Research Article Scheduling Method of Data-Intensive ...downloads.hindawi.com/journals/mpe/2015/605439.pdf · Research Article Scheduling Method of Data-Intensive Applications in

Research ArticleScheduling Method of Data-Intensive Applications in CloudComputing Environments

Xiong Fu,1 Yeliang Cang,1 Xinxin Zhu,1 and Song Deng2

1School of Computer Science and Technology, Nanjing University of Posts and Telecommunications, Nanjing 210023, China2Institute of Advanced Technology, Nanjing University of Posts and Telecommunications, Nanjing 210023, China

Correspondence should be addressed to Xiong Fu; [email protected]

Received 5 January 2015; Accepted 29 March 2015

Academic Editor: Emilio Insfran

Copyright © 2015 Xiong Fu et al.This is an open access article distributed under the Creative CommonsAttribution License, whichpermits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The virtualization of cloud computing improves the utilization of resources and energy. And a cloud user can deploy his/her ownapplications and related data on a pay-as-you-go basis. The communications between an application and a data storage node,as well as within the application, have a great impact on the execution efficiency of the application. The locations of subtasksof an application and the data that transferred between the subtasks are the main reason why communication delay exists. Thecommunication delay can affect the completion time of the application. In this paper, we take into account the data transmissiontime and communications between subtasks and propose a heuristic optimal virtual machine (VM) placement algorithm. Relatedsimulations demonstrate that this algorithm can reduce the completion time of user tasks and ensure the feasibility and effectivenessof the overall network performance of applications when running in a cloud computing environment.

1. Introduction

Cloud computing has been presented as a brand new sharingnetwork computation model of commercial resources inrecent years. It has also been universally recognized as thethird technology revolution, and it will continue to lead thebusiness revolution in the coming two decades. According tothe survey conducted by market research company Gartnerin 2010, cloud computing has become one of the mostcrucial technologies for IT users. Meanwhile many IT giantshave built large amount of data centers and provide cloudcomputing services for outside users. For example, Googlehas deployed 36 data centers and millions of computationnodes all over the world. The number of Microsoft’s serverswill be doubled in every 14 months, and there are hundredsof thousands of servers in its cloud computing data centers[1].

Currently, cloud computing centers based on the vir-tualization technology have become the most widely usedhosting platforms for composite applications. As such, alarge amount of the communication intensive applicationsand data-intensive applications has emerged. An application

is usually divided into different subtasks. These subtasksare then allocated to virtual machines (VMs). The VMsare finally placed in specific computation nodes. Therefore,the communication traffic among subtasks and the datatransmission rate between a computation node and a storagenode have a great impact on the reaction and executionefficiency of the application. So guaranteeing high networkperformance has become the top issue in a cloud computingsystem and it has attracted a lot of attention [2, 3]. Asthe network topologies of cloud systems may be differentand visualization technology itself affects the communicationdelay in a cloud, the execution of an application could largelydepend on the network performance [4].

There are many types of applications in cloud computingsystems, such as web applications (These applications aredivided into different layers: web layer, application layer, anddata layer, and there are communications between differentlayers.) and distributed applications (e.g., applications relatedto e-commerce or scientific computation, these applicationsare generally divided into multiple subtasks; there are com-putation behaviors within a subtask and data exchangesbetween one subtask and another). Communication capacity

Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2015, Article ID 605439, 8 pageshttp://dx.doi.org/10.1155/2015/605439

2 Mathematical Problems in Engineering

affects the execution time and reaction time of subtasks orapplications, and it is the bottleneck of multitask execution ina cloud system. As there are constraints of physical resources(mainly about CPU andmemory resources), an application isusually divided into multiple subtasks. And the subtasks aredistributed in computation nodes on a large scale. Therefore,the communication capacity among physical devices cangreatly affect the completion time of an application.

In a current cloud computing environment, an applica-tion could be divided into multiple subtasks when deployedinto a cloud computing system. Then these subtasks areassigned to VMs based on the subtasks’ types; differentsubtasks may access different data in data centers, and thereshould be specific data exchanges between subtasks. As thetypes of applications vary, the communication behaviorsamong subtasks and storage nodes are different as well. Forexample, subtasks of a scientific computation application arerunning in a dependency relationship; namely, one certainsubtask cannot be executed unless another specific subtaskis finished. For someMPI applications, subtasks are executedwith data being transmitted between them at the same time.

However, in a particular cloud system, VMs’ configura-tions, such as the processor power, disk sizes, and memorycapacities, are different. Moreover, there should be dataexchanges between different VMs to finish an application,and each VM needs to access files in data centers. Forexample, in a web application, the data layer needs to accessfiles in a database. The data transmission time can greatlyaffect the execution of an application. So it is pretty clear thathow to determine the physical or logical locations of VMsplays an important role in a cloud computing system.

