Improving Resource Utilization in Cloud Environments

using Application Placement Heuristics

Atakan Aral and Tolga Ovatman

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

Keywords:

Application Placement, Resource Allocation, Cloud Computing, Software as a Service.

Abstract:

Application placement is an important concept when providing software as a service in cloud environments.

Because of the potential downtime cost of application migration, most of the time additional resource acquisi-

tion is preferred over migrating the applications residing in the virtual machines (VMs). This situation results

in under-utilized resources. To overcome this problem static/dynamic estimations on the resource require-

ments of VMs and/or applications can be performed. A simpler strategy is using heuristics during application

placement process instead of naively applying greedy strategies like round-robin. In this paper, we propose a

number of novel heuristics and compare them with round robin placement strategy and a few proposed place-

ment heuristics in the literature to explore the performance of heuristics in application placement problem.

Our focus is to better utilize the resources offered by the cloud environment and at the same time minimize the

number of application migrations. Our results indicate that an application heuristic that relies on the difference

between the maximum and minimum utilization rates of the resources not only outperforms other application

placement approaches but also signiﬁcantly improves the conventional approaches present in the literature.

1 INTRODUCTION

In addition to software delivered as services over the

Internet, hardware and software systems that make

the delivery possible are referred as cloud comput-

ing (Armbrust et al., 2010). In cloud computing, re-

sources such as CPU, memory, bandwidth and stor-

age are treated as utilities that can scale up and down

on demand. It also allows per-usage metering and

billing of these resources. By means of cloud comput-

ing, cloud users can handle unexpected high demands

without over-provisioning and they do not need to in-

vest for abundant hardware resources initially. Cloud

providers, on the other hand, have the opportunity of

reallocating idle resources for other cloud users.

Cloud computing services are categorized into

three layers according to the level of abstraction they

provide. These are, in ascending order of abstraction:

(1) Infrastructure as a Service (IaaS) that provide vir-

tual raw resources, (2) Platform as a Service (PaaS)

that provide virtual development environments, and

(3) Software as a Service (SaaS) that provide online

applications on demand (Buyya et al., 2011).

One general research challenge in cloud comput-

ing is the efﬁcient allocation of cloud resources to

users since cloud providers should satisfy quality of

service (QoS) objectiveswhile minimizing their oper-

ational cost (Zhang et al., 2010). In this paper, we ap-

proach this challenge from the perspective of a SaaS

provider. Provider receives user applications with het-

erogeneous resource requests and assigns these appli-

cations to the virtual machines (VMs) with ﬁnite ca-

pacities. Multiple requests can be assigned to a single

VM so the SaaS provider would prefer to fully utilize

the VM resources at hand before hiring new ones. It

is also possible to move applications between VMs,

but this would cause a temporary halt of the migrating

application and would reduce QoS. Consequently, the

objective is to maximize the number of applications

assigned to VMs and at the same time, minimize the

number of application migrations.

For the sake of simplicity, we make the following

four assumptions regarding the characteristics of the

applications and VMs.

1. Cloud provider possesses constant number of

VMs with homogeneous capacities. These capac-

ities are four dimensional (namely CPU, memory,

bandwidth and storage). It is not possible to add

new VMs or resize existing ones.

2. Similarly, applications have the same four con-

stant consumptions (requests).

527

Aral A. and Ovatman T..

Improving Resource Utilization in Cloud Environments using Application Placement Heuristics.

DOI: 10.5220/0004848005270534

In Proceedings of the 4th International Conference on Cloud Computing and Services Science (CLOSER-2014), pages 527-534

ISBN: 978-989-758-019-2

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

3. Requests of the applications are foreknown.

4. Cloud provider receives one application at a time

and after this application is assigned to a VM, it

runs inﬁnitely (i.e., it does not have a life span).

However, it may temporarily suspend, migrate to

another VM and continue its execution there.

