A Performed Load Balancing Algorithm for Public

Cloud Computing U sing Ant Colony Optimization

Awatif RagmaniC ), Amina EI Omri, Noreddine Abghour, Khalid Moussaid, Mohammed Rida
Modeling and Optimization for Mobile Services Research Team
Faculty of Sciences Ain chock
University Hassan 11 Casablanca, Morocco

Abstract- In the space of a few years, Cloud computing has will focus particularly on the optlIIllzation 01' Cloud
experienced remarkable growth. Indeed, its economic model performance. Specifically, we analyze the impact of load
based on demand of hardware and software according to balancing algorithms on Cloud performance. In fact, an optimal
technical criteria (CPU utilization, memory, bandwidth ...) or response time will have the advantage of offering users the
package has strongly contributed to the liberalization of ability to work on almost identical way 01' remote resources to
computing resources in the world. However, the development of local resources. The study of performance in the Cloud can be
Cloud computing requires the optimization of performance of the
done through different aspects as identified in the research
different services offered by Cloud providers in order to ensure a
works [1]-[4].
high level of security, availability and responsiveness. One major
aspect for dealing with performance issues in Cloud computing is In summary, we note that Cloud is a distributed system,
the load balancing. In fact, an efficient load balancing contributes which shares thousands of computing resources. The Cloud
to the decrease of costs and maximizes availability of resources. uses Internet links to connect users to providers. The
Through this paper, we study various research which has treated distribution 01' workloads between different nodes within the
the load balancing in the Cloud computing. In a second step, we Cloud network is a crucial step in the process of optimizing the
propose an improved load balancing architecture and algorithm
overall workloads. Hence the importance 01' load balancing to
for Cloud computing which aims to allow a better response time.
allocate efficiently the various available resources including
Cloud network links, central processing units, disk drives, or
Keywords- Cloud computing; Load balancing; Scheduling;
other resources. Particularly, using a unique hardware load
Ant colony optimization
balancer for central administration 01' hundreds 01' processors
I. INTRODUCTION proves costly in terms of time and money. In addition,
oversized hardware remains underutilized if traffic declines [2].
Cloud computing is a highly promising technology concept.
With its fundamental concept of resource use on demand, The objectives of our study include the realization of
Cloud computing has greatly contributed to the liberalization 01' comparative analysis 01' the different load balancing algorithms
IT resources. Thus, wide kinds of organizations as weil as within the Cloud in order to propose a more efficient solution.
individuals become able to access IT resources that were The remainder 01' the paper is organized as folIows. The main
previously beyond their means. The concept of pay-as-you-go works on the load distribution problem in the Cloud are
applied by Cloud providers fits in both economic and discussed in Sections 3 and 4, Section 5 introduces a load
technological trends of different organizations. Particularly, the balancing system for Cloud computing and we will conclude
use 01' virtualization facilitates the sharing 01' resources and the with Section 6.
administration of the whole Cloud system. In despite of the
reluctance 01' some actors to adopt the Cloud, all economic 11. CLOUD COMPUTING

trends predict strong growth of Cloud applications within the The Cloud computing is defined as an advancement of the
coming years [1]. However, the development 01' Cloud cannot previous distributed model like Grid Computing. The main
be done without the continuous improvement of its goal behind its expansion is to transform the computing
performance including response time 01' Cloud services and the resource to service. Cloud computing is based on the concept
energy consumption of data centers. Indeed, it was noted that of offering to users hardware, software applications, data
each increase in the average response time induces a decrease management, and storage as a service. In other words, the users
of users' number and revenue [2]. do not need to know the location of servers or any other
technical aspects 01' the computing infrastructure. Cloud
The excitement experienced by the Cloud was the initiating
computing aims at scalability by using load balancing. In
factor of several studies conducted by researchers as weil as the
Cloud computing, services offered to users is not limited to
giant Cloud providers like Amazon, Google, IBM and
processing power and include storage, website hosting,
Microsoft. These studies have as key objectives to develop the
database support and more others [5].