To handle this problem, some of the current VM place-ments focus on the consumption of energy; the placementof assigning VMs to the same physical node, for instance,can limit the number of running physical machines andcut down the energy consumption [5]. Other placementseither focus on the CPU dimension of physical machinesonly [6] or concern the problem from the user’s perspectivelike the VM placement based on SLA [7]. There are alsosome placements in which only physical resources are takeninto account [8, 9]. However, few of them are about thecommunication capacity. The works in [2, 3] only focus onthe communication capacity and design an overall networktopological structure to improve the network performance ofa cloud system, but no concrete VM placements are foundin these two papers. The effect of data transmission ratebetween a computation node and a storage node is discussedin [10], but the communications between VMs are not. Thencommunication delays among VMs are also not discussedin [11] which only focuses on the data transmission betweenVMs and storage nodes. All of the placement algorithmsthat we discussed above are not taking into account thecommunication traffic between VMs, and therefore there ismuch higher completion time of applications.

In this paper, a heuristic VM placement algorithm isproposed. Not only the communications between subtasksbut also the data transmissions between computation nodesand storage nodes are concerned in this algorithm, so it

can effectively shorten the overall completion time of anapplication.

The remainder of this paper is organized as follows. Wecover related work in Section 2. Section 3 presents the cloudsystem model. Then the heuristic VM placement algorithmis presented in Section 4. Next, we show some simulationresults and analysis in four different algorithms in Section 5.Finally, Section 6 presents future work.

2. Related Work

In a cloud system, the subtasks of an application are allo-cated to VMs. And the key problem of dynamic resourcemanagement in a cloud computing system is how to placeand manage VMs effectively. In the infrastructure layer ofcloud computing, the problem of how to place VMs canbe adapted into the classic problem—Bin-Packing problem.And it satisfies the condition that the number of the runningphysical nodes is at its minimum level and resources requiredby VMs in a host should not exceed the host’s capacity at thesame time.

The placement problem of VMs is an NP-hard variantof the N-dimensional Bin-Packing problem which has nopolynomial algorithms for optimal solutions [12]. Manyheuristic or greedy algorithms are introduced to get closeto the global optimal solution of this problem. A lot ofadditional simple rules are also presented on the basis ofthese heuristic algorithms such as suboptimal fit, first fit, andoptimal fit. In addition, many heuristic algorithms use theconventionalmethodwhich features single-point search.Thismethod tends to fall into the local area and thus cannot getthe global optimal but partial solution. In some cases, thismethod cannot get a solution at all. Another way to get theoptimal solution is using a Constraint Programming (CP)engine [8, 9, 13, 14]. EmployingCP to optimizeVMplacementis a convenient technique of elegance and flexibility whichsimplifies the way of getting solutions of the problem. But thequality of constraint conditions directly affects the quality ofthe final solutions obtained by CP.

Current solutions to VMplacement in a cloud computingenvironment can be divided into two categories: one focuseson the optimization of single objective while another focuseson that of multiple objectives. The single objective includesthe minimum number of host nodes [5], ensuring highefficiency of service level [7], reducing VM migration times[15], cutting down the energy consumption of data centers[16], promising high availability for users [7], and decreasingthe use of network I/O in cloud computing systems [17].But defects of the single objective are also evident and someproblems that the single objective deals with above are inconflict with themselves. For example, we can allocate moreVMs to fewer physical hosts and shut down the idle hosts toreduce the energy consumption and management expenses.But this will lead to more VM migration times. On thecontrary, to achieve the minimum migration times, theremust bemore physical hosts. So the strategy ofmultiobjectiveoptimization has arisen [18, 19]. It is an optimization processthat concerns all of these optimal conditions and makes atrade-off among them. Many of the existing VM placements

Mathematical Problems in Engineering 3

of multiobjective optimization phase to solve the VM place-ment problem. They only consider one objective at a time.Few of them can take multiple objectives into account at thesame time.Therefore, they cannot get the globally but locallyoptimal solution most of the time. The work in [18] dividesthe VM placement problem into combinational optimizationproblems. It adopts genetic algorithm in dealing with themultiobjective optimization problem of VM placement andthen optimizesmultiple objectives which includeminimizingthe overall resource waste, energy consumption, and heatdissipation. But it does not take into account the overheadof VMmigrations.

The performance of the solutions to the multiobjectiveoptimization presented above, whatever Bin-Packing solu-tions or multistage solutions, is not good as we may wish inthe aspect of time complexity. All things considered, we canmake an acceptable balance between the time complexity ofthe algorithm and the precision of the result depending onreasonable regulations of a heuristic algorithm.

