On Optimizing Resource Allocation and Application Placement Costs in

Cloud Systems

Cihan Sec¸inti and Tolga Ovatman

Department of Computer Engineering, Istanbul Technical University, Maslak, Istanbul, Turkey

Keywords:

Virtual Machine Resource Allocation, Application Live-migration, Utilization Optimization.

Abstract:

Resource utilization problem has been widely studied for cloud systems where a number of virtualized re-

sources are shared among applications hosted as services. Resource utilization optimization can be confronted

at two levels: allocating resources to virtual machines(VM) across physical machines or assigning applications

to virtual machines present in the host. With the improving capabilities on virtualization technologies, realiz-

ing resource allocation at both levels is becoming more viable using VM reconﬁguration and live-migration

of applications. In this paper we investigate applying resource allocation optimization at these two levels and

the emerging trade-off in deciding the appropriate technique to be used. We ﬁrst analyze the effect of gradu-

ally increasing the amount of resources assigned to a virtual machine using VM reconﬁguration and compare

our results with fully assigning host’s resources without reconﬁguration. Later, we investigate the amount of

utilization revenue when application live-migration is used for applications having smaller/larger performance

needs. Finally, we compare the host utilization for different amounts of cost rates between live-migration

and reconﬁguration. Consequently our analysis results identify the cost rate and application granularity levels

where it is optimal to apply live-migration or VM reconﬁguration.

1 INTRODUCTION

The tremendous increase in employment of cloud sys-

tems through the recent years triggered the evolution

of software towards cloud services. As the technolog-

ical advancement continue improving the capabilities

of virtualization and application hosting, it became vi-

able to develop more efﬁcient application placement

strategies that can be used in provisioning cloud sys-

tems. Latest developments in virtual machine(VM)

technology and application live-migration techniques

added new dimensions to the problem of managing

hosts’ resources effectively.

Virtual machine(VM) reconﬁguration can be de-

scribed as the process of changing the amount of re-

sources(e.g. CPU, memory, bandwidth, storage) as-

signed to the VM by the host(Verma et al., 2010).

VM reconﬁguration can simply be performed by shut-

ting down the VM and tuning the amount of resources

allocated for the VM. However some resources, like

bandwidth and storage, are reconﬁgurable without

shutting down VM and additionally recent advances

are enabling “hot add” features for CPU and memory

reconﬁguration. This situation reveals on-the-ﬂy re-

source allocation problem between the host and VM

layers in a cloud system.

On the other hand, application live-migration is

another development supported by most of the VM

managers which enables moving a running applica-

tion from a virtual machine to another without discon-

necting the client or application(Clark et al., 2005).

Application live-migration includes warm-up, copy-

ing and memory migration phases and requires a

substantial amount of VM down-time during live-

migration process. Being able to migrate applications

with different levels of SLA requirements also reveals

the problem of resource allocation by moving appli-

cations between VMs to utilize the host more effec-

tively.

Although the resource management problems de-

scribed above are encountered at different layers

in the cloud system, both situations directly effect

the resource utilization of the host. Moreover they

both have their own costs(e.g., downtime, processing)

which may change according to the technology and/or

VM type being used.

In this paper, we investigated the resource alloca-

tion problem in both layers and reveal the situations

where it is meaningful to apply VM reconﬁguration

or application live-migration. For VM reconﬁgura-

535

Seçinti C. and Ovatman T..

On Optimizing Resource Allocation and Application Placement Costs in Cloud Systems.

DOI: 10.5220/0004849605350542

In Proceedings of the 4th International Conference on Cloud Computing and Services Science (CLOSER-2014), pages 535-542

ISBN: 978-989-758-019-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

tion we compared the case where available resources

are assigned equally among the available VMs and

the resources needed by a VM is gradually assigned

to the VM as the applications continue to be deployed.

We also investigated the effect of amount of initial re-

sources assigned to VMs over the amount of ﬁnal uti-

lization of the host. We found out an initial resource

allocation threshold can be identiﬁed where the ﬁnal

utilization starts to decrease as the initial resources as-

signed to each of the VMs start to increase.

We also analyzed the effect of application live-

migration over the ﬁnal utilization of the host with

respect to the granularity of the applications being de-

ployed to the VMs. We used round robin placement

and mixed integer linear programming(MILP) in de-

ciding the placement of applications having resource

requirements inside certain intervals. This way we

analyze if the effect of live-migrating smaller/bigger

applications make any difference in utilization of the

host.

