A Green Model of Cloud Resources Provisioningpdfs.semanticscholar.org/990a/7c751cda4d75ef5876b277ecaa0ec2… · A Green Model of Cloud Resources Provisioning Meriem Azaiez 1, Walid

A Green Model of Cloud Resources Provisioning

Meriem Azaiez1, Walid Chainbi1 and Hanen Chihi21National Engineering School of Sousse, University of Sousse, Sousse, Tunisia

2Institute of Computer Sciences of Ariana, Ariana, Tunisiafazaiez.myriam, walid.chainbi, [email protected]

Keywords: Cloud Computing, Optimization, Scheduling, Green Computing, CloudSim, Genetic Algorithm.

Abstract: The evolution of network technologies and their reliability on the one hand, and the spread of virtualizationtechniques on the other hand, have motivated the use of execution and storage resources allocated by distantproviders. These resources may progress on demand. Cloud computing deals with such aspects. However,these resources are greedy in energy because they consume huge amounts of electrical energy, which affectsthe invoicing of Cloud services which depends on run-time and used resources. The environment is affectedtoo due to the emission of greenhouse gas. Therefore, we need Green Cloud computing solutions that reducethe environmental impact. To overcome this Challenge, we study in this paper the relationship between Cloudinfrastructure and energy consumption. Then, we present a genetic algorithm based solution that schedulesCloud resources and optimizes the energy consumption and CO2 emissions of Cloud computing infrastructurebased on geographical features of data centers. Unlike previous work, we propose to optimize the use of Cloudresources by scheduling dynamically the customers applications and therefore reduce energy consumptionas well as the emission of CO2. The optimal solution of scheduling is found using multi-objective geneticalgorithm. In order to test our model, we extended the CloudSim simulator with a module implementingthe dynamic scheduling of customers applications. The experiments show promising results related to theadoption of our model.

1 INTRODUCTION

Cloud computing is an emerging field which becomesincreasingly popular. But, this technology is identi-fied as one of the fastest growing consumers of en-ergy. This consumption of energy will reach in 2020,more than three times compared to today (Relaxnews,2010). This problem is caused by the energy con-sumed by data centers. Indeed, the data centers en-ergy consumption increases with the number of cen-ters and with data center workload. This consumptionis amplified when the cooling infrastructure and aux-iliary equipment are included which represents morethan 50% of the power consumption (Zhang et al.,2012).

Another problem with energy is the emission ofgreenhouse gas which reached the 2% of the totalamount of CO2 emissions in the world (TUAL, 2013)and will quadruple in 2020 (Thrash, 2012). This prob-lem brings a huge impact to the environment. There-fore, a new challenge appears to deal with the energyconsumption and greenhouse gas emissions.

Many studies have addressed the green provision-

ing of resources to reduce energy consumption inCloud environment. Most of these works use staticallocation methods such as FCFS (First-Come-First-Serve) to ensure performance and quality of services(Calheiros et al., 2011). These strategies are very sim-ple, but the problem with them is the need of manyavailable resources. The energetic efficiency of theseresource provisioning methods depends on the num-ber of customers applications. Other methods usedin the Cloud environment are deployed pre allocationstrategies (Nair and Jayarekha, 2012). But the prob-lem here is the prediction of required resources, whichis difficult.

Unlike theses works, we propose to include greenaspect in Cloud environment to support the GreenCloud computing. Indeed, the present study al-lows finding a solution to the green provisioning byscheduling customers applications to optimize the useof Cloud resources and therefore reduce energy con-sumption as well as the emission of CO2.

We use the information of the Cloud infrastructureresources and their relations with energy consumptionand CO2 emissions. The main objectives are to min-

135

imize these two factors on the one hand, and to re-duce the cost of services on the other hand. To opti-mize these objectives, we include our parameters in amulti-objective genetic algorithm and we execute theoptimal solution in our Cloud which is tested as anextension of the CloudSim framework.