3. Problem Statement

3.1. Scenario. In a classic cloud computing system, there arelots of flexible data storage nodes, computation nodes, andbrokers, and all of them can communicate with each otherbased on the network topology structure. Cloud users caninteract with each other through the broker.They can transferdata and deploy their own applications. In this case, cloudusers only need to care about theworking state of applicationsand they do not need to care about the allocated locations oftheir own tasks. The scenario is shown in Figure 1.

In this scenario, cloud users request the cloud broker todeploy applications and upload data that they will use; thecloud broker allocates the related subtasks to the correspond-ing VMs in the computation nodes and uploads the data tothe storage nodes.

3.2. Cloud Model. Before deploying applications, we assumethat the related data has been uploaded to the storage nodesin advance. And the bandwidths between storage nodes,computation nodes, and the broker are already known.

As is shown in Section 3.1, there are many computationnodes and data nodes in a cloud computing system.Wedefine𝑆 as a set that includes all of the storage nodes. Let 𝑆

𝑖denote

a certain storage node (𝑆𝑖∈ 𝑆, 1 ≤ 𝑖 ≤ 𝑀).𝑀 is the number

of storage nodes. Let𝐻 denote the set of computation nodes.And𝐻

𝑗is a physical host that belongs to𝐻 (𝐻

𝑗∈ 𝐻, 1 ≤ 𝑗 ≤

𝑁).𝑁 refers to the number of computation nodes.𝐻𝑖cpu refersto the available CPUs of host 𝐻

𝑗and 𝐻𝑗ram is the remaining

memory resource in host𝐻𝑗. The sign𝐻𝑗disk denotes the disk

size of host𝐻𝑗.

The data transmission rates or bandwidths among allthe physical devices in a cloud can be calculated using thefunction Rate (ps, Δ𝑡). ps refers to the size of the package andΔ𝑡 is the package transfer time slot.

UserBroker

Storage nodes

Storage nodes

Storage nodes

Deploy applications

Assign VMsUpload user

data

Computation nodes

Computation nodes

Computation nodes Pu

blic

ne

twor

k

Figure 1:The structure of self-adapting resourcemonitoringmodel.

The matrix CSH represents data transmission ratesbetween storage nodes and computation nodes. Each ele-ment CSH

𝑖,𝑗in CSH denotes the data transmission rate

between computation node𝐻𝑗and storage node 𝑆

𝑖. Similarly,

CHH represents data transmission rates among computationnodes, and CHH

𝑎,𝑏is the data transmission rate between

computation nodes 𝑎 and 𝑏where 1 ≤ 𝑎 ≤ 𝑀 and 1 ≤ 𝑏 ≤ 𝑀.CHH𝑎,𝑏= ∞ when 𝑎 = 𝑏. In other words, we disregard the

data transmission time among VMs (or subtasks) in the samehosts. And the equation CHH

𝑎,𝑏= CHH

𝑏,𝑎means that the

data transmission rate from host 𝑎 to host 𝑏 is the same asthat from host 𝑏 to host 𝑎.

3.3. TaskModel. Thepresent brokerwill divide an applicationinto multiple subtasks when it receives the requests of theapplication. Let 𝐴 denote the subtask set of the presentapplications, and each subtask is allocated to the corre-sponding VM whose resources of CPU, memory, and diskare already known. The VM is then placed on a specificcomputation node. What we should do is to determine thefinal computation node that a subtask is allocated to. V𝐴

𝑙∈ 𝐴

denotes the subtask that belongs to the application 𝐴 where1 ≤ 𝑙 ≤ 𝐿 and 𝐿 denotes the number of subtasks. And thesigns of 𝑅𝑙cpu, 𝑅

𝑙

ram, and 𝑅𝑙

disk denote the processor capacity,memory capacity, and disk size of the subtask V𝐴

𝑙, respectively.

And 𝐹 is the set of files that the application 𝐴 will request. 𝑓𝑟

is the element of 𝐹 (1 ≤ 𝑟 ≤ 𝑅). 𝑅 represents the number offiles in 𝐹. We define 𝑇𝑙exec as the completion time of subtaskV𝐴𝑙.Let 𝑝𝐴 = {𝑝

1, 𝑝2, 𝑝3, . . . , 𝑝

𝑙, . . . , 𝑝

𝐿} denote the distri-

bution path of the subtasks of 𝐴 where 𝑝𝑙represents the

allocated host of subtask V𝐴𝑙(𝑝𝑙∈ 𝐻).


We can define the file storage matrix𝐷 as

𝐷 =

[[[[[[[[[

[

𝑑1,1

𝑑1,2

𝑑1,3

⋅ ⋅ ⋅ 𝑑1,𝑀

𝑑2,1

𝑑2,2

𝑑2,3

⋅ ⋅ ⋅ 𝑑2,𝑀

𝑑3,1

𝑑3,2

𝑑3,3

⋅ ⋅ ⋅ 𝑑3,𝑀

.