area 01' applications covered by the Cloud, optimizing
performance and improving security. During this paper, we

978-1-4673-8894-8/16/$31.00 ©2016 IEEE

hardware and software partitioning. the importance 01' choosing the users. the first feature 01' Cloud computing is the scalability by III. B. different organizations which are motivated by shared As consequence. the Private Clouds are implemented on behalf of a simulation. the virtualization incIudes multiple concepts such as computing on private networks within a specific organization. Finally. The quality 01' service like CPU performance. operating systems and required services tor a particular application [6]. 1. configuration is a hybrid Cloud environment which includes There are two types 01' virtualization in case 01' Cloud. the particular user in order to maintain the full control over data. 11 On-Demand Sell- service J applications remotely from the Cloud. Indeed. but includes in addition. Particularly. Briefly. virtualization has a crucial role on system optimization by security. Otherwise. The first implantation of Cloud computing does not require any algorithm is used to allocate the jobs and divide tasks into sub particular capital expenditure because users pay only for tasks as three-tier architecture. and quality of service. Virtualization the Internet via web applications and services. they proposed a scheduling algorithm Finally. Visual model of NIST working definition of Cloud Computing [7] There are four types of Cloud computing. the Cloud computing systems can be presented as autonomous systems and are administrated transparently to In [10]. machine Particularly. The most popular is the public Cloud where computing resources are shared over D. the Platform-as-a-Service (PAAS) layer which is similar to IAAS. Types oi Cloud Computing ee ee Fig.A. computer hardware into multiple execution environments. The second type of virtualisation is the Para C. the balancing (OLB) and the load balance Min-Min. distributed system. Cloud Computing Feature virtualization where the hardware allows multiple operating The main strength of Cloud computing lies in the new systems to operate on unique machine by efficient use of characteristics it introduces compared to previous computing system resources such as memory and processor [9]. The Software-as-a­ Resource Pooling I Service (SAAS) layer which offer to users the possibility to run I Broad Network Access Jl Rapid Elasticitv J! M easured Servlce . The second kind is the Private Cloud which refers to Cloud Thus. and Cloud computing possible. Public Clouds The virtualization is the key technology which makes the services and applications are provided by third parties. The second algorithm in turn. first category is the full virtualization where an entire Lastly. Private Cloud can be developed reducing and simplifying the administration of server and administrated by the internal IT department 01' the infrastructure or being a mean 01' modifying the global organization or by an external Cloud provider. Those tools and methods are mostly centric interface which means that Cloud interfaces and user's inspired from the previous techniques applied within location are independent and can be used through web services. This concept can be defined as a Cloud computing resources are shared between different global framework and process of sharing the resources of customers. The second layer is the Infrastructure-as-a-Service (IAAS) Service models which refers to computing resources as a service. bandwidth appropriate node in the Cloud for the execution of a task. This kind of Cloud allows multiple entities to software present in the actual server. the virtual machine obtained will have all the requirements. emulation and quality of service. The and memory capacity are guaranteed by a service level authors have also introduced the use 01' agents in collecting agreement (SLA). services and capacity they use [6]. The third management approach of data center [8]. model like resources sharing and IT cost optimization. system among multiple users and separating users from each other. time sharing. Deployment models Finally. several researchers have attempt to find out special techniques for optimizing load balancing and The second feature of Cloud can be summarized as user­ distributing workload. serves to improve the execution time of each node. 1). available information such CPU capacity and remaining memory. RELATED WORK offering resources and services for users on demand. the pricing model of Cloud services offers the users gathering the qualities of two algorithms the opportunistic load the opportunity to optimize their IT investment. Layers oi Cloud Computing Essential Characteristics ! Cloud computing can be defined as set 01' services organized on three layers (see Fig. the Community Cloud is a platform used by installation 01' one machine is performed on another machine. within the Cloud system. the authors stressed. The both public and private Cloud [6]. . This includes [ [ II I SAAS PAAS I AAS virtualized computers with defined processing power and Software as a service Platform as a service I Infrastructure as a service bandwidth for storage and Internet access. Full virtualization has share specific business applications or rent in common been doing weil for different reasons Iike sharing a computer hardware resources. The Due to the importance and complexity of load balancing resources are scalable over several data centers.