Figure 1 summarizes the approach being used

throughout the paper. The provisioning engine is re-

sponsible of resource allocation to virtual machines

and carrying on application placement decisions. As

increasing number of applications are being submit-

ted to the host machine provisioning engine may mi-

grate applications across VMs to generate additional

resources. We use MILP to check for a possible ap-

plcation migration scheme that can discover neces-

sary amount of resources. If this solution doesn’t

work additional resources can be assigned from the

host to VMs if present. Our further analysis focus on

making the decision of application migration versus

ressource addition to VMs by reconﬁguration. The

tradeoff is analyzed regarding different rates of costs

between migration and reconﬁguration processes.

As mentioned above both of the resource allo-

cation activities(migration and reconﬁguration) de-

scribed above have their own costs when applied. In

our last stage of experiments we investigated the ef-

fect of different cost rates between these activities and

discussed the thresholds when it becomes meaningful

to apply each of these activities. Our results reveal

the certain situations that can be used in provisioning

process during the decision of using VM reconﬁgura-

tion or live-migration technologies. Also by analyz-

ing the trade-off between applying these technologies

under certain cost rates we were able to explore condi-

tions that can aid in deciding in which layer(host-VM,

VM-application) it is meaningful to apply on-the-ﬂy

resource allocation.

In the next section, we continue by summarizing

the related work in the literature. In Section 3 we ex-

plain the trade-off between VM resource allocation

versus application live-migration in more detail and

present the MILP model used in decision of applica-

tion migrations in more detail. Section 4 presents the

details of the experiments performed and a discussion

of the results. Finally in Section 5 we conclude the

study and present possible future directions.

2 RELATED WORK

Resource allocation is an important concept on sys-

tems that exhibits frequent application deployment.

Nowadays, with development of cloud systems re-

source allocation regain importance on the problem of

increasing utilization of the servers that are being run

on the cloud infrastructure. Even before the advances

in cloud computing resource allocation is vastly stud-

ied for clusters and data centers(Bennani and Menasc,

2005; Zhu and Singhal, 2001).

Migration algorithms can be applied to increase

utilization of systems having a diverse range of char-

acteristics. There were several studies based on in-

creasing live-migration performance. In a recent

study(Kikuchi and Matsumoto, 2012), live-migration

is considered to be a serious factor in causing the

response time degradations for applications. Thus,

live-migration operations should be controlled to min-

imize the effect on the performance of applications

running on the cloud system. In the study, the impact

of transmission control protocol’s behavior on live-

migration is experimented per se to estimate the best

live-migration strategy to be used. In our study in-

stead on performing complicatedestimation computa-

tions we have focused on detecting thresholds based

on relative ratios which allows to perform decisions

before the deployment process in a simple way.

Following the improvements of the cloud systems,

live-migration is being used to increase utilization of

the virtual machines on the cloud infrastructure and

different algorithms are developed to increase usabil-

ity of live-migration method. Yang(Yang et al., 2011)

considers the different VM(and hence application)

migration strategies adopted by the host machines that

have different load states by taking into account four

different resource types. The migration is done con-

sidering the amount of resource allocated to each VM,

however the amount of resource utilization has not

been taken into account. In a more recent study(Wang

et al., 2013), migration is performed regarding the

queue model of the time of application deployment

requests from the clients. A centralized control man-

agement mechanism has been deﬁned to inform the

client which server is available at a time. However,

Wang et al. assumed each request to be asigned to a

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

536

Figure 1: Overall MILP based provisioning process.

new virtual machine which ignores multiple applica-

tions deployed to a common virtual machine.

On the other hand, cloud reconﬁguration research

is mostly based on physical machine reconﬁguration.

In this particular area, the challange is more focused

on ﬁguring out how to place virtual machines to phys-

ical machines. In a recent study by He et al. it

is claimed that utilization of the physical machines

would be increased when the reconﬁguration is ap-

plied to physical machines(He et al., 2012). In their

study they focused on virtual machines placement on

physical hosts rather than taking into account applica-

tion placement as well.