.

....

.

.

....

.

.

.

𝑑𝑅,1

𝑑𝑅,2

𝑑𝑅,3

⋅ ⋅ ⋅ 𝑑𝑅,𝑀

]]]]]]]]]

]

. (1)

The element 𝑑𝑟,𝑚

in𝐷 represents the size of the data blockof file 𝑓

𝑟in storage node 𝑠

𝑚(𝑠𝑚∈ 𝑆, 1 ≤ 𝑟 ≤ 𝑅, 1 ≤ 𝑚 ≤ 𝑀).

Let 𝐷𝑙 denote the set of files that subtask V𝐴𝑙will request,

and 𝑑𝑙𝑠is the element in 𝐷𝑙 (𝐷𝑙 ⊆ 𝐹, 1 ≤ 𝑠 ≤ 𝑅). The size of

the file 𝑑𝑙𝑠can be calculated by the function Size (𝑑𝑙

𝑠); namely,

Size (𝑑𝑙𝑠) =

𝑀

∑

𝑚=1

𝑑𝑑𝑙𝑠,𝑚. (2)

Let matrix CS denote the communication data sizesamong all the subtasks that belong to 𝐴; namely,

CS =

[[[[[[[[[

[

0 cs1,2

cs1,3

⋅ ⋅ ⋅ cs1,𝐿

cs2,1

0 cs2,3

⋅ ⋅ ⋅ cs2,𝐿

cs3,1

cs3,2

0 ⋅ ⋅ ⋅ cs3,𝐿

.

.

....

.

.

. cs𝑥,𝑦

.

.

.

cs𝐿,1

cs𝐿,2

cs𝐿,3

⋅ ⋅ ⋅ 0

]]]]]]]]]

]

. (3)

Element cs𝑥,𝑦

represents the size of data that needs to betransferred between subtask 𝑥 and subtask 𝑦, and cs

𝑥,𝑦= 0

when 𝑥 = 𝑦, 1 ≤ 𝑥, 𝑦 ≤ 𝐿. The equation cs𝑥,𝑦= cs𝑦,𝑥

denotesthat the size of data that needs to be transferred mutuallybetween the two subtasks is the same.

Then we can define the total data transmission timebetween the subtask V𝐴

𝑙and all the related data files as

𝑇file = ∑

𝑑𝑙𝑠∈𝐷𝑙𝑠

𝑀

∑

𝑚=1

𝑑𝑑𝑙𝑠,𝑚

CSH𝑚,𝑝𝑙

,

s.t. 1 ≤ 𝑗 ≤ 𝑁, 𝑅𝑙

cpu ≤ 𝐻𝑗

cpu, 𝑅𝑙

ram ≤ 𝐻𝑗

ram,

𝑅𝑙

disk ≤ 𝐻𝑗

disk.

(4)

And the total communication time between subtask V𝐴𝑙

and other subtasks that belong to 𝐴 can be denoted as

𝑇task =𝐿

∑

𝑦=1

cs𝑙,𝑦

CHH𝑝𝑙,𝑝𝑦

,

s.t. 𝑝𝑦, 𝑝𝑙∈ 𝑃𝐴, 𝑅𝑙

cpu ≤ 𝐻𝑗

cpu, 𝑅𝑙

ram ≤ 𝐻𝑗

ram,

𝑅𝑙

disk ≤ 𝐻𝑗

disk.

(5)

3.4. Problem Definition. The problem of an applicationdeployment can be adapted to the problem of reducing

completion time of all the subtasks as much as possible. 𝑇𝐴denotes the overall completion time of application𝐴. We canshorten the overall completion time based on the definitionsof subtasks and the cloud model that we discussed aboveand finally get the allocation path 𝑃𝐴. The subtasks are thenallocated to computation nodes based on 𝑃𝐴.

We can get the completion time of all subtasks that belongto 𝐴 using formulas (4) and (5); namely,

𝑇𝐴=

𝐿

∑

𝑙=1

( ∑


𝑀

∑

𝑚=1


CSH𝑚,𝑝𝑙

+

𝐿

∑

𝑦=1

cs𝑙,𝑦

CHH𝑝𝑙,𝑝𝑦

+ 𝑇𝑙

exec) (6)

s.t.