The remainder of this paper is organized as fol-lows. Section 2 discusses related work. A detaileddescription of our solution is presented in section 3.Then in section 4, we provide technical details of ourwork. Section 5 deals with the experiments and theresults of the simulations which are produced withcomparisons and detailed analysis. Finally, the paperends with brief conclusive remarks and discussion onfuture studies directions.

2 RELATED WORK

Cloud resources optimization is difficult to meet be-cause of the uncertainty of future consumer demandand resource prices. It has been a topic of research in-terest and development for many years. To address thegrowing challenge, techniques from many disciplineswere integrated synergistically. Next, we present thestate of the art of cloud resources optimization meth-ods.

Cloud computing’s usage-based pricing modelcreates an incentive for subscribers to optimize theutilization of the rented resources. Borovskiy et al.(Borovskiy et al., 2011) devise a formal approach fordistributing workload among a minimum number ofservers. They model this problem as a linear program-ming problem and describe two solution approaches.The first one generates a set of candidate blocks andthen composes an optimal partition by solving an in-teger programming problem. The second approachsolves the set partitioning problem with column gen-eration technique. The disadvantage of such methodis its difficulty to consider the purpose of Cloud re-sources optimization because of the nonlinear charac-teristics of users’ demands distribution.

Chaisiri et al. (Chaisiri et al., 2012) propose amethod to optimize Cloud resources cost. The underprovisioning problem can occur when the reserved re-sources are unable to fully meet users’ demands dueto its uncertainty of the workload distribution. Toaddress this problem, the authors propose an opti-mal cloud resource provisioning algorithm based onstochastic programming model.

Regarding the problem of the description of theCloud resources characteristic with nonlinear equa-tion, some researchers propose the use of a stochas-tic optimization approach. For example, Li proposes

a model based on stochastic integer programming forCloud resources optimization (Li, 2012). He proposesto address the SLA-aware resource composition prob-lem. He defines a stochastic integer programmingmodel for resource composition and provides an algo-rithm that implements Grbner based theory to solvethis problem (Buchberger, 2001).To solve the mini-mization problem of Cloud infrastructure cost, Zhaoet al. developed a deterministic model for resourcereservation planning, using a mixed integer linear al-gorithm, to generate optimal decisions given fixed pa-rameters (Zhao et al., 2012). In addition, they pro-posed a stochastic model of resource rental planningwhich explicitly takes into account the uncertainty ofresources and users’ demand in the decision makingprocess. One major disadvantage of such approaches,especially in dynamic environments where the opti-mal solution changes over time, is that the parameterestimation phase significantly delays the implementa-tion of an optimal solution.

Other researchers use the constraint satisfactionproblem (CSP) approach to solve the problem ofCloud resources optimization. Van et al. proposea two-level based architecture that defines a clearseparation between application specific functions anda generic global decision level (Nguyen Van et al.,2009). They use utility functions to map the currentstate of each application for a scalar value that quan-tify the ”satisfaction” of each application in termsof its performance targets. These utility functionsare also means of communication with the layer ofglobal decision which builds a global utility functionincluding the costs of resource management. Thestage of provisioning of virtual machines has beenseparated from the stage of placement of virtual ma-chines within the global decision layer loop and for-mulates both problems as constraint satisfaction prob-lem. Doughertya et al. propose a model driven ap-proach to optimize the configuration, the energy con-sumption and the cost of infrastructure for Cloud in-frastructure self-scaling to create green IT environ-ments that reduce emissions resulting from the use ofredundant resources unused (Dougherty et al., 2012),(Dougherty et al., ). They proposed to decompose themodel to four sub-problems to ensure infrastructureauto-scaling: explaining how virtual machine config-urations can be captured; describe how these modelscan be transformed in constraint satisfaction problemsfor the configuration and optimization of energy con-sumption; showing how optimal auto-scale configura-tions can be derived from these CSPs with a constraintsolver and present a case study of energy consumptionand cost reduction of production of this model-drivenapproach. The main drawback of these methods is

CLOSER�2014�-�4th�International�Conference�on�Cloud�Computing�and�Services�Science

136

there exponential complexity.