These assumptions may seem too restrictive at

ﬁrst but our approach can be easily generalized to real

world cloud scenarios. We use static values (VM ca-

pacities, counts etc.) because we aim to use current

resources at hand in most efﬁcient way. When these

resources are almost fully utilized, SaaS provider pos-

sibly increases them, and our algorithm immediately

adapts and starts to make decisions according to the

new conﬁguration. Similarly, when resource con-

sumption changes over time or even terminates, new

decisions will be made considering the current utiliza-

tion. Third assumption of foreknown resource con-

sumptions is based on the Service Level Agreements

(SLAs) made between the cloud provider and cloud

user where expected performance criteria is deﬁned.

Fundamentally, the problem of optimally placing

applications to VMs can be formulated as an NP-hard

multidimensional bin packing problem and solved via

optimization techniques such as linear programming

(Zhang et al., 2010). However, it would be computa-

tionally expensive to run the optimization algorithm

at each application arrival especially for more than a

few VMs. Moreover, optimization would incur large

number of application migrations. Our solution is

to distribute applications to VMs using an intelligent

heuristic until the point that a migration is inevitable

(i.e. when no VM has enough capacity to accept in-

coming application). After this point, an optimization

technique can be employed to rearrange the applica-

tions.

Contributions in this paper may be summarized as

follows.

1. A Mixed Integer Programming (MIP) formulation

of the application placement problem that mini-

mizes the number of migrations

2. Novel heuristics to assign applications to VMs ef-

ﬁciently and intelligently

3. Comparison of the heuristics to the existing ones

4. Analysis of the heuristic approach on various VM

counts and capacities

The following section gives an overview of previ-

ous work in resource allocation and application place-

ment literature. Section 3 explains the details of

our method including MIP formulation and heuristics,

while Section 4 contains the experiments, their results

and discussion. Finally, we conclude the paper with

ﬁnal remarks in Section 5.

2 RELATED WORK

Allocation of a cloud provider’s resources to its cus-

tomers is entitled as resource allocation, resource se-

lection or application placement depending on the

context and what is meant by the term resource. In

this section, we review different approaches to the

problem in the cloud computing literature. It should

be noted that similar problems are encountered and

extensively studied before the arrival of cloud com-

puting (Urgaonkar et al., 2005). However cloud com-

puting brings some unique challenges to the problem

such as inclusion of network as a resource and as a re-

quest, geo-diversity and location limitations (jurisdic-

tion), multi-tenancy, more heterogeneous resources

and less predictable demands.

2.1 VM Assignment Strategies

Selection of physical machines to host virtual ma-

chines while fulﬁlling the requirements and optimiz-

ing resource usage is a problem that can be solved via

optimization and approximation algorithms. Some

examples use heuristics, linear programming, artiﬁ-

cial intelligence, nature inspired computing and game

theory (Endo et al., 2011).

In a recent study (Papagianni et al., 2013), the

problem is formulated as two phases that must be op-

timized together: node mapping and link mapping.

User request come in the form of VMs with deﬁned

capacities and their connectivity as a graph. The aim

is to optimally assign these capacity request to phys-

ical machines (node mapping) and connectivity re-

quests to network entities (link mapping). They apply

a relaxed MIP approximation for the node mapping

and shortest path (or minimum cost ﬂow) algorithm

for the link mapping in a coordinated way.

Another approach (Xiao et al., 2013), introduced

the skewness metric to measure unevenness in the uti-

lizations of various resources within a server. They

aim to ﬁnd a trade-off between overload avoidance

and green computing concepts. Using a set of thresh-

old based heuristics and a prediction algorithm they

dynamically create a list of migrations that relieves

overloaded servers to ensure QoS and evacuates un-

derloaded ones to exploit green computing.

Following the improvements of the cloud systems,

VM migration is being used to increase utilization

of the virtual machines on the cloud infrastructure.

(Yang et al., 2011) considers the different VM migra-

tion strategies adopted by the host machines that have

different load states by taking into account four dif-

ferent resource types. In a more recent study (Wang

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

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

528

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.

2.2 Application Assignment Strategies

Literature in the area of allocation within a VM is not

as extensive as the one in allocation within a server,

as the most studies consider resource allocation for

the beneﬁt of IaaS providers. A detailed model of

SLAs is given in the study (Wu et al., 2011) which

falls into the application placement category. SLAs

between cloud providers and users are taken into con-

sideration to quantitatively measure the service qual-

ity. They also suggest a simple heuristic: ‘assign the

request to the VM with the lowest (or the highest in

other variation) utilization’.