highlight the fact that the Definition relying on a priori view of the system state at proposed algorithm is more efficient than the previous knowledge about the run-time and some algorithms. Execution of a dynamie lt distributes processes to load balancing algorithm Principally. system failed . LOAD BALANCING ALGORITHMS' CATEGORlES objective ant colony system algorithm in order to ensure an efficient placement of virtual machine (VM). Its optimization is based on the balancing of satisfy two classifications based on the way that load is workloads between different processes and reducing inter­ distributed and how processes are allocated to nodes and processes communication. Thus. partitioning and mapping steps. the authors summarized the role of load balancing assigning each task to a process. In other words. Particularly. As the number of processes depending on the status of the nodes. load balancing (see Table I. algorithm. This step is called the in the Cloud. the authors aim to implement solutions which reduce both the total resource wastage and power consumption. In [14]. the authors announce that JIQ algorithm incurs no balancing algorithms should: overhead communication between dispatchers and processors at job arrivals. The objective of this perform one or more tasks of the program. the authors of [17] attend to study the load balancing quality and static load balancing is balancing in Cloud computing by introducing a semi run-time overhead the only compatible centralized and multi cluster architecture. The authors highlight the fact parallel programs for overhead that the pheromone update has been proved as an optimized which the execution • They are aiming for a times of processes and means to balance the load. However. In the first case it is generated by a program is often different from the number of presented as centralized vision. the large and complex Cloud environment. the authors presented a load balancing model for • Keep up the system stability. the task study is to propose a hybrid algorithm combining the properties is the smallest division of a parallel program. • The design and • It is not satisfactory for security and availability. A heuristic algorithm using ant colony any run-time overhead They incur non- optimization concept has been proposed to implement a service • They are attractive for • negligible run-time load distribution for the Cloud. dynamic or adaptive overloading or under loading of a processor by processes. Consequently. the of biased random sampling algorithms and honeybee foraging development of a parallel program requires at first. extremes representing a predicted certain tradeoff among • In some situations. the optimization of Cloud resources cannot be done without an efficient load balancing strategy [18]. Another research study [11] have coped with the analysis of IV. Finally. distributed or hybrid and in the processor in charge of processing the request. in most cases maintaining a coherent two single-objective algorithms. analysis of dynamic' s parallel programs that Disadvantage load balancing are of the dynamic or algorithms is a unpredictable kind complex process . public Cloud. static or dynamic balancing strategy. A main controller chooses the suitable partitions for arriving jobs while the According to their working strategy. processes and the system negotiation policy for on which they run process migrations across Another research study [16]. there is a risk of second approach it is defined as static. the proposed technique Advantages limited balanced state their communication does not address the fault tolerance issues and more works are between the two requirements can be necessary to fulfill this aspect. the comparison of performance of the proposed processors at cOlnpile requires some means for algorithm to the existing multi-objective genetic algorithm and time. Through this Static Dynamic paper. the load Particularly. the authors indicate that an extension of • Enhance the performance significantly. In Through the research study [13]. Thus. approach. tried to find out an optimized processors load balancing algorithm for the Cloud environment by using • They will not cause ant colony optimization. LOAD BALANCING ALGORITHM various load balancing algorithms applied in the Cloud in order A parallel program is a set of processes which aims to to define a new more powerful algorithm. Otherwise.) [18]. The authors of the research study [15] introduced a multi­ TABLE I. the basic algorithm JIQ should cope with very high loads using • Ensure a partial or complete backup plan in case of only local information of server load. The main idea of this choice paper is to find out a system which guarantees a highest level of performance and to keep an efficient load balancing. the model divides global Cloud into several Cloud partitions in order to simplify load balancing of • Put up future update of the system. load balancing balancer tür each Cloud partition decides the best load algorithms are classified on two categories. the major goal of load balancing is to develop algorithm entitled 10in-Idle-Queue (JIQ) for load balancing in performed algorithms able to achieve both efficiently the distributed systems which introduce different advantages. the authors presented an summary. decomposing the overall program into a collection of tasks and In [12]. They note that load balancing algorithms must partitioning.