Bio-inspired scheduling algorithms are often usedin heterogeneous computing environments. Chaisiriet al. propose an optimal VM placement algo-rithm which optimally allocates VM to multiple cloudproviders and follows optimally in advance the re-sources reservation (Chaisiri et al., 2009). This algo-rithm minimizes the cost virtual machines hosting inan environment of multiple cloud providers. Van et al.show that the ability to automate the provisioning anddynamic placement of virtual machines, taking intoaccount both the software-level’ SLA and resourcecost with high-level handles for the monitor to spec-ify compromise between the two (Van et al., 2009).The model defines a support for heterogeneous appli-cations and workloads including both enterprise on-line applications with stringent QoS requirements andbatch-oriented CPU intensive applications. It is notfocused on optimization problems that are NP-hard intheir general form.

Kessaci et al. (Kessaci et al., 2011) proposeto optimize the allocation of VM requests using aPareto-based meta-heuristic approach. In fact, theyuse a multi-objective genetic algorithm and proposeto adopt new mutation and crossover strategies to pro-duce new generations. The three objectives are con-sidered in the optimization process: minimize bothenergy consumption and CO2 emissions of the cloudinfrastructures and maximize the profit of the suppli-ers. To formulate the problem, they use a real en-coding. Each individual is a vector representing theresult of processing a pool of applications receivedduring the scheduling cycle. The used encoding iden-tifies three main features: the index of the vector rep-resents the applications, the value of each cell identi-fies the VM on which the application will be sched-uled and the maximum number of application. TheInitialization of the MOGA population is divided intothree steps: read the application pool with the greedymethod (Black, ), initializes one or two elements ofthe population by the result of the first step and ran-domly initializes the rest of the population. The prob-lem of this approach is the use of the greedy algo-rithm to initialize the population of the MOGA algo-rithm. Despite their simplicity, greedy algorithms canbe subtle and they are costly and mostly provide a lo-cal minimum. All of the presented approaches takeinto account the optimization of the Cloud resourcesbut they do not consider the relationship between thesatisfaction of users’ requests and the optimization ofproviders’ infrastructures cost. They do not pay at-tention on how each one of those parts can affect theother. In fact, the optimization of cost and responsetime of clients’ requests are closely related to the

number of available and active resources. Our objec-tive is to minimize the number of the active resourcesthat minimize energy consumption. To resolve thisproblem, we propose to use a multi-objective opti-mization. The initialization of the MOGA populationis a real time process. It amounts to read the applica-tion pool and extracts useful information for the opti-mization algorithm.

3 THE PROPOSED APPROACH

Cloud computing can be represented in three maincategories which are based on the capacity of abstrac-tion and the paradigm of services. Thus we have theSaaS, PaaS and IaaS. In our work, we focus on IaaSas our goal is Cloud requests scheduling.

Iaas provides processing capacities and storage aswell as network components as standardized services.These services manage a workload requested by cus-tomers applications.

3.1 The Mathematical Model

The Cloud model adopted by this study is IaaS withtwo-tier architecture as shown in figure 1. On the onehand, we have the Cloud services provider and on theother hand, the customer applications which need ser-vices.

Figure 1: The architecture of our Cloud model.

The optimization of our objectives, which are theenergy consumption and the CO2 emission, is owedto the exploitation of features offered by geographicaldistribution of data centers. Indeed, the parameters ofenergy model as well as CO2 model are different fromone geographical site to another.

In the Cloud, there are two sources of energy con-sumption: energy from computing equipment, whichis the energy required for calculation and energy fromauxiliary equipment, which is the energy required forcooling.

