Skip to Main content Skip to Navigation

Data centers energy optimization

Abstract : To ensure both good data center service performance and reasonable power consumption, a detailed analysis of the behavior of these systems is essential for the design of efficient optimization algorithms to reduce energy consumption. This thesis fits into this context, and our main work is to design dynamic energy management systems based on stochastic models of controlled queues. The goal is to search for optimal control policies for data center management, which should meet the growing demands of reducing energy consumption and digital pollution while maintaining quality of service. We first focused on the modeling of dynamic energy management by a stochastic model for a homogeneous data center, mainly to study some structural properties of the optimal strategy, such as monotony. Afterwards, since data centers have a significant level of server heterogeneity in terms of energy consumption and service rates, we have generalized the homogeneous model to a heterogeneous model. In addition, since the data center server's wake-up and shutdown are not instantaneous and a server requires a little more time to go from sleep mode to ready-to-work mode, we have extended the model to the purpose of including this server time latency. Throughout this exact optimization, arrivals and service rates are specified with histograms that can be obtained from actual traces, empirical data, or traffic measurements. We have shown that the size of the MDP model is very large and leads to the problem of the explosion of state space and a large computation time. Thus, we have shown that optimal optimization requiring a MDP is often difficult or almost impossible to apply for large data centers. Especially if we take into account real aspects such as server heterogeneity or latency. So, we have suggested what we call the greedy-window algorithm that allows to find a sub-optimal strategy better than that produced when considering a special mechanism like threshold approaches. And more importantly, unlike the MDP approach, this algorithm does not require the complete construction of the structure that encodes all possible strategies. Thus, this algorithm gives a strategy very close to the optimal strategy with very low space-time complexities. This makes this solution practical, scalable, dynamic and can be put online.
Document type :
Complete list of metadata
Contributor : Abes Star :  Contact
Submitted on : Monday, January 25, 2021 - 4:35:09 PM
Last modification on : Tuesday, January 26, 2021 - 3:08:38 AM
Long-term archiving on: : Monday, April 26, 2021 - 7:23:10 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03120640, version 1



Léa Bayati. Data centers energy optimization. Operations Research [cs.RO]. Université Paris-Est, 2019. English. ⟨NNT : 2019PESC0063⟩. ⟨tel-03120640⟩



Record views


Files downloads