1 ≤ 𝑖 ≤ 𝑀, (7)

1 ≤ 𝑙 ≤ 𝐿, (8)

1 ≤ 𝑦 ≤ 𝐿, (9)

1 ≤ 𝑚 ≤ 𝑀, (10)

∑

𝑝𝑙=ℎ𝑖

𝑅𝑙

cpu ≤ 𝐻ℎ𝑖

cpu, ∀𝑖, 𝑙, (11)

∑

𝑝𝑙=ℎ𝑖

𝑅𝑙

ram ≤ 𝐻ℎ𝑖

ram, ∀𝑖, 𝑙, (12)

∑

𝑝𝑙=ℎ𝑖

𝑅𝑙

disk ≤ 𝐻ℎ𝑖

disk, ∀𝑖, 𝑙, (13)

𝑙 ≤ 𝑦, (14)

𝑝𝑙∈ 𝐻. (15)

Condition (11), (12), and (13) mean that the sum of thephysical resources requested by all subtasks in computationnode ℎ

𝑖does not exceed the remaining resource capacity,

such as processor capacity, memory capacity, and disk size.And condition (14) means that the size of data transferredmutually between the two subtasks is the same, so we onlyneed to calculate a single transmission time.The problem thatwe ought to solve is how to get the allocation path 𝑃𝐴 so thatwe can minimize the overall completion time 𝑇𝐴.

4. Placement Algorithm of Subtasks

The key to solve the optimization problem presented informula (6) is to settle the placement problem of everysubtask so that we can get the final path 𝑃𝐴 of the application𝐴. We can use a heuristic placement algorithm of subtasks toget the overall optimal placement of the application and thisalgorithm is called HRVP for short.

Step 1. Initialize the computation node setH, storage node setS, and data transmission rate matrices CHH and CSH basedon the characteristic of the cloud system.

Step 2. Initialize the subtask set 𝐴, including computationcapacity 𝑅𝑙cpu, memory capacity 𝑅𝑙ram, and disk size 𝑅𝑙disk of


each subtask. Finally initialize the file set 𝐷𝑙 and communi-cation data matrix CS.

Step 3. We can initialize the file storage matrix 𝐷 by travers-ing all of the storage nodes in 𝑆 and the file information ofeach storage node.

Step 4. Let the cyclic variable Count = 0 denote the numberof subtasks that have been allocated. And one more variable 𝑙is defined.

Step 5. Initialize the variable 𝑙 = Count + 1 and traverse theset𝐻 to find a host𝐻

𝑖∈ 𝐻 satisfying the formula

Min( ∑


𝑀

∑

𝑚=1


CSH𝑚,ℎ𝑖

+

Count∑

𝑦=1

cs𝑙,𝑦

CHHℎ𝑖 ,𝑝𝑦

+ 𝑇𝑙

exec)&&

𝑅𝑙

cpu ≤ 𝐻𝑖

cpu&&𝑅𝑙

ram ≤ 𝐻𝑖

ram&&𝑅𝑙

disk ≤ 𝐻𝑖

disk.

(16)

Step 6. Calculate formula (16), get 𝐻𝑖, and save its value;

namely,𝑝Count = 𝐻𝑙. Update the remaining physical resourcesof related computation nodes using the following formulas:

𝐻𝑖

cpu = 𝐻𝑖

cpu − 𝑅𝑙

cpu, 𝐻𝑖

ram = 𝐻𝑖

ram − 𝑅𝑙

ram,

𝐻𝑖

disk = 𝐻𝑖

disk − 𝑅𝑙

disk.

(17)

Step 7. Update the variable Count = Count+1. If Count = 𝐿,go to the next step or else go back to Step 5.

Step 8. Get the final allocation path 𝑃𝐴.

Complexity Analysis. The time complexity of initializing thematrices 𝐷 and CS is 𝑂(𝑀 ⋅ 𝑅) and 𝑂(𝐿2), respectively. Sothe time complexity of the heuristic algorithm in Step 3 is notmore than max{𝑂(𝑀 ⋅ 𝑅), 𝑂(𝐿

2)}.

5. Performance Evaluation

Themodel we presented is based on the IaaS cloud computingsystem which supplies cloud users with visually endlessresources. The placement of the VM allocated to a subtaskplays an important role in the system. It is a great challengeto implement and experiment an algorithm repeatedly in alarge-scale computing infrastructure. We chose the simula-tion toolkit CloudSim [20, 21] as our experimental platform.This toolkit can simulate not only a variety of cloud physicalresources and user tasks but also network topology of thewhole cloud, and it plays an important role in our experiment.

To show the superiority of the proposed algorithm,we compare it with another two already known placementalgorithms of VMs. One of the algorithms is energy efficientplacement of VMs [5] whose objective is to minimize thenumber of running physical hosts and increase resourceutilization, namely, allocating all of the VMs to the runningand qualified hosts and switching off the idle hosts as manyas possible. We call this algorithm MRP for short. Another

VM placement policy of CloudSim, known as VMSimpleAl-locationPolicy or simply as SAP, allocates the VM to the leastutilized host. And all of the results obtained by using the threealgorithms above are compared with the optimal solutionobtained by using integer linear programming approachwhich is called ILP for short.