To make the energy equation, we use the model ofCMOS (Complementary Metal-Oxide Semiconduc-tor) processors by adding two constraints of the prob-lem which are run time (trun) and number of proces-sors required to run the customer application (nbpr).Thus, the final formula of the energy required for cal-culation is

Ecal = (a f 3 +b)� trun�nbpr (1)

A�Green�Model�of�Cloud�Resources�Provisioning

137

The energy performance coefficient (COP) is theratio between the produced heat and the energy con-sumed in the treatment. This factor leads us to deductthe cooling energy which is

Eaux =Ecal

COP(2)

Therefore, the energy to minimize is the total en-ergy:

Etotal = Ecal +Eaux (3)Regarding the second objective, which is the emis-

sion of greenhouse gas, we use in its formula CO2rate, which depends on geographical locations. Theformula of this second objective is

CO2 = Etotal�RateCO2 (4)

To optimize simultaneously these two objectives,we use the multi-objective optimization techniques(Goldberg, 1996).

3.2 The Multi-objective Optimization

A multi-objective optimization problem is to optimizeseveral objective functions simultaneously. Our opti-mization problem is defined as follows:

Minimize F(x) = ( f1(x); f2(x)) where both func-tions to optimize f1 and f2 are respectively the func-tion of energy and the function of CO2, x = (x1; ;xn)is the vector of decision variables which are the runtime and the number of required processors, F(x) isthe vector of objectives which will be optimized.

The single optimal solution in the mono-objectiveoptimization problem is replaced by the concept ofPareto optimal solutions in multi-objective optimiza-tion problem. Therefore, the optimal solution is not asingle solution but a set of solutions. To find the rightbalance of solutions, it is necessary to identify the re-lation between these objectives. The most used rela-tion is the relation of Pareto dominance. For this rela-tion, all efficient solutions are called the Pareto Front.This set of solutions is a balance where no improve-ment can be made on an objective without degrada-tion of at least another objective. So the purpose is toobtain the Pareto front or converge as much as possi-ble on this front. Figure 2 shows an example of dom-inance relation in case there are two objectives to bemaximized.

To solve our problem of multi-objective optimiza-tion based on Pareto-solving methods, we use theheuristic algorithms. These algorithms are used toexplore the possible solutions space seeking the bestsolution. Among these algorithms, we use multi-objective genetic algorithm (MOGA).

Different models have been proposed for multi-objective genetic algorithms including VEGA (Vector

Figure 2: The relation of dominance.

Evaluated Genetic Algorithm), NPGA (Niched ParetoGenetic Algorithm), NSGA (Non Dominated SortingGenetic Algorithm) etc. We adopt in our project, theNSGA-II because it uses an elitist approach that savesand injects the best solutions found in previous gen-erations in the new generations (Melcher, 2007),(Debet al., 2002). It uses a sorting procedure based on thenon-dominance which is faster. It requires no param-eter setting. It uses also a comparison operator basedon a calculation of the Crowding distance. This dis-tance is calculated from nearby solutions.

In our study, this algorithm makes the schedulingof the execution of customers applications with re-source optimization. And to assess the effectivenessof the algorithm as well as the solution, we calculatethe energy consumption and the emission of CO2 afterthe execution of the applications.

4 SIMULATION ENVIRONMENT

4.1 CloudSim Extension

Since CloudSim provides an extensible frameworkfor modeling and simulation of Cloud computing in-frastructures and services, the present work is in-cluded in CloudSim as an extension. In fact, to cre-ate our application, we leverage some basic functionsof CloudSim and we extend some characteristics andfeatures.

In CloudSim, the management of Cloud resourcesand all allocation policies are standard and don‘t con-sider some characteristics of data centers. So, in thepresent work, we use this specification to investigatenew model for resource allocation. This new tech-nique is in accordance with ecological standards.