There are a few studies which examine the trade-

offs between reconﬁguration and migration technolo-

gies. A study by Chen et al. proposes a framework

that uses live-migration to reduce cost of resource re-

conﬁguration(Chen et al., 2012). They conduct ex-

periments to illustrate the performance of the frame-

work in maximizing utilization and reducing cost of

the runtime reconﬁguration. Also they propose a two-

level runtime reconﬁguration strategy to reduce vir-

tual machine live-migration and shorten the total live-

migration time. Earlier studies by Verla et al. ex-

amine the performance impact of reconﬁguration in

cloud systems and build a model to predict the dura-

tion and performance effects of such activities (Verma

et al., 2010; Verma et al., 2011).

In our study we also focus on maximizing perfor-

mance by considering the effect of migration and re-

conﬁguration under different VM and application re-

source consumption characteristics. We have adopted

the strategy to determine resource requirements of ap-

plications in a random way limited inside certain in-

tervals. In the study by Hwang et al. it is considered

that the exact resource requirements of any applica-

tion could not be found easily(Hwang and Pedram,

2012). Therefore, calculating resource requirements

in a randomized way is not a trivial approach. They

also argue that by determining application resources

randomly, error calculation can be done depending on

standard deviation of randomized variables. We adopt

this strategy by limiting our random resource deter-

mination process under certain limits which we use

to represent the characteristics of the application size.

This way we were able to fuzzily quantize the appli-

cation sizes in our experiments.

As a ﬁnal case to mention, we have used MILP

techniques in deciding the optimal placement of ap-

plication among VMs in our experiments. Utilization

of ILP techniques is a known approach in resource op-

timization of cloud systems. A study by Borovskiy et

al. treats the workload distribution issue as a set par-

titioning problem and devise a formal approach based

on integer programming techniques(Borovskiy et al.,

2011). In an earlier study a similar ILP based ap-

proach was applied on workload scheduling (Van den

Bossche et al., 2010). Later, Alvarez et al. applied

ILP techniques in data allocation problem by taking

into account stroage and computational costs (Ruiz-

Alvarez and Alvarez, 2012). Although we haven’t

dig into a single type of resource allocation problem

deeply, we have built a simple model regarding four

different types of resources altogether.

OnOptimizingResourceAllocationandApplicationPlacementCostsinCloudSystems

537

3 RESOURCE ALLOCATION

AND APPLICATION

LIVE-MIGRATION

An obvious problem in assigning a host’s resources

to a number of VMs is fragmentation caused by vary-

ing resource requirements of applications that are go-

ing to be placed in VMs. A straightforward strat-

egy might be allocating all the resources present in

the host equally among the VMs once and for all

which requires no more reconﬁguration in the future.

A more sensible approach can be assigning a certain

amount of host resources equally to the VMs and allo-

cating remaining resources in appropriate amounts as

the VMs require more resources in order to accept in-

coming applications. However this solution requires

a reconﬁguration cost, which may include shutting

down a VM for a certain amount of time.

Another solution for more effective resource al-

location is to use application live-migration. In this

case, a reallocation of the present applications resid-

ing in VMs needs to be computed which contains two

distinct challenges. The ﬁrst challenge is leveraging

the cost of VM down-time during the migration pro-

cess where the second one is exploring a new place-

ment for applications that contain minimal number of

application migrations. Live migration of an applica-

tion can be achieved in two different ways. Each VM

in the system can be afﬁliated with a single applica-

tion so that the application migration is achieved by

migrating VMs across hosts. Another possibility is to

redeploy the application to another VM which also-

makes it possible for a VM to host multiple applica-

tions. Our analysis consider the second case however

the ﬁrst case can also be supported by minor mod-

iﬁcations by slightly transforming the problem from

application migration among VMs to VM migration

among hosts.

It can be crucial to decide which strategy to be

used in resource allocation by considering different

aspects of the system. These aspects include re-

source needs of the applications at hand as well as the

cost rate between VM reconﬁguration and application

live-migration. In our experiments we use cost ratios

in detecting the optimal results since each type of cost

can change over time and VM type being used in the

host system. By using proﬁling techniques the cost

rate can be determined and appropriate strategy can

be chosen accordingly. Moreover a priori knowledge

on the resource consumption behavior of the applica-

tions can be used in deciding any of those resource al-

location approaches as well. This a priori knowledge

do not need to be very detailed; identifying applica-

tions as heavily resource consuming or lighter should

be enough.