int1uential parameters of Cloud system. ) that could be modified in order to Fig. [20]. through this algorithm the user Load balancing algorithms in Cloud can be classified as first communicates with the load balancer to find out an immediate mode scheduling and batch mode scheduling. Firstly. 3. we introduce: time minimum is selected and assigned to the adequate node. Its invocation operation is defined by decision must weigh its three rules: location rule. which represents the machine manager which maintains a set 01' the tasks and number of tasks achieved in a given time. In this case. minimum completion time is estimated for all the Several algorithms of load balancing are cited in the literature available nodes. Each physical machine may host overhead cost against its distribution rule and multiple virtual machines. This weight is fixed on the basis of load capacity of each machine. virtual machines and resource provisioning policies. . maximum completion time is chosen and assigned to the related machine. Simulation Framework adequate machine based on its weight. If immediate mode scheduling organizes tasks based on its arrival an adequate machine is identified then the load balancer by applying a minimum execution time and minimum accepts the request 01' the client and assigns that virtual completion time algorithms. the essentially. This toolkit is An index of the status of developed on the basis 01' CloudSim [22] simulator by • • lt defines how to collect processor taking on a and maintain the improving and extending its characteristics. . Briet1y. In the case of short tasks are higher • Weighted Round Robin: this algorithm offers an in number than the long one. the data v. it is essential to be able to understand the processing times per individual requests. then each node is allowed to use a particular time • Max Min: this algorithm is similar to the min-min interval. Secondly. This algorithm is more adapted for the case where • Round Robin: this algorithm is an elementary the small tasks are larger in number than the great tasks. organizes tasks in a Meta task set based upon their arrival and • Min-Min: this algorithm copes with a task set which are they are mapped at prescheduled times to their related machines [19]. each virtual machine. However. The stability measures processes deal with priorities. the separation of the programming and experimentation functions and secondly to An initiation rule Aload balancing carry out repeatable experiments by varying simulation It dictates when to operation • parameters. load balance policy. and processors necessary for simulation and modeling of Cloud components such as data positive values as the making load balancig load i ncreases decisions centers. the Cloud components are identified by different parameters (user region. the tasks which have the machines basing on a rotating process. There are two main performance indicators of load • Equally Spread Current Execution: this algorithm balancing: stability and efficiency. this algorithm performed improved resources allocation by defining a weight to better than the Min-Min algorithm. scheduling method which applies the concept 01' time Min-Min is a simple and fast algorithm able to slices. the completion [19]. The main aspect 01' this distribution lies in the fact 01' paying no attention Before the definition of an efficient and optimized load to the advanced load balancing requirement like balancing algorithm. The CloudAnalyst zero value if the workload information of is an efficient toolkit which enables modelers to achieve processor is idle. virtual machine. the analysis 01' Cloud resources operating under real and varied conditions seems difficult. The efficiency the volume and relocates the load to those virtual indicates the time necessary to either reduce the variance or machines which is lightly loaded or should process that arrive at the equilibrium state [19]. In other words. In brief. storage . The adequate virtual machine for the required operation. This simulator allows firstly. significant key performance indicators in a distributed system • Throttled load balancer: this algorithm is fully based on [5]. Firstly. the capability of an algorithm to force any initial algorithm dispatches the load randomly by analyzing workload distribution into equilibrium state. the available tasks are sent to the selection of the virtual machines. After the calculation. need less time to fulfill the task and maximize indicated the time necessary from registering the input to throughput. Principally. composed 01' several hosts. the presented by four components as summarized on Fig. Otherwise. In this time interval the node will execute its tasks. This toolkit provides the ability to model each initiate a load balancing • A load balancing Cloud by defining one or more data centers which are operation. The response time. 2. Load Balancing Algorithms components . The Batch mode scheduling machine to the client. expected performance selection rule benefit As depicted in Fig. and throughput. this algorithm operates on random algorithm. the request is queued. the time is separated into various segments guarantee enhanced performance. represents the most their resources needed. 2. which task easy. Thus. The datacenter system and minimum completion times for all 01' them controller allocates the tasks to the available virtual are calculated. Each Load balancing algorithm can be primarily not allocates to any 01' the nodes. To overcome this difficulty we opt for the use Aload measu rement An information rule exchange rule of the simulation platform CloudAnalyst [21]. The tasks are transferred to the virtual providing a response to it. PROPOSED LOAD BALANCING SYSTEM center controller will assign a number 01' requests to A. memory.