Physical Resources.There are 10 data storage nodes.The capac-ity of each node is 1 T and the average data transmission ratebetween storage nodes and computation nodes is 100Mb/s.There are 15 computation nodes, and the configuration ofeach computation node is 4 CPUs, 4GBmemory, 100GB disksize. Each computation node uses Xen as the visualizationplatform [22]. The average data transmission rate betweencomputation nodes is 100Mb/s.

We tested two applications 𝐴 and 𝐵 in our experi-ment. Application 𝐴 is the Workflow App provided by theCloudSim framework and it includes three subtasks. Twoof them send data package to the other subtask and eachsubtask has its own computation phase. There are two datafiles and the sizes are 500MB, respectively. The values ofthe three subtasks [𝑅𝑙cpu, 𝑅

𝑙

ram, 𝑅𝑙

disk] of application 𝐴 are[1, 1024, 5120], [1, 2048, 2048], and [2, 1024, 5120], respec-tively. The corresponding units of the three parameters arethe number of CPUs, MB, and MB.

To prove the performance and efficiency of the pro-posed algorithm in various environments, we assess theperformance of each application from two aspects: changethe communication traffic between subtasks and files grad-ually under the condition that the communication trafficbetween subtasks keeps stable; change the communicationtraffic between subtasks gradually under the condition thatthe communication traffic between subtasks and files keepsstable.

In Figure 2, we assume that the overall communicationtraffic between the three subtasks of application 𝐴 andfiles is fixed to 2GB. And when we change the overallcommunication traffic between the three subtasks gradually,we can see the final complete time of the application afterconducting the algorithms we presented above.

The optimal solution is obtained by using integer linearprogramming approach which is called ILP for short. FromFigure 2, we can see that HRVP has the best performancecompared with MRP and SAP.

In Figure 3, the overall communication traffic betweenthe three subtasks of application 𝐴 is 2 GB. The communi-cation traffic between the three subtasks and files is changinggradually.

From Figures 2 and 3, the HRVP has the performancemost close to that of ILP comparedwith other two algorithms.In Figure 2,HRVPhas an overwhelming advantage overMRPand SAP.MRP can improve physical resources utilization, butit has longer completion time compared with SAP andHRVP.On the contrary, the application completion time of SAP isclose to HRVP in Figure 3. But SAP allocates VMs to the leastutilized hosts, and therefore it will cause low utilization ofphysical resources.


2025303540455055606570

2 3 4 5 6

Com

plet

ion

time (

s)

Overall communication traffic (GB)

MRPSAP

HRVPILP

Figure 2: The completion time of application 𝐴 when the commu-nication traffic between subtasks and files keeps stable.

20253035404550556065

2 3 4 5 6

Com

plet

ion

time (

s)


MRPSAP

HRVPILP

Figure 3: The completion time of application 𝐴 when the commu-nication traffic between subtasks keeps stable.

Application 𝐵 is an extensional application. The descrip-tion is as follows: the application in this experimentincludes 6 subtasks, and the physical resource value of eachsubtask is [1, 2048, 2048], [2, 2048, 4096], [1, 1024, 5120],[1, 2048, 2048], [2, 1024, 5120], and [2, 2048, 4096] where thecorresponding units of the three parameters are the numberof CPUs, MB, and MB.

The experiment is conducted in five groups (𝑎𝑏𝑐𝑑𝑒). Thecommunication traffic between all the subtasks and files ineach group is unchangeable while the data communicationtraffic between VMs is not. And the relationship of thecommunication traffic between VMs in the five groups is𝑎 < 𝑏 < 𝑐 < 𝑑 < 𝑒. The related five data files are distributedon the ten storage nodes in advance. And the sizes of thefive data files are 657MB, 350MB, 500MB, 400MB, and750MB, respectively. For each subtask, its computation phaseis included.

Assuming that the overall communication traffic betweensubtasks of 𝐵 and files is fixed to 4GB, the applicationcompletion time of 𝐵 is shown in Figure 4. When the overallcommunication traffic between the 6 subtasks is changing

0

50

100

150

200

250

2 4 10 12 16

Com

plet

ion

time (

s)


MRPSAP

HRVPILP

Figure 4: The completion time of application 𝐵 when the commu-nication traffic between subtasks and files keeps stable.

0

50

100

150

200

250

2 4 10 12 16

Com

plet

ion

time (

s)


MRPSAP

HRVPILP

Figure 5: The completion time of application 𝐵 when communica-tion traffic between subtasks keeps stable.

gradually, we can see the final complete time of the applica-tion after conducting the algorithms we presented above.

