Skip to Main content Skip to Navigation

Datacenter management for on-site intermittent and uncertain renewable energy sources

Abstract : In recent years, information and communication technologies (ICT) became a major energy consumer, with the associated harmful ecological consequences. Indeed, the emergence of Cloud computing and massive Internet companies increased the importance and number of datacenters around the world. In order to mitigate economical and ecological cost, powering datacenters with renewable energy sources (RES) began to appear as a sustainable solution. Some of the commonly used RES, such as solar and wind energies, directly depends on weather conditions. Hence they are both intermittent and partly uncertain. Batteries or other energy storage devices (ESD) are often considered to relieve these issues, but they result in additional energy losses and are too costly to be used alone without more integration. The power consumption of a datacenter is closely tied to the computing resource usage, which in turn depends on its workload and on the algorithms that schedule it. To use RES as efficiently as possible while preserving the quality of service of a datacenter, a coordinated management of computing resources, electrical sources and storage is required. A wide variety of datacenters exists, each with different hardware, workload and purpose. Similarly, each electrical infrastructure is modeled and managed uniquely, depending on the kind of RES used, ESD technologies and operating objectives (cost or environmental impact). Some existing works successfully address this problem by considering a specific couple of electrical and computing models. However, because of this combined diversity, the existing approaches cannot be extrapolated to other infrastructures. This thesis explores novel ways to deal with this coordination problem. A first contribution revisits batch tasks scheduling problem by introducing an abstraction of the power sources. A scheduling algorithm is proposed, taking preferences of electrical sources into account, though designed to be independent from the type of sources and from the goal of the electrical infrastructure (cost, environmental impact, or a mix of both). A second contribution addresses the joint power planning coordination problem in a totally infrastructure-agnostic way. The datacenter computing resources and workload management is considered as a black-box implementing a scheduling under variable power constraint algorithm. The same goes for the electrical sources and storage management system, which acts as a source commitment optimization algorithm. A cooperative multiobjective power planning optimization, based on a multi-objective evolutionary algorithm (MOEA), dialogues with the two black-boxes to find the best trade-offs between electrical and computing internal objectives. Finally, a third contribution focuses on RES production uncertainties in a more specific infrastructure. Based on a Markov Decision Process (MDP) formulation, the structure of the underlying decision problem is studied. For several variants of the problem, tractable methods are proposed to find optimal policies or a bounded approximation.
Document type :
Complete list of metadata

Cited literature [210 references]  Display  Hide  Download
Contributor : ABES STAR :  Contact
Submitted on : Friday, March 6, 2020 - 11:49:10 AM
Last modification on : Monday, July 4, 2022 - 9:05:31 AM
Long-term archiving on: : Sunday, June 7, 2020 - 1:52:50 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02500692, version 1


Léo Grange. Datacenter management for on-site intermittent and uncertain renewable energy sources. Systems and Control [cs.SY]. Université Paul Sabatier - Toulouse III, 2019. English. ⟨NNT : 2019TOU30176⟩. ⟨tel-02500692⟩



Record views


Files downloads