locations.26 0.49 1 1 1 1 1 1 1 2 2 2 2 2 2 50.87 • Data Center Region • Bandwidth 1 1 1 2 2 2 2 2 2 2 2 1 1 50... ..85 Fig..:od '0 = . .� � " = � �W . a few additional experiments are achieved in order to highlight the fact that the most intluential factor in optimizing validate the combination predicted.. we highlight two facts.65 0. we choose the Taguchi L16 array size 01' the user's request.38 Host 1 2 2 2 2 1 1 2 2 1 1 1 1 500.complete the various experiments in economic and easy way..::! " '" = 0 � .11 � eData s ize per request 1 1 1 1 1 1 1 1 1 1 1 1 1 50. The Table 11 and the ranking 01' intluential factors presented in Taguchi's concept utilizes an optimal number 01' planned Table III.::! Ö � Q 0 '" 0 . CloudAnalyst platform. Yu: performance characteristic To conduct our study we apply the Taguchi experience measurements per trial. The CloudAnalyst apply a round robin algorithm as default load Ni 2 balance policy and proposes also a throttled load balance policy "\" Yu SNj = -10 10g(L.37 0.s CJl� o� � " " 0 = . describes request allocation within a data center.63 2 1 2 2 1 2 1 2 1 2 1 1 2 499. TAGUCHI ARRAY APPLIED AND RESULTS OBTA1NED Once the combination 01' control parameter is defined by the Taguchi array.� Ee " .. .85 0.66 1 2 2 2 2 1 1 1 1 2 2 2 2 500. the service broker policies parameter allows the noise ratio (SNR) to analyze the results obtained.. '" " achieved measurements are summarized in Table 11. 0 .17 0. � " '" 0 " � p '" -. it is established that the response figures and tables to analyze the obtained measurements and to time may be multiplied by 10 in the case 01' two remote predict the best combination to optimize the effect 01' factors. load balancing because it has the same number of parameters as the predefined system should make it possible to affect the user's request to simulation model (see Fig.41 Data Center 1 1 1 2 2 2 2 1 1 1 1 2 2 50. Memory • Number ofVirtual Machine 2 1 2 1 2 1 2 2 1 2 1 2 1 500.s rJ) " CJl U P 0 Ei Ei � � " Q � == " "C == '" � Cloud Computing Q .: . 0 ... plans [23].. This methodology inspired from the industrial sector offers the advantage 01' organizing and making easier the According to the simulation results summarized on the study of a complex system such Cloud computing.25 0. � '" .73 2 1 2 2 1 2 1 1 2 1 2 2 1 500.16 0.66 • NumberofProce'isor Speed ofProcessor • 2 1 2 1 2 1 2 1 2 1 2 1 2 500. the of the user who initiated the request have a strong impact on appropriate experiments are achieved and the performance the response time. u: trial number.25 0. In other words. .. measurements are recorded..35 Virtuall\llachine • NumberofProce'isor 2 2 1 1 2 2 1 1 2 2 1 1 2 50. Secondly.� e .51 0.39 2 2 1 2 1 1 2 1 2 2 1 2 1 50. = . TABLE 11. " ö =: IOIJ IOIJ� :.23 0... The load balance policy is the parameter that function defined on (1). '" CJl . .70 0... " Z Cloud User " Z == � Ei � � � == .. .69 0. . Ni: Number of trials for experiment. this approach use As mentioned on Table 11. the Ll6 array the closest data center and to optimize the data size and has 13 factors and each factor has two levels. 0 CJl . the factors effect ranks (see Table III) Lastly. = .� . Each array cell frequency of request performed..45 • Speed ofProcessor • Storage. Cloud Computing simulation model .. Memory 2 2 1 1 2 2 1 2 1 1 2 2 1 50...12 0. 3).. B. The Taguchi experiment plans make use of the signal-to­ Particularly.. rJ) Load balance policy 0- • ·User Region • Service brokerpolicy " Z -Requestperuser/ HOUT • Data Center 1... ti) (1) which constraints the requests being processed to a throttling u=l 1 threshold [21]. 3. The L16 array applied and the CJl " " ... define a qualitative or quantitative factor levels (1 or 2).46 2 2 1 2 1 1 2 2 1 1 2 1 2 50.08 0...29 1 2 2 1 1 2 2 2 2 1 1 2 2 500..90 0. In fact..19 0.. which determine the type of parameter values for each experiment. For each set 01' factors... the response time is the request per user followed by the data To fulfill our analysis.30 • Storage... .� � 0 � '" IOIJ � .2 .S "C � . " � " Ei 0 " � "C . In other words... the geographical experiments in order to examine the settings of influential location 01' the Data Center which performs the request and that process control parameters. Analysis of Influential Cloud Computing Factors Where i: experiment number. ..43 • Cost per service • Host number 1 2 2 1 1 2 2 1 1 2 2 1 1 499. Firstly. �. we have proceeded to simulation on the . the control of the data center which performs the user request at a optimization of outputs can be obtained by minimizing the given time. Afterward.