Before continuing to experiments performed to

analyze the mentioned trade-offs we would like to ex-

plain an important step that takes a crucial role in de-

termining migration costs. As mentioned above, se-

lecting the migrating applications in determining a

better placement for applications that reside in the

host plays a key role in determining live-migration

costs. The determined placement should include min-

imum number of application migrations in order to

achieve a reordering process with minimal cost. In

our experiments we mainly focus on cost rates be-

tween migration and reconﬁguration, thus we have

chosen to use symbolic costs for each of the concepts

and perform our analysis based on different cost rates

instead of speciﬁc cost values.

In order to determine an optimal application

placement scheme we have used MILP techniques.

By using MILP we minimize the number of migra-

tions when determining the place of each application

to the pool of VMs. In achieving this, for each solu-

tion we compare the place(VM) of each application

with its former place and try to maximize the num-

ber of applications that doesn’t change their place.

To perform this, we provide MILP with Equation 1

to maximize. In Equation 1 i index is used to select

among VMs and j index is used to select among ap-

plications. Since an application is either assigned or

not to a VM we used assignment matrices of boolean

variables. We try to maximize the sum of old and new

application-VM assignment products which produces

a value of 1 if the assignment didn’t change and 0 oth-

erwise if a migration is present.

#VMs

∑

i=0

#Apps

∑

k=0

(newAssign[i][k] × oldAssign[i][k]) (1)

In addition we also do apply a number of con-

straints in order to drive the MILP to produce mean-

ingful results. In Equation 2 we guarantee that each

application has been assigned to only one VM. Addi-

tionally in Equation 3 resource needs of each applica-

tion that has been assigned to a speciﬁc VM for each

type of resources are summed up to be less than the

available resource assigned to the VM.

#VMs

i=0

#Apps

∑

k=0

newAssign[i][k] = 1

(2)

#VMs

i=0

#Res

j=0

#Apps

∑

k=0

newAsgn[i][k]× resNeed[ j][k]

≤ resAv[i][ j]

(3)

In our experiments we have used four different

types of resources which are CPU, memory, band-

width and storage requirements. We also use a total

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

538

of four VMs which greatly reduces the computational

complexity of the MILP. A more detailed discussion

of the experimental environment is presented in the

following section.

4 EXPERIMENTS ON

OPTIMIZING PROVISIONING

TRADEOFF

4.1 Experimental Setup

Our experiments on VM reconﬁguration and appli-

cation live-migration initially focus on separate as-

pects about applying resource allocation using each

technique respectively. Initially, we focus on the dif-

ference in the host utilization when application mi-

gration is used for various application resource con-

sumption characteristics. Secondly, we try to pinpoint

the amount of initial resources need to be assigned to

VMs to increase the ﬁnal utilization of the host. And

ﬁnally we combine both of our ﬁndings and analyze

when it is meaningful to use each technique regarding

the cost rate between migration and reconﬁguration.

During our experiments we have used some as-

sumptions which can be listed as follows:

- We have used four kinds of resources in our sim-

ulations: CPU, memory, bandwidth and storage.

Each of our applications have four constant val-

ues indicating their resource need for a speciﬁc

kind of resource. These values may point to SLA

requirements of the applications need to be hosted

by the cloud system.

- Each application’s resource needs are set inde-

pendently using a uniform random distribution.

On the other hand in our experiments we produce

those random values inside a number of intervals

resembling cloud systems with applications in dif-

ferent granularity levels. For instance in Figure 2,

10 different granularity levels are used where ran-

domly produced application resource needs vary

around 5, 10, 15, ...units respectively.

- An important assumption is an application can be

deployed on the cloud or migrated but it is not

undeployed. A more detailed simulation can be

performed by producing the incoming and leav-

ing applications using a predeﬁned distribution.

In our simulations we mainly focus on fully uti-

lizing the host so we only handle the case where

applications never leave the host machine.

Our simulations are developed using JAVA program-

ming language. We have used the native library

lp solve and Java ILP as API to solve the MILP prob-

lem introduced earlier. We have simulated four VMs

residing in a host machine which let us to run our sim-

ulations on a decent laptop. In our simulations we

have used POJOs to resemble hosts and virtual ma-

chines involving integer variables(resource variables)

that represent the resources being used in simulations.