In another study (Tang et al., 2007) optimal load

balancing for dynamically changing demands is con-

sidered. Their approximation method aim to maxi-

mize satisﬁed demand and minimize application mi-

grations. Proposed application placement controller

estimates future demand and migrates applications

accordingly to avoid overloading.

A recent work (Espadas et al., 2013) offers a

resource allocation mechanism that exploits multi-

tenancy and provides tenant isolation, load balancing

and VM allocation. They also formally measure over

and underprovisioning of VM resources.

3 APPLICATION PLACEMENT

ALGORITHM

3.1 Overview

A high level pseudo code of both parts is given in Al-

gorithm 1. In part I (lines 4 to 12), migrations are not

allowed, and each arriving application is assigned to

the VM determined by an heuristic of evenness. Goal

of this part is to keep the utilization of four resources

in a VM approximately even. By increasing the re-

source evenness of all VMs, overall utilization of the

resources are maximized (Xiao et al., 2013).

To decide which VM is the best for a given appli-

cation, it is transiently considered to be assigned to

the VMs that has enough remaining capacity one by

one and evenness is calculated. Then, the application

is actually assigned to the VM with the best evenness

value (line 12). At one point of the part I, incoming

application’s resource request will not ﬁt into the re-

maining capacity of any VM (the application is not

directly assignable and condition in line 11 is not sat-

isﬁed). However, rejected application can still be as-

signed to a VM if some applications are re-assigned.

Part II (lines 13 to 16) migrates some of the appli-

cations and tries to make room for the new applica-

tion. If part II succeeds to assign the application (line

16), part I of the algorithm continues to receive new

applications; otherwise, it is certain that there is no

placement that can assign received applications to the

VMs and the algorithm terminates (line 17).

Algorithm 1: Pseudo code for the application

placement strategy.

1 rejected ← false

2 while rejected = false do

3 receive application A

4 assignable ← false

5 foreach virtual machine VM do

6 if VM has enough capacity for A then

7 assignable ← true

8 assign A to VM

9 calculate unevenness of VM

10 remove A from VM

11 if assignable = true then

12 assign A to VM with min unevenness

13 else

14 run optimization algorithm

15 if optimization succeeds then

16 assign A and migrate others

17 else rejected ← true

One variation of part I can be obtained by calcu-

lating unevenness twice, i.e. before and after tenta-

tively assigning the application. In that case, actual

assignment is made to the VM whose unevenness de-

creases the most, instead of the one whose unevenness

becomes the minimum.

3.2 Heuristics

We propose four heuristics to approximate the un-

evenness in the utilization of the resources on a VM.

We also adapt the skewness heuristic originally sug-

gested for assigning VMs to servers (Xiao et al.,

2013). Finally we implement the greedy approxima-

tion algorithm, round-robin, to use as the baseline.

3.2.1 Standard Deviation (SD)

Standard deviation is a natural choice to ﬁnd the un-

evenness of data points since it shows the amount of

dispersion from the average. SD of the resource uti-

lizations on a VM v can be calculated as follows.

ImprovingResourceUtilizationinCloudEnvironmentsusingApplicationPlacementHeuristics

529

SD(v) =

∑

i=1

− ¯r)

(1)

Here, r

denotes the utilization of the ith resource,

while ¯r is the average utilization of all resources of v.

m is the number of resources.

3.2.2 Span (SP)

For the trivial case of only four instances, a simpler

heuristic may be used. Since we are interested more

in outliers than inliers, difference between the max-

imum and minimum utilization rates can be a good

candidate for an unevenness heuristic.

SP(v) = max

i∈[1,m]

) − min

i∈[1,m]

) (2)

3.2.3 Cumulative Difference (CD)

Another candidate is the total difference of the uti-

lization rates of all resource pairs on v, as given by

the formula below. This heuristic also considers the

inliers similar to SD, but it is simpler.

CD(v) =

∑

i=1

∑

j=1

− r

(3)

3.2.4 Cumulative Difference from Minimum

(DM)

A similar heuristic to CD is to calculate differences

from the minimum utilization rate. This approach

tries to ensure that no resource is under-utilized (low

outlier) relative to others, while previous approaches

are trying to prevent both low and high outliers.

