Modélisation, analyse et prédiction de l’efficacité énergétique des centres de données

Abstract

The applications and technologies enabled by various concepts and paradigms of cloud computing and Internet of Things, and their combination, have led to growing volumes of data to process, store, and exchange. This in turns led to a growing need for data centers (DC). With the rapid expansion of DCs, many problems are emerging, and high energy consumption is the first. Also, trends in increasing data center capacity and energy use by IT equipment, reaffirm the need to verify and maintain the energy efficiency of DCs. As a result, DC are among the most complex and energy-intensive environments due to high internal loads, low temperature, and humidity settings, and continuous operation. DCs are complex systems composed of Information and technology (IT) and non-IT (i.e., mechanical, and electrical) sub-systems. The variety of configurations and the interdependencies of the different data center sub-systems leads to enormous challenges in understanding and optimizing energy efficiency. Designing a DC is a difficult engineering task due to the complexity of the subsystems that form it, as well as its various interactions and dependencies of different subsystems. Within this context, a metamodel DCMM (Data Center MetaModel) is proposed which presents the different sub-systems constituting the physical and IT infrastructure, their interconnections and the heterogeneous structure of DC, the main characteristics, and its various constraints. This metamodel is used as a basis to give a precise and generic redefinition of PUE. Then, we analyze the behavior of deep machine learning-based model to automatically calculate and predict the data center energy efficiency metric Power Usage Effectiveness (PUE). A sensitive analysis is used to fix the key parameters and evaluate the actions of various data center subsystems on the PUE variation. We implement three deep machine learning-based models: 1) DNN model with the "Resilient Back-Propagation: ReBP" algorithm as training algorithm, 2) Attention-based LSTM model and 3) Google MLP base line model for comparison. The comparative study is based on three metrics: mean square error (RMSE), mean absolute error (MAE) and the correlation coefficient R-squared (R2). The validation of the proposed approach is done through experimentations with real datasets from two real case studies. The obtained results of this experimentation indicate that our proposed LSTM attention-based model improves the PUE, and consequently, shows its promise for a practical implementation.

Date	11 Jun 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Abdelouahed Gherbi (Supervisor) & Nadjia Kara (Co-supervisor)