Applications are added to system as discrete events

causing the resource variables to be decreased accord-

ing to the resource consumption values of the appli-

cations. The resource consumption values are deter-

mined randomly inside speciﬁc intervals. These inter-

vals are indicated in the result charts presented later in

the paper. Heavier analysis on a large number of VMs

residing in federated cloud environments can also be

simulated however MILP performance exponentially

decrease for such cases.

4.2 Experiment Results

Our ﬁrst experiment consists of assigning applica-

tions having different resource needs to the equally

sized VMs in a round robin manner. When using

round robin placement the applications are placed in

VMs following a predeﬁned order. This order can be

determined using priorities but in our case all the vir-

tual machines are assumed to have the same priority.

When placing an application, if the next VM deter-

mined by the round robin is out of resources the appli-

cation will be placed to the next VM in line. This pro-

cess is carried on until an appropriate VM is found,

otherwise the application is rejected due to the lack of

resources.

In Figure 2 we consider using MILP to select a

set of applications to be migrated across VMs when

round robin algorithm is unable to ﬁnd enough re-

source for any of the VMs to satisfy SLA needs of

the submitted application. X-axis represents the in-

tervals which is used to produce random values for

application resource needs and y-axis shows the im-

provementof the host utilization before an application

is rejected because of the unavailability of resources.

The values in the y-axis are the ratio between pre-

migration and post-migration host utilization. The in-

tervals in the x-axis is determined resembling a slid-

ing window interval. In each interval the median of

the produced values is shifted 5 units, which we be-

lieve is granular enough to analyze the variations in

the utilization.

For each set of experiments applications are pro-

duced with random resource needs inside the speciﬁc

interval and assigned to an available VM using round

robin. In one set of experiments we have also applied

MILP and application migration after round robin is

OnOptimizingResourceAllocationandApplicationPlacementCostsinCloudSystems

539

0−10 5−15 10−20 15−25 20−30 25−35 30−40 40−50 50−60 60−70

Size of Apps

Improvement of Host Utilization(%)

Migration Uti/RR Uti

Figure 2: The improvement in host utilization when migra-

tion is used for different sized applications.

unable to ﬁnd a VM with necessary amount of re-

sources available. In our experiments we have used

VMs having 125 units of resources available, so it can

be seen that migration is applicable when the applica-

tion sizes being deployed to the system is less than

about 25% of VM sizes. The peak in 2535 point re-

sults from the uniform distribution around 30 unit ap-

plication size being a factor of the 120 unit resource

reserved for the host.

Our second experiment is to reveal the amount

of initial resources that needs to be assigned to each

VM in order to perform a better resource allocation.

In this experiment instead of allocating all the re-

sources equally among the VMs we start by allocating

a predetermined amount of resource equally among

VMs. Later when any of the VMs becomes full, in-

stead of rejecting an application required amount of

resources is granted to the VM so that the applica-

tion can be deployed. This gradual increase contin-

ues until the host runs out of resources and reject

incoming application. In Figure 3 we compare the

amount of utilization of the host when the ﬁrst appli-

cation is rejected for applications with different re-

source needs. The applications used in the experi-

ment can be either small(requiring 0%-25% of each

resource), large(requiring25%-50% of each resource)

or randomly mixed applications of both. We can see

that independent of the application size the host uti-

lization starts to decrease after the case where 84% of

initial resources are assigned.

Our ﬁnal experiment compares the costs of appli-

cation migration and VM reconﬁguration. We have

used only small applications(requiring 0%-25% of

each resource) and compare the total cost for differ-

ent cost rates assigned to reconﬁguration and migra-

tion cases. Instead of using real cost values we rather

chose to iterate over cost rates because cost values can

40 50 60 70 80 90 100

100

Initial Resources(%)

Host Utilization(%)

Small app

Large app

Mixed app

Figure 3: The impact of initial VM resource assignment

over host utilization.

vary signiﬁcantly based on the hardware and software

conﬁgurations of the system and technologied neing

used. However cost rates can be used independently

of these concepts because there should always be a

rate between migration cost and reconﬁguration cost.

In Figure 4 we indicate two thresholds which con-

strains the use of each concept during resource alloca-

tion process. The ﬁrst threshold is the vertical initial

resource allocation threshold which is the boundary

when the total utilization starts to decrease, in other

words it becomes less meaningful to apply reconﬁgu-

ration. The horizontal threshold is the cost induced by

solely using migration without reconﬁguration, which

marks the boundary where it becomes less feasible to

apply migration and reconﬁguration at the same time.