DM(v) =

∑

i=1



− min

j∈[1,m]

)



(4)

3.2.5 Skewness (SK)

Formulation of SK is given in (Xiao et al., 2013).

SK(v) =

∑

i=1



¯r

− 1



(5)

3.2.6 Round-Robin (RR)

Round-robin algorithm is a very straightforward solu-

tion to the application placement problem. For each

application, the greedy algorithm attempts to assign it

to the next VM. If that VM is not available, consecu-

tive ones are tested in order. In the ideal case where

all VMs are always available, it assigns equal number

of applications to each of them. Naturally, it does not

consider the evenness of resource types within a VM

as opposed to the heuristics mentioned above.

3.3 MIP Formulation

To determine an optimal application placement

scheme, we use MIP techniques. We minimize the

number of migrations when determining the place

(VM) of each application from the pool of VMs.

In achieving this, for each solution we compare the

place of each application with its former place and

try to maximize the number of applications that don’t

change their place.

We have used the native library lp

solve and Java

ILP API to solve the MIP problem introduced earlier.

We have simulated four VMs residing in a host ma-

chine which let us to run our simulations on a decent

personal computer. Heavier analysis on a large num-

ber of VMs residing in federated cloud environments

can also be simulated, however MIP performance ex-

ponentially decrease for such cases.

3.3.1 Variables

When using MIP, it is required to build formulae that

represents the constraints of the system and an ob-

jective function to be optimized during the process.

In constructing such formulae variables that repre-

sent the overall properties of the system shall be used.

Below, the variables we used in our model and their

meanings are explained brieﬂy:

• #VMs: Number of VMs present in the cloud en-

vironment.

• #Apps: Number of applications to be placed upon

the present VMs.

• oldAsgn: A boolean matrix that holds the present

assignment of each application upon a VM.

• newAsgn: A boolean matrix that holds the result-

ing assignment by MIP.

• resNeed: An integer matrix that holds the amount

of resource needed by each application for the

four different type of resources mentioned before.

• resAv: An integer matrix that holds the amount of

resource available for each VM and each resource.

3.3.2 Objective Function

In modeling resource consumption for a cloud envi-

ronment, we provide the MIP solver with Equation 6

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

new application-VM assignment products which pro-

duce a value of 1 if the assignment didn’t change and

0 if a migration is present.

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

530

#VMs

∑

i=0

#Apps

∑

j=0

(newAsgn[i][j] × oldAsgn[i][j]) (6)

3.3.3 Constraints

Naturally, we also apply a number of constraints in

order to drive the MIP to produce meaningful results.

In Equation 7 we guarantee that each application has

been assigned to exactly one VM. Additionally in

Equation 8 resource needs of each application that has

been assigned to a speciﬁc VM for each type of re-

sources are summed up to be less than the available

resource assigned to the VM.

#VMs

i=0

#Apps

∑

j=0

newAsgn[i][j] = 1

(7)

#VMs

i=0

#Res

j=0

#Apps

∑

k=0

newAsgn[i][k]× resNeed[ j][k]

≤ resAv[i][ j]

(8)

4 EXPERIMENTS

A simple simulator is implemented to conduct the ex-

periments explained in this section. Simulator gen-

erates applications with uniformly random resource

requests within a given range and manages VMs

with user deﬁned count and capacity. For all ex-

periments, applications are generated with resource

requests within the range of [0, 30] units while VM

count and capacities varied.

4.1 Postponement of Migration

This ﬁrst experiment aims to compare the power of

heuristics to assign maximum number applications

before a migration is necessary. As mentioned before,

migrations should be avoided since they may cause

temporary downtimes for an application. So, a good

heuristic should place applications in a way that mi-

grations become compulsory as late as possible.