Figure 4 shows a trend that the more the overall com-munication traffic increases, the closer the execution resultof HRVP gets to that of the optimal solution (ILP). Butthe performance of MRP and SAP is getting worse. That isbecause MRP and SAP are not bandwidth related, and theycannot adapt to the bandwidth change.

In Figure 5, the overall communication traffic betweensubtasks of application 𝐵 is fixed to 4GB. Change the com-munication traffic between the subtasks and files gradually,and the result is shown as in Figure 5.

From the descriptions of the figures we learned that theHRVP algorithm can ensure the ideal complete time of anapplication in each case. The complete time is close to theoptimal level. In Figures 2, 3, 4, and 5, because the bandwidthsbetween computation nodes and storage nodes are different,the performances of MRP and SAP are unstable. But HRVPkeeps a better property over the previous two algorithms.That is for the reason that HRVP is bandwidth efficient


and allocates VMs to computation nodes in consideration ofthe bandwidths between physical equipments like hosts orstorage nodes.

6. Algorithm Analysis

In the beginning of the proposed algorithm,matrixes like CS,𝐷, CSH, and CHH should be initialized.The time complexityof initializing thesematrixes is𝑂(𝐿2),𝑂(𝑀⋅𝑅),𝑂(𝑀⋅𝑁), and𝑂(𝑁 ⋅𝑁), respectively. Among these matrixes, CS and𝐷 willchange according to a specific application. Because differentapplications will have different communication data sizesbetween their subtasks these applications may also accessdifferent files. So themain initialization part of this algorithmis about initializing matrices CS and 𝐷. When the numberof subtasks of an application increases, the communicationmatrix will become more complex and more files will beaccessed. Therefore, the complexity of the initialization timeand execution time will increase.

To solve this problem,we can divide files into bigger parts,and the number of the subfileswill decrease. So the expression𝑂(𝑀⋅𝑅)will get a smaller value. However, there is not a betterway to simplify the process of initializing matrix CS becauseof the increasing number of subtasks. To compensate thisweakness, the elements in CS can be sorted so that they canbe accessedmore quickly when used in Step 5 of the proposedalgorithm.

7. Conclusions

Aheuristic algorithm that targets the task placement problemof data-intensive applications in a cloud computing systemis proposed in this paper. Not only the data transmissiontime between subtasks and storage nodes but also thecommunication traffic between the subtasks is taken intoaccount. It can obtain shorter completion time of applicationscompared with other several algorithms. And the applicationcompletion time of the proposed algorithm is pretty closeto that of the optimal solution obtained by using linearprogramming approach.

Although the heuristic algorithmwe proposed can reducethe completion time of an application, there remains aproblem that needs to be solved, namely, the increasingcomplexity of the initialization time and execution timebecause of the excessive number of subtasks. What we willdo next is to simplify the initialization process and cut downthe dimension of subtasks.Therefore the execution efficiencyand accuracy of the algorithm can be improved.

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper.

Acknowledgments

This work is sponsored by the National Science Foundationof China (nos. 61202354 and 61272422) and technological

innovation fund for technology-based enterprises of JiangsuProvince in Jiangsu (BC2014195).

References

[1] C. Guo, H. Wu, K. Tan, L. Shi, Y. Zhang, and S. Lu, “DCell: ascalable and fault-tolerant network structure for data centers,”ACMSIGCOMMComputer Communication Review, vol. 38, no.4, pp. 75–86, 2008.

[2] A. Greenberg, J. Hamilton, and D. Maltz, “The cost of a cloud:research problems in data center networks,” ACM SIGCOMMComputer Communication Review, vol. 39, no. 1, pp. 68–73,2009.

[3] S. Ijaz, E. U. Munir, W. Anwar, andW. Nasir, “Efficient schedul-ing strategy for task graphs in heterogeneous computingenvironment,” The International Arab Journal of InformationTechnology, vol. 10, no. 5, pp. 75–86, 2013.

[4] G. Wang and T. S. E. Ng, “The impact of virtualizationon network performance of Amazon EC2 Data Center,” inProceedings of the IEEE INFOCOM, pp. 1–9, IEEE, San Diego,Calif, USA, March 2010.

[5] A. Beloglazov and R. Buyya, “Energy efficient allocation ofvirtual machines in cloud data centers,” in Proceedings of the10th IEEE/ACM International Symposium on Cluster, Cloud, andGrid Computing (CCGrid ’10), pp. 577–578, May 2010.

[6] B. Urgaonkar, A. L. Rosenberg, and P. Shenoy, “Applicationplacement on a cluster of servers,” International Journal ofFoundations of Computer Science, vol. 18, no. 5, pp. 1023–1041,2007.