A logical decision is to use the lower left region in the

ﬁgure to use both concepts collaboratively.

Finally we would like to compare the performance

of MILP in terms of host utilization and execution

time for different number of VMs andusing small

scale applications. In Figure 5 utilization achieved

by applying MILP based replacement of the applica-

tions is compared with the execution time based per-

formance of MILP. As expected, MILP performance

is degraded as the number of VMs increase since there

exist a larger number of possibilities to be checked

during optimization process. However because of the

same reason the optimal placement of applications on

the VMs increases the host utilization further as the

number of VMs increases. There exist a larger num-

ber of possible VMs to place the applications upon

causing a better application scheme to be found by

the MILP.

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

540

76 78 80 82 84 86 88 90 92 94 96

Initial Resources(%)

Cost

Migration Threshold

cost rate=1

cost rate=1.5

cost rate=2

cost rate=3

Reconfiguration Threshold

Figure 4: The trade-off between cumulative migration-reconﬁguration cost and initial resource assignment for different cost

rates between migration and reconﬁguration.

4 6 8 10 12 14 16 18 20

Number of VMs

Host Utilization (%)

4 6 8 10 12 14 16 18 20

0.5

Performance(1/ms)

MILP Performance

MILP Utilization

Figure 5: Utilization rate versus performance of MILP optimization for different number of VMs.

4.3 Evaluation and Analysis

By the experiment results explained above we have

provided guidelines to be used in VM reconﬁguation

and application live-migration in cloud systems. By

analyzing our ﬁrst set of results in Figure 2 we have

revealed an interval of 0% to 25% for application re-

source needs where using live-migration increases the

total utilization of the host. This result makes sense

because as the size of the application becomes larger

it is less likely to allocate a large enough resource

space from the leftovers in the VMs. By proﬁling or

estimating resource need characteristics of the appli-

cations to be deployed to cloud system it is possible

to decide on live-migration use strategy where such a

strategy makes difference where smaller applications

are dominant.

On the other hand a similar analysis can be made

for reconﬁguration purposes by examining Figure 3

as well. It is clear that instead of fully assigning all

the resources equally to the VMs at hand, it is more

preferable to assign around 80% of the resources and

gradually reconﬁgure each VM according to needs.

An interesting fact can be the stable utilization rate

until 80%’s which may be tuned by taken the recon-

ﬁguration cost into account since assigning smaller

amount of resources initially will clearly require a

larger amount of reconﬁguration.

Taking into account the costs of reconﬁguration

and live-migration techniques Figure 4 identiﬁes a re-

gion where it is possible to leverage the trade-off be-

tween applying these two techniques. Since the cost

of each operation can change from system to system

we have chosen to use different cost rates in our anal-

ysis. We can interprete the regions formed by the

thresehold values in the ﬁgure as follows: the upper

left quadrant contains the cases where it is advanta-

geous to perform conﬁguration but the cost of using

reconﬁguration and live-migration together exceeds

the cost of using migration solely. In this case it is

meaningful to chose and use only one of the tech-

niques at all. The lower left quadrant is the ideal

case where both of the techniques can be used col-

laboratively. For the lower right quadrant using live-

migration is preferrable since gains by reconﬁgura-

tion starts to decrease. However, it can still be mean-

ingful to use reconﬁguration as well for lower per-

centages of initial resource allocation regarding the

cost rates. The upper-right quadrant is the area where

using solely live-migration should be preferred.

OnOptimizingResourceAllocationandApplicationPlacementCostsinCloudSystems

541

5 CONCLUSIONS AND FUTURE

WORK

Advances in the cloud resource allocation technolo-

gies led to apply resource allocation strategies at dif-

ferent layers of a cloud system. In this paper we inves-

tigated resource allocation challenges at two different

levels, more speciﬁcally between host and VM layer

and between VM and application layer. Our analyses

show that when the resource consumption of appli-

cations gets larger it becomes less meaningful to ap-

ply application live-migration at VM layer. Moreover

it also becomes less feasible to use VM reconﬁgura-

tion as the amount of initial resource assigned to each

VM start to go beyond 84%. In deciding the level

to apply resource allocation strategy VM reconﬁgu-

ration is preferable when the amount of cost rate be-

tween reconﬁguration and live migration is below 2.