02 0. third layer is composed by two levels of nodes. 5.00 -44. while (value of Timer ) < T do The second layer embraces regional load balancers.02 -44.01 -44. components receive various tasks sent by the main controller if RegionalBalancerState == Idle then and schedule tasks received through the different nodes of the Cloud following an ant colony optimization algorithm. main controller load. Through different agents. they Search tür next RegionalBalancer.03 -44.02 Rank 10 11 I 2 7 8 3 Level NbrProcessor SpeedProcessor NbreVM Bandwidth Memory Storage I -44.01 -44.02 -44. FACTORS EFFECT RANKS FOR RESPONSE TIME INDICATOR Level UserRegion DCRegion RequestUser Datasize BrkPolicy BIcPolicy NbrHost I -44.98 -44.02 0.00 Rank 4 9 13 6 5 12 Load C.01 -44. Then it informs the main controller about the while Task do modified state of a node. Those SearchcloserLoadBalancer (Task). divided the tasks received on elementary subtasks in order to Ca1culateWaiting Time (Task).00 2 -44.00 -44.00 -44. This configuration aims to lighten the CalculateTechnicalratio(Task).01 -34.00 Delta 0.01 -44. the proposed load balancing architecture is organized on three layers.01 0.01 0. The first layer includes the main controller and the second controller.02 0. this organization allows ant colonies to achieve complex tasks which exceed usually the individual capabilities of a single ant. the second balancer updates the load level of the different nodes according begin to a frequency T.01 -44.00 19. The Send Task to RegionalBalancer. 5). region and technical characteristics and then dispatches users' tasks on regional load balancers as mentioned hereafter (see Algorithm land Fig.02 Delta 0.00 -44. This approach is driven by the conclusion of the Taguchi analysis which point out that the Ca1culateExpectingCompletionTime (Task) . The first level else receives the tasks sent by regional load balancer. "Request size" factor has a high impact on the performance if Waiting Time> Expecting Completion Time then response time. 4.00 -44. As depicted in Fig. Indeed. the Fig. TABLE TTT.01 0.00 0.01 2 -44.96 0. 4. In addition.01 -44. The Fig.00 -44.00 -53. else Put Task on the RegionalBalancer queue end if end if end while end while end D. The main controller is the first element to receive the users' tasks.00 0. Proposed Load Balancing Architecture Heavy state Balancer According to the above analysis. Ant Colony Optimization Ant colonies represent distributed systems which has a well structured social organization. Send Task to NextRegionalBalancer. an efficient load balancing State policy should allocate users' request to closest data center and allows optimization of request size.00 -44.00 -44.02 0. Then. As consequence. Task scheduling schema second controller has to communicate frequently with the regional load balancers in order to refresh the information Algorithm I: Task scheduling of the Main Controller about system state. The proposed load balancing architecture . the main End controller communicates regularly with the second controller in order to update information about the state of the system. minimize the size of users' requests.01 -44.01 -44. The main controller classifies users' tasks by priority.

(3) solutions to the optimization problem. . The initial main steps [26]. The first step 01' the proposed algorithm is the initialization The standard ant colony algorithm is composed by five of the pheromone per virtual machine VMj. as ants deposit pheromone on the elements or maintain an optimized level 01' load balancing. . Artificial ants in ACO represent the procedure 01' stochastic The different rewards are saved on a pheromone table. and the • Pheromone table theoretical understanding 01' properties 01' ACO algorithms [25]. the fifth Where Pecapacity is processing capacity. we decide to apply the ACO due to the Begin fact that this algorithm is highly compatible with the behavior Step 1: Initialize the pheromone 01' distributed network. Each node VMj 01' the Cloud system will updated based on the resuIts of ants' search experience and be identified by: existing heuristic information. each ant k select a VMj for three procedures. As consequence. Otherwise the intensity 01' pheromone update pheromone increases the probability that those elements that is given hereafter: were either used by many ants or that were used by at least one ant and which produced a very good solution will be used (4) again by future ants. The artificial pheromone information is Cloud environment. Memory is the enable memory and BW is bandwidth condition. balancing in Cloud environment. ACO apply the second procedure that is the process Where Ctj is the cost parameter 01' the VMj and LBj is the by which the pheromone trails are updated. or decline. Concerning the load