4.2 Class Diagram

Our application divided into two classes diagram:The first class diagram is presented in figure 3. In

this diagram we have three types of classes.To meet the needs of our application, we add two

classes. The first one is Localisation. This class con-tains all locations specifications of Datacenter. The

CLOSER�2014�-�4th�International�Conference�on�Cloud�Computing�and�Services�Science

138

Figure 3: Project infrastructure class diagram.

Figure 4: MOGA implementation class diagram.

second class is Process. This class is in charge of allcommunications between graphical interfaces, dataaccess, Cloud configuration and simulation. Also, itallows to manage all Cloud components.

We modify the class DatacenterBroker to adaptit to our application. This class denotes a Cloud bro-ker. It is a mediator between users requests and Cloudinfrastructure. Since DatacenterBroker contains theprocess of creation of Cloud infrastructure and man-agement of resource allocation policies, we changesome details and we implement our new model in thisclass.

The other classes are those that we use in this workto create the environment of our new extension. The

infrastructures of Cloud are represented by the classDatacenter which manages physical machines. Tech-nical static properties of datacenters are in Datacen-terCharacteristics class and desired functionalities ofa storage system are in Storage Class.The class Host represents a physical machine whichhas one or more processing element. Also, it hasmemory and bandwidth, managed by allocation poli-cies, storage capacity and provisioning policy for as-signment to one or more virtual machines. Each pro-cessing element of hosts, represented by Pe class, hasa processing capacity. The allocation in virtual ma-chines depends on physical characteristics of hosts.User applications, represented by the class Cloudlet,

A�Green�Model�of�Cloud�Resources�Provisioning

139

are deployed in virtual machines by provisioning pol-icy CloudletSchedulerTimeShared. Virtual machinesare distributed in Datacenters using the VmAllocation-PolicySimple class.

The class diagram presented in figure 4, shows theclasses used to implement MOGA namely NSGAII.The DatacenterBroker class from the first diagram isused for linking this group of classes to the other andis used also for running the algorithm.

In this diagram, we have 2 types of classes.Classes of library for the NSGA-II algorithm that

we use: the core of the algorithm is implemented inNSGA2 class. In this class, we create an instance ofNSGA2Configuration to specify and store all neces-sary parameters for genetic algorithm.The instance of the class NSGA2Event is created ineach generation during the run of the algorithm. Inthis class, we store information about the current sta-tus of the algorithm. We use this class in Assign-mentNSGA2Listener to extract information and printit.

Classes that we use to personalize our algorithm:AssignmentIndividual class implements the class In-dividual. Its used to describe the populations of ge-netic algorithm. Each individual present a possiblesolution of the resource assignment problem.For each fitness function, we create a specificclass which implements FitnessFunction interface.Since we have 2 fitness functions, we implement2 classes. The first one is EnergyFitnessFunction,which contains the energy function and the second isCO2FitnessFunction which contain CO2 function.To observe the performance of our genetic algorithm,we implement the NSGA2Listener interface using As-signmentNSGA2Listener class. This class shows fit-ness function values and other detailed data of the bestindividuals found during the run of the algorithm.

When the genetic algorithm finish, it returns thebest found populations which are non-dominated so-lutions.

5 RESULTS ANALYSIS

In this section, we present the experiments and theevaluation that we undertook in order to study theefficiency of our extension of CloudSim in terms ofCloud computing environments optimization.We deploy two sets of tests. In the first one, we de-cide on the value of MOGA parameters. Then, inthe next test, we simulate and we analyze the Cloudenvironment by taking into account the extension ofCloudSim and we compare the results with the initialexisting approach in CloudSim.

5.1 Parameters of the Algorithm

Genetic algorithms have four parameters. To maxi-mize the efficiency of our algorithm, we have to makea good choice of its parameters values.

Regarding the first parameter which is the size ofthe population, it is equal to the number of customersapplications to optimize. By varying the values ofstop criterion of the algorithm, which is the maxi-mum number of generation, we noticed that the bestindividuals are always found before the 1000th gen-eration. Accordingly, we consider this value as themaximal generation number and as sufficient to findthe solution.