Moreover, the optimal cost is observed when an ini-

tial resource assigned to each VM is inside 80%-84%

area. As a future study we plan to realize the cloud

system environment in an actual testbed and replicate

our results in actual workloads. Additionally we plan

to extend our approach to support horizontal scaling

and multiple hosts which require more advanced(even

non-linear) optimization technniques.

ACKNOWLEDGEMENTS

This work is performed in joint with “mCloud”

project of Simternet Iletisim Sistemleri Reklam San.

ve Tic. Ltd. Sti. “mCloud” project is supported

by The Scientiﬁc and Technological Research Coun-

cil of Turkey(TUBITAK)-TEYDEB project number

7130115. This work is also partly supported by ITU

HP Cloud Computing Center.

REFERENCES

Bennani, M. N. and Menasc, D. A. (2005). Resource alloca-

tion for autonomic data centers using analytic perfor-

mance models. In Autonomic Computing, 2005. ICAC

2005. Proceedings. Second International Conference

on, pages 229–240.

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 CLOUD

COMPUTING 2011, The Second International Con-

ference on Cloud Computing, GRIDs, and Virtualiza-

tion, pages 127–132.

Chen, W., Qiao, X., Wei, J., and Huang, T. (2012). A

proﬁt-aware virtual machine deployment optimization

framework for cloud platform providers. In Cloud

Computing (CLOUD), 2012 IEEE 5th International

Conference on, pages 17–24.

Clark, C., Fraser, K., Hand, S., Hansen, J. G., Jul, E.,

Limpach, C., Pratt, I., and Warﬁeld, A. (2005). Live

migration of virtual machines. In Proceedings of the

2nd conference on Symposium on Networked Systems

Design & Implementation - Volume 2, NSDI’05, pages

273–286, Berkeley, CA, USA. USENIX Association.

He, S., Guo, L., Ghanem, M., and Guo, Y. (2012). Improv-

ing resource utilisation in the cloud environment using

multivariate probabilistic models. In Cloud Comput-

ing (CLOUD), 2012 IEEE 5th International Confer-

ence on, pages 574–581.

Hwang, I. and Pedram, M. (2012). Portfolio theory-based

resource assignment in a cloud computing system. In

Cloud Computing (CLOUD), 2012 IEEE 5th Interna-

tional Conference on, pages 582–589.

Kikuchi, S. and Matsumoto, Y. (2012). Impact of live

migration on multi-tier application performance in

clouds. In Cloud Computing (CLOUD), 2012 IEEE

5th International Conference on, pages 261–268.

Ruiz-Alvarez, A. and Alvarez, M. (2012). A model and

decision procedure for data storage in cloud comput-

ing. In Cluster, Cloud and Grid Computing (CCGrid),

2012 12th IEEE/ACM International Symposium on,

pages 572–579. IEEE.

Van den Bossche, R., Vanmechelen, K., and Broeckhove, J.

(2010). Cost-optimal scheduling in hybrid iaas clouds

for deadline constrained workloads. In Cloud Com-

puting (CLOUD), 2010 IEEE 3rd International Con-

ference on, pages 228–235. IEEE.

Verma, A., Kumar, G., and Koller, R. (2010). The cost

of reconﬁguration in a cloud. In Proceedings of the

11th International Middleware Conference Industrial

track, Middleware Industrial Track ’10, pages 11–16,

New York, NY, USA. ACM.

Verma, A., Kumar, G., Koller, R., and Sen, A. (2011).

Cosmig: Modeling the impact of reconﬁguration in

a cloud. In Modeling, Analysis Simulation of Com-

puter and Telecommunication Systems (MASCOTS),

2011 IEEE 19th International Symposium on, pages

3–11.

Wang, Y., Chen, S., and Pedram, M. (2013). Service

level agreement-based joint application environment

assignment and resource allocation in cloud comput-

ing systems. In Green Technologies Conference, 2013

IEEE, pages 167–174.

Yang, K., Gu, J., Zhao, T., and Sun, G. (2011). An op-

timized control strategy for load balancing based on

live migration of virtual machine. In Chinagrid Con-

ference (ChinaGrid), 2011 Sixth Annual, pages 141–

146.

Zhu, X. and Singhal, S. (2001). Optimal resource assign-

ment in internet data centers. In Modeling, Analysis

and Simulation of Computer and Telecommunication

Systems, 2001. Proceedings. Ninth International Sym-

posium on, pages 61–69.

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

542