optimization and distributed control problems [24]. In a different way. Numerous researehes have • Cost and technical capacity focused on the implementation of high-performing algorithmic possibilities and successful applications of ACO algorithms to • Probability to be chosen a large variety 01' computationally hard problems. ability indicators of the VMj. This solution construction which put up possible solutions for the control system should improve the load balancing system in optimization problem. end end . n . Stcapacity is storage step incIudes the updating of the global pheromone until end capacity. task. computing resources are Particularly. This value is equivalent to the quantity of pheromone. The main steps of the ACO load arch spaces. Secondly. Moreover. the deposit 01' new pheromone trial. ß and yare respectively the weight 01' the component 01' From a practical point 01' view. the ants' search experience while (termination condition not met) do can be applied to influence reminiscent of strengthening Step 2: Construct solutions for each ant k learning. Proposed ACO Algorithm examination 01' real ants' behavior. The first procedure administrates a colony 01' achieving a task with a probability defined as folIows: ants that concurrently and asynchronously progress through [Tj(tl([Ctl [LBt neighbor nodes. As the solution is being o otherwise built. pheromone evaporation implements a useful form of forgetting: it avoids a The pheromone value can increase or decrease and value of too rapid convergence 01' the algorithm toward a suboptimal region. Furthermore. the use of a colony of ants should Step 3: evaluate the fitness 01' each solution increase the algorithm robustness and in many ACO Step 4: save the best solution applications the collective interaction of a population of agents Step 5: update the pheromone value is needed to proficiently find out a solution to problem. Through this paper. the solution construction in future iterations 01' the algorithm. and use these models as a Let a Cloud be presented by a set 01' nodes which represent basis for implementing novel algorithms dedicated to the the available virtual machine VMj. the ant estimates the solution that will be used by procedure to decide how much pheromone to deposit. The first step is the initialization of the pheromone value 'Cj (0) of VMj is caIculated as folIows: Tj (0) variables. connections used. Each virtual machine VMj estimates a reward per trail laying and the behavior of a number of ant colonies. therefore favoring the exploration of new areas of these p vary between 0 and 1. An ACO algorithm can be summarized as the interaction of After the initialization step. The second and third steps take into account the = (Pecapacity + Stcapacity + Memory + BW) (2) different iterations. To finish. a. the third procedure is applied to put into balancing algorithm are summarized hereafter: practice centralized actions which cannot be performed by Algorithm 11: Ant Colony Optimization single ants. ant colony optimization (ACO) is a Meta dynamically increased or decreased following the users heuristic which is developed on the basis 01' the pheromone demand. due to pheromone evaporation. ants incrementally put up f I[Tk(tW[Ctklß [LBklY i j 1. Finally. Those values can load balancer indicator for VMj which allows the algorithm to either amplify.different research tried to study the models resulting from the E. The fourth step allows the building of solution and the updating 01' local pheromone.

Garcia."Load Balancing In Cloud Computing. Briefly. 2012. 56. Lau. Systems (JNS2). Chang and R. Web Semantic Techno!.K-Q. Firstly. Sei. N. 2012 IEEE 4th International [24] T. pp.pp. [9] A. S. Furht and A." Int." in 146.2009. initiative using modified ant colony framework.-C. London. Through this 2013. location 01' both user and data center in order to choose the 1. Yan. 1. R. M. 2010 Second [26] S. Apr. and each level has a specific role and specific algorithm. Lu.11 networks. 68. CONCLUSION AND FUTURE WORK [7] P." in Network Security and the updating state 01' the system through various agents. Technoi.pp." World Acad. we aim to optirnize the whole response times 01' [16] R.pp.. Evai. Taguchi." Softw. 2015. 79.. Load balancing in parallel computers: theory and practice. NJ. [20] M. Wu. Mare.. 2007. Through this paper we propose an improved [10] S.pp. Buyya. 108-113. London: Springer London. the "Join-Idle-Queue: A novel load balancing algorithm for dynamically proposed architecture 01' the Cloud is organized on three levels scalable web services. 2012. "The NIST definition of cloud computing. Handbook of Cloud Computing.-S." Peiform. and A. "Experimental validation of [25] M.Dec. vol.. Sei. 2010. 6. 