In theory, it was found that the values of the pa-rameters of evolutionary algorithms vary in a specificinterval. Indeed, former studies (Mais et al., ) haveshown that the best results are achieved by a valueof crossover probability between 0.45 and 0.95, and avalue of mutation probability between 0.01 and 0.005.To fix these two parameters, we kept the same struc-ture of the Clouds environment, the same resourcesand the same customers applications. Then, we varythe two remaining parameters of MOGA to choosethe values that give the best results.

To test the effect of the different values of mu-tation probability on our Cloud environment, we as-sume that the value of crossover probability is 0.9which is a predetermined value. Also for the test ofcrossover probability, we use the predetermined valueof mutation probability which is 0.05.

The evaluation of tests show that the variationof probability values of the two parameters in thetheoretical range gives almost stable results. Hence,we keep the predefined value of crossover probabilityand we choose the upper border of the interval ofmutation probabilities. Consequently, values of0.9 as crossovers probability and 0.01 as mutationsprobability may give a satisfactory result. We keepthese values to simulate the different test cases.

5.2 Simulation Results

The purpose of this experiment is to discuss the per-formance analysis of our approach compared withstatic allocation method in terms of the efficiency ofresource utilization for the same workload. The staticallocation method used is the method deployed in theframework CloudSim before extension.

In these experiments, we calculate 2 metrics. Thefirst one is the total energy consumption by the phys-ical resources of data centers caused by customer ap-plications workloads which is presented in figure 5.

CLOSER�2014�-�4th�International�Conference�on�Cloud�Computing�and�Services�Science

140

The second one is the total emission of greenhousegas caused by energy consumption of data centerswhich is shown in figure 6.

Figure 5: Energy consumption.

Figure 6: The emission of CO2.

To demonstrate the amount of this optimization,we calculate the percentage of decrease of energyconsumption as well as the emission of CO2. The re-sult of this demonstration is presented in figure 7.

The experimental results show that our proposedapproach provides better results compared to resultsobtained by static method of resource allocation. In-deed, from these results we can conclude that for theuse of our solution on Cloud environment, the energyconsumption is reduced to 50% and CO2 emissionsup to 60%. This reduction depends strongly on thecharacteristics of available resources and the amountof applications which will be run on the Cloud. Theobtained results are due to the efficient scheduling ofcustomers applications, and to the reduction of the useof energy-intensive resources.

Figure 7: Percentage of decrease.

6 CONCLUSION

To address the problem of energy efficiency, we haveproposed in this paper a new approach for a Cloudcomputing environment that schedule resources allo-cation based on energy optimization functions. Moreprecisely, the presented work has optimized the re-sources in the Cloud, and has minimized the rate ofenergy consumption and CO2 emissions by the man-agement of Cloud computing resources and the effec-tive choice of our genetic algorithm parameters. Theexperimental results have shown that the proposedapproach leads to a significant reduction in energyconsumption and CO2 emissions in comparison withstatic techniques of resource allocation.

This study opens more challenges to reduce theimpact of new technologies on the environment, en-courage the ecosystems, and support energy effi-ciency.

In future work, we plan to integrate in our project,an agent system in order to make the Cloud environ-ment auto-adaptive. A study is underway in order tofulfill this objective.

REFERENCES

Black, P. Greedy algorithm. dictionary of algorithms anddata structures (online), us national institute of stan-dards and technology, february 2005.

Borovskiy, V., Wust, J., Schwarz, C., Koch, W., and Zeier,A. (2011). A linear programming approach for opti-mizing workload distribution in a cloud. In CLOUDCOMPUTING 2011, The Second International Con-ference on Cloud Computing, GRIDs, and Virtualiza-tion, pages 127–132.

Buchberger, B. (2001). Grobner bases: A short introduc-tion for systems theorists. In Computer Aided SystemsTheoryEUROCAST 2001, pages 1–19. Springer.