[7] D. Breitgand and A. Epstein, “SLA-aware placement of multi-virtual machine elastic services in compute clouds,” in Proceed-ings of the 12th IFIP/IEEE International Symposiumon IntegratedNetwork Management (IM ’11), pp. 161–168, May 2011.

[8] C. Tang, M. Steinder, M. Spreitzer, and G. Pacifici, “A scalableapplication placement controller for enterprise data centers,” inProceedings of the 16th International World Wide Web Confer-ence (WWW ’07), pp. 331–340, May 2007.

[9] H. N. Van, F. D. Tran, and J.-M. Menaud, “Autonomic virtualresource management for service hosting platforms,” in Pro-ceedings of the ICSE Workshowp on Software Engineering Chal-lenges of Cloud Computing (CLOUD ’09), pp. 1–8, Vancouver,Canada, May 2009.

[10] J. T. Piao and J. Yan, “A network-aware virtual machineplacement andmigration approach in cloud computing,” inPro-ceedings of the 9th International Conference on Grid and CloudComputing (GCC '10), pp. 87–92, Nanjing, China, November2010.

[11] K. Zamanifar, N. Nasri, and M.-H. Nadimi-Shahraki, “Data-aware virtual machine placement and rate allocation in cloudenvironment,” in Proceedings of the 2nd International Confer-ence on Advanced Computing and Communication Technologies(ACCT ’12), pp. 357–360, January 2012.

[12] X. Zhu, D. Young, B. J.Watson et al., “1000 islands: an integratedapproach to resource management for virtualized data centers,”Cluster Computing, vol. 12, no. 1, pp. 45–57, 2009.

[13] R. Dechter, Constraint Processing, Morgan Kaufmann, 2003.[14] S. C. Brailsford, C. N. Potts, and B. M. Smith, “Constraint

satisfaction problems: algorithms and applications,” EuropeanJournal of Operational Research, vol. 119, no. 3, pp. 557–581, 1999.

[15] T. S. Kang, M. Tsugawa, J. Fortes, and T. Hirofuchi, “Reducingthemigration times ofmultiple VMs onWANs using a feedback


controller,” in Proceedings of the IEEE 27th International ParallelandDistributed Processing SymposiumWorkshops&PhDForum(IPDPSW ’13), pp. 1480–1489, Cambridge, Mass, USA, May2013.

[16] J. Yuan, X. Jiang, L. Zhong, and H. Yu, “Energy aware resourcescheduling algorithm for data center using reinforcementlearning,” in Proceedings of the 5th International Conference onIntelligent Computation Technology and Automation (ICICTA’12), pp. 435–438, January 2012.

[17] K. Sato, H. Sato, and S.Matsuoka, “Amodel-based algorithm foroptimizing I/O intensive applications in clouds using vm-basedmigration,” in Proceedings of the 9th IEEE/ACM InternationalSymposium on Cluster Computing and the Grid (CCGRID ’09),pp. 466–471, May 2009.

[18] J. Xu and J. A. B. Fortes, “Multi-objective virtual machine place-ment in virtualized data center environments,” in Proceedingsof the IEEE/ACM International Conference on Green Computingand Communications, pp. 179–188, December 2010.

[19] S. Wang, H. Gu, and G. Wu, “A new approach to multi-objective virtual machine placement in virtualized data center,”in Proceedings of the IEEE 8th International Conference onNetworking, Architecture and Storage (NAS '13), pp. 331–335,Xi’an, China, July 2013.

[20] S. K. Garg and R. Buyya, “NetworkCloudSim: modelling par-allel applications in cloud simulations,” in Proceedings of the4th IEEE/ACM International Conference on Utility and CloudComputing (UCC ’11), pp. 105–113, December 2011.

[21] R. N. Calheiros, R. Ranjan, A. Beloglazov, C. A. F. de Rose, andR. Buyya, “CloudSim: a toolkit for modeling and simulationof cloud computing environments and evaluation of resourceprovisioning algorithms,” Software: Practice and Experience, vol.41, no. 1, pp. 23–50, 2011.

[22] P. Barham, B. Dragovic, K. Fraser et al., “Xen and the art ofvirtualization,” in Proceedings of the 19th ACM Symposium onOperating Systems Principles (SOSP ’03), pp. 164–177, October2003.

Submit your manuscripts athttp://www.hindawi.com

Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttp://www.hindawi.com

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation http://www.hindawi.com Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Discrete MathematicsJournal of

Hindawi Publishing Corporationhttp://www.hindawi.com

Volume 2014 Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014

Stochastic AnalysisInternational Journal of

Documents

Research Article Scheduling Method of Data-Intensive ...downloads.hindawi.com/journals/mpe/2015/605439.pdf · Research Article Scheduling Method of Data-Intensive Applications in