4.1.1 Experimental Setup

Part II of the algorithm is not required for this experi-

ment since we are only interested in the performance

of the heuristics before the optimization starts. Ex-

cluding the computationally expensive part of the al-

gorithm allowed to test heuristics for a great number

of conﬁgurations and repetitions. Two variants (mini-

mum and most decreasing) of ﬁve heuristics as well as

round-robin are compared on 100 different VM con-

ﬁgurations. This consists of 10 VM counts ranging

from 3 to 12 and 10 VM capacities ranging from 75

to 300 units per resource.

All 11 strategies are run on each conﬁguration

30.000 times to avoid randomness of applications af-

fect the results. Execution of a strategy on a conﬁg-

uration is stopped when no VM is available for the

received application. After that, the total number of

applications assigned to all VMs are logged and the

average of 30.000 runs are calculated.

4.1.2 Results and Discussion

Due to the great number of conﬁgurations tested in

this experiment, it is not possible to include all results

here. We instead present ﬁve conﬁgurations with a

ﬁxed VM count (8 VMs) on the left half of the Table

1 and ﬁve conﬁgurations with a ﬁxed VM capacity

(200 units per resource) on the right half. Remaining

90 conﬁgurations follow the same pattern as them.

For each conﬁguration, Table 1 contain average

number of applications assigned by 11 strategies, ex-

pected number of applications to fully utilize VMs

and the rate of improvement by the best performing

strategy in comparison to the baseline (RR). Sub-

scripts min and dec stand for the minimum and most

decreasing variants of the strategies, respectively.

Our results indicate that in 477 of 500 cases

(95,4%), min variant outperformed the dec variant of

the same strategy. In most of the 23 cases where min

variant performed worse, tiny VM capacities and the

DM strategy are used. This clearly demonstrates that

arriving application should be assigned to the VM that

becomes the most even with the assignment, instead

of the one whose evenness increases the most.

It is also evident that intelligent heuristics always

perform better than the greedy assignment. In all

100 conﬁgurations, all 10 strategies manage to assign

more applications than the round-robin. The worst

improvement rate is 4, 3% (3 VMs with 300 units of

capacity per resource) while the best one is 12, 1%

(12 VMs with 75 units of capacity per resource). Uti-

lization rate of round-robin is between 73, 8% and

90, 2% while utilization rate of the heuristics is be-

tween 80, 5% and 96, 4%.

Improvement rate gets better as the number of

VMs increase (as seen in the right half of the ﬁnal

row in Table 1) but their capacities decrease (as seen

left half of the same row). With high capacities and

small number of VMs, randomness of resource re-

quests tend to cover imperfect assignments while with

low capacities and large number of VMs, decisions

are more critical and should be made with more care.

In cases that are ideal for cloud computing beneﬁts

(i.e. many distributed VMs with relatively low capac-

ities), heuristics show their real strength.

ImprovingResourceUtilizationinCloudEnvironmentsusingApplicationPlacementHeuristics

531

Table 1: Partial result of the application count experiment.

VM Capacity 100 150 200 250 300 200 200 200 200 200

VM Count 8 8 8 8 8 4 6 8 10 12

RR 42,0 67,2 92,7 118,4 144,3 46,0 69,3 92,7 116,2 139,7

min

45,8 72,3 98,8 125,3 151,9 48,1 73,4 98,8 124,3 149,9

dec

45,5 72,0 98,6 125,2 151,9 48,1 73,2 98,6 124,1 149,7

min

46,2 73,2 100,2 127,2 154,3 48,7 74,4 100,2 126,2 152,3

dec

46,0 72,6 99,3 126,0 153,2 48,4 73,8 99,3 124,9 150,7

min

46,2 73,2 100,2 127,3 154,3 48,7 74,4 100,2 126,2 152,3

dec

45,8 72,3 98,9 125,5 152,1 48,2 73,4 98,9 124,4 150,1

min

46,2 73,2 100,2 127,3 154,3 48,8 74,4 100,2 126,2 152,3

dec

45,9 72,5 99,1 125,9 152,7 48,3 73,7 99,1 124,8 150,5

min

45,7 72,6 99,6 126,5 153,6 48,3 73,8 99,6 125,4 151,5

dec

45,8 72,5 99,2 126,2 153,1 48,3 73,6 99,2 125,0 150,8

Expected 53,3 80,0 106,7 133,3 160,0 53,3 80,0 106,7 133,3 160,0

Improvement 10,0% 8,9% 8,1% 7,5% 6,9% 6,1% 7,4% 8,1% 8,6% 9,0%

Since it is already demonstrated that heuristics

make better assignments than the greedy algorithm

and the min variants are better than dec variants, we

can ﬁnally make a comparison among them. In triv-

ial cases where capacity per resource is less than 100

units, all heuristics perform more or less the same,

while in all other cases SP, SD and CD perform sig-

niﬁcantly better than SK and DM.