A�Green�Model�of�Cloud�Resources�Provisioning

141

Calheiros, R. N., Ranjan, R., Beloglazov, A., De Rose,C. A., and Buyya, R. (2011). Cloudsim: a toolkit formodeling and simulation of cloud computing environ-ments and evaluation of resource provisioning algo-rithms. Software: Practice and Experience, 41(1):23–50.

Chaisiri, S., Lee, B.-S., and Niyato, D. (2009). Opti-mal virtual machine placement across multiple cloudproviders. In Services Computing Conference, 2009.APSCC 2009. IEEE Asia-Pacific, pages 103–110.IEEE.

Chaisiri, S., Lee, B.-S., and Niyato, D. (2012). Opti-mization of resource provisioning cost in cloud com-puting. Services Computing, IEEE Transactions on,5(2):164–177.

Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002).A fast and elitist multiobjective genetic algorithm:Nsga-ii. Evolutionary Computation, IEEE Transac-tions on, 6(2):182–197.

Dougherty, B., White, J., and Schmidt, D. C. Model-drivenconfiguration of cloud computing auto-scaling infras-tructure.

Dougherty, B., White, J., and Schmidt, D. C. (2012).Model-driven auto-scaling of green cloud computinginfrastructure. Future Generation Computer Systems,28(2):371–378.

Goldberg, D. (19 janvier 1996). Algorithmes genetiques:exploration, optimisation et apprentissage automa-tique. Addison-Wesley France, SA, Kluwer AcademicPublisher.

Kessaci, Y., Melab, N., Talbi, E.-G., et al. (2011). Optimi-sation multi-critere pour l’allocation de ressources surclouds distribues avec prise en compte de l’energie. InRencontres Scientifiques France Grilles 2011.

Li, Q. (2012). Applying stochastic integer programming tooptimization of resource scheduling in cloud comput-ing. Journal of Networks, 7(7):1078–1084.

Mais, H.-R., Bloch, C., RAMDANE-CHERIF, W., andChatonnay, P. Differentes operateurs evolutionnairesde permutation: selections, croisements et mutations.

Melcher, J. (11 September 2007). Jnsga2 1.1 - a free javalibrary for the nsga-ii algorithm.

Nair, T. and Jayarekha, P. (2012). Pre-allocation strate-gies of computational resources in cloud computingusing adaptive resonance theory-2. arXiv preprintarXiv:1206.0419.

Nguyen Van, H., Dang Tran, F., and Menaud, J.-M. (2009).Autonomic virtual resource management for servicehosting platforms. In Proceedings of the 2009 ICSEWorkshop on Software Engineering Challenges ofCloud Computing, pages 1–8. IEEE Computer Soci-ety.

Relaxnews (02-04-2010). Data centers et cloud computingmettent trop de co2 selon greenpeace.

Thrash, B. (27-02-2012). The green data center opportunity.TUAL, M. (14-04-2013). Data centers: la donne colo.Van, H. N., Tran, F. D., and Menaud, J.-M. (2009). Sla-

aware virtual resource management for cloud infras-tructures. In Computer and Information Technology,2009. CIT’09. Ninth IEEE International Conferenceon, volume 1, pages 357–362. IEEE.

Zhang, Y., Wang, Y., and Wang, X. (2012). Electricitybill capping for cloud-scale data centers that impactthe power markets. In Parallel Processing (ICPP),2012 41st International Conference on, pages 440–449. IEEE.

Zhao, H., Pan, M., Liu, X., Li, X., and Fang, Y. (2012). Op-timal resource rental planning for elastic applicationsin cloud market. In Parallel & Distributed ProcessingSymposium (IPDPS), 2012 IEEE 26th International,pages 808–819. IEEE.

CLOSER�2014�-�4th�International�Conference�on�Cloud�Computing�and�Services�Science

142