275-280.2005." in Cloud Technologies and Applications (CloudTech). and X. and proposed algorithm.. Exp. 1. vol." Tsinghua Sei. Liu. Technology and Science (CloudCom). Calheiros. secondly in a real environment and finally to improve the [17] M. Kotz. "Analytical Performance Modeling for Computer computing environments and evaluation of resource provisioning Systems. Following this work. Y. Communication Systems and Networks (COMSNETS). K Garg. vol. 427-432. "Load load balancers. C. Kartit. [5] K Erciyes. 2012. Kliot. on the CloudSim simulator and Apr. R. and R. 2010. [4] S. Comput. Geiler. 8. Otherwise. Tay.pp. Pathari and K Calicut. Xu and F. 2010. XenServer." 2011. 2004." in Cloud Computing Mich: John Wiley & Sons . P. Boston. Liao.Handbook of Metaheuristics. 3. Bradford Book The MIT Press. high-performance computing on clouds.-Y." vol. 11.2015. I. 381. allocate the appropriate nodes to users' tasks. 1. Xie. 34-39. 1-6. vol. Ant Colony Optimization.. 1230-1242. De Rose.pp. H. "A multi-objective ant second policy based on ant colony optirnization algorithm to colony system algorithm for virtual machine placement in cloud computing. the auxiliary controller will insure balancing cloud computing: state of art.1996. cIosest load balancer. K Mahanti. Secondly. Pang. no. "A performance estimation model for Taguchi's quality engineering handbook. Gao. Virtualization with Xen: including geographie scope 01' data centers. ASI Consulting Group.pp. and H. Qi. Taguchi.pp. The authors would like to put on record their thanks to the [19] P. Sei. "An In-depth Analysis and anonymous referees for their precious suggestions. Lect. Guan. 2010. [21] B. Wang. "A seeured load balancing architecture for cloud computing based on multiple clusters. 446-452. Chowdhury.MA: Springer US. Eng. EI Bakkali. J. C. Eds. S.no. "SMICloud: A Framework for [22] R. R. 52-58. "CloudAnalyst: REFERENCES A CloudSim-Based Visual Modeller for Analysing Cloud Computing Environments and Applications. Comput. Kumar. 18. pp." Procedia Comput. and L. 1-8. 210-218. S. Yano. Belkhouraf. R. and R. Pract. we choose to combine different algorithms in order [13] Y.2013. 2010 3rd IEEE Cloud into two parts. 50. Larus. vol. Environment. K Vasudevan." 2011. The regional load balancer applies a [15] Y. D. Escalante. 2011. Wickremasinghe. G. K Idrissi. 41. Buyya. and D. E. Chang. Bane�iee. MA: difficult to maintain a satisfactory level 01' availability.pp.pp. Mukherjee. EI Marraki. and R. it becomes more and more XenEnterprise. and S.. Kant. H. : Livonia.Jan. no. Gendreau and J. H. Comparing and Ranking Cloud Services. Z. " [12] A. The [14] G. M. ACKNOWLEDGMENT [18] C." J. vol. Ranjan. 106-109. Versteeg. Potvin. 2. Q. A." in Computer functions 01' the main controller representing the entry point to Science and Information Technology (ICCSIT). no. "Ant colony Optimization: A Solution of Load balancing services in the Cloud. Xu. 23-50. Grance. Given the increasing number 01' users in the Cloud and the [8] D. Boston. A. Springer Science & Business Media. [6] B. pp. Z. and A. approach. pp. -S. Raghuwanshi. "CloudSim: a toolkit for modeling and simulation of cloud [2] Y. Marco Dorigo. Calheiros. function 01' partitioning users' tasks through different regional [11] V.-S. Khiyaita. solutions proposed firstly. and P. Techno!." 2010. 2011. Tripathi and S. C. Distributed Graph Algorithms for Computer Networks. analytical performance models for IEEE 802. vol. Mishra. N. no. the main controller will keep the International Conference on. [3] J. EI Kettani. Sei. Y." Synth. Beloglazov. England: A Conference on.MA: Springer US. 33-50. J. "A load balancing model based on cloud main controller applies an algorithm based on the geographie partitioning for the public cloud.2015. 2015 International Conference on. "Load Balancing in Cloud Computing. Fu. vol. Greenberg. . Hoboken. 2. Berwal and C. no. algorithms. 1056- 1071. VI. 221-224. J.-P. Ouahmane. F. A. Hou."Towards a load architecture for load balancing by offering to share the balancing in a three-Ievel cloud computing network. Gopinath and S. Kartit.Nov. Burlington.. [1] S. "Cloud computing International Conference on.vol. G. Mell and T. H. Williams and J. pp. " balancing. A. security Syngress. "Cloud Task Scheduling in Hybrid and responsiveness through conventional models 01' load Environment using Ant Colony Optimization. 2012 National Days of. Eds. pp.. which have Study of Load Balancing Techniques in the Cloud Computing improved the value 01' the paper over its initial version. and XenExpress. Syst. S. W. 1-116. vol.2013. Wang. Zbakh. we plan to test the in Cloud.. to optirnize the partitioning and mapping steps. 2.2010. [23] G.no. vol. Buyya.

Master your semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master your semester with Scribd & The New York Times

Cancel anytime.