and nonlinear interdependency create challenges in understanding and optimizing energy efficiency. Doing analytical or experimental approach to determine optimum configuration is very challenging however, a learning based approach has proven to be effective for optimizing complex operations. Machine learning methodologies have proven to be effective for optimizing complex systems. In this thesis, we utilize a learning engine that learns from operationally collected data to accurately predict Power Usage Effectiveness (PUE) and creation of intelligent method to validate and test results. We explore new techniques on how to design and implement Internet of Things (IoT) platform to collect, store and analyze data. First, we study using machine learning framework to predictively detect issues in facility side systems in a modular midsize data center. We propose ways to recognize gaps between optimal values and operational values to identify potential issues. Second, we study using machine learning techniques to optimize power usage in facility side systems in a modular midsize data center. We have experimented with neural network controllers to further optimize the data suite cooling system energy consumption in real time. We designed, implemented, and deployed an Internet of Things framework to collect relevant information from facility side infrastructure. We designed flexible configuration controllers to connect all facility side infrastructure within data center ecosystem. We addressed resiliency by creating reductant controls network and mission critical alerting via edge device. The data collected was also used to enhance service processes that improved operational service level metrics. We observed high impact on service metrics with faster response time (increased 77%) and first time resolution went up by 32%. Further, our experimental results show that we can predictively identify issues in the cooling systems. And, the anomalies in the systems can be identified 30 days to 60 days ahead. We also see the potential to optimize power usage efficiency in the range of 3% to 6%. In the future, more samples of issues and corrective actions can be analyzed to create practical implementation of neural network based controller for real-time optimization.