Although being still much better than greedy,

skewness algorithm is outperformed by the best three

heuristics by 1, 2% on average. Our observations

show that both variants of SK are sensitive to aver-

age resource utilization of the VMs. As a result, they

tend to assign new applications to the VMs that con-

tain more applications although better alternatives in

terms of evenness are available. This causes their

placements resemble greedy ﬁrst-ﬁt algorithm.

4.2 Minimization of Migration Count

Goal of this experiment is to assess the quality of ap-

plication placement of the strategies. A good place-

ment is expected to allow assignment of a new appli-

cation after a small number of migrations, or in the

ideal case, fully utilize VMs without any migrations.

It should be noted that the objective of the optimiza-

tion algorithm is to minimize the number of migra-

tions so the migration counts are guaranteed to be the

smallest possible values.

4.2.1 Experimental Setup

Both part I and II are included in this experiment and

that forces us to decrease repetition count. That’s

why, only the best three heuristics of the ﬁrst exper-

iment, round-robin and skewness algorithms are exe-

cuted for a single conﬁguration: three VMs with ca-

pacities of 300 units per resource. Greater number of

VMs causes a signiﬁcant increase in running time.

An execution is stopped on the following two con-

ditions: (1) In part I, when remaining capacities of all

VMs are less than the expected value of the four re-

source requests (i.e., 15 units), (2) In part II, when the

optimization algorithm runs once. We call the ﬁrst

condition a perfect placement since all VMs are al-

most fully utilized without any migrations. Second

condition is only considered when a perfect place-

ment is not achieved and the algorithm continues to

part II. In that case, if the algorithm succeeds to as-

sign the application to a VM, number of migrations

are logged. Otherwise, if the application is rejected, it

is also considered as a perfect placement because MIP

proved that it was impossible also for the heuristic to

make that placement.

Each strategy is executed 1.000 times calculating

the average number of migrations and the number of

perfect placements. Number of migrations for a per-

fect placement is regarded zero.

4.2.2 Results and Discussion

Table 2 gives the average number of migrations in-

troduced by the optimization after each ﬁve strategies

(RR, SK

min

and three best performers of experiment

1) failed to assign any more applications. The table

also contains the number of cases (among 1.000 ex-

ecutions) where heuristics made a perfect placement

so an optimization was not required.

Results demonstrate that placements made by the

heuristics allow assignment of new applications with

less changes since SD

min

, SP

min

and CD

min

strate-

gies required 34, 5%, 33, 3% and 31, 0% less migra-

tions than RR, respectively. Similar to the ﬁrst experi-

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

532

Table 2: Results of the migration count experiment.

Strategy Avg. Migr. Count Perfect Count

RR 8,4 26

min

5,5 108

min

5,6 100

min

5,8 86

min

7,2 88

ment SK

min

performed somewhere in the middle with

14, 3% better than RR but 27, 8% worse than the oth-

ers in average. SD

min

and SP

min

made perfect place-

ments in 10, 8% and 10, 0% of executions, followed

by SK

min

with 8, 8% and CD

min

with 8, 6%. Coin-

cidental perfect placements are made by RR in only

2, 6% of cases.

It is clear that intelligent heuristics make bet-

ter placements that can be optimized with signiﬁ-

cantly less migrations and that are already optimum

in roughly four times more incidents than the greedy

approach. Suggested heuristics also outperformSK

min

clearly with the average number of migrations. As a

result, they allow applications to run with less suspen-

sion as well as decreasing the need for time consum-

ing optimization algorithm.

Due to its simplicity, we choose SP

min

heuristic

to represent heuristic approach in the future perfor-

mance analysis.

4.3 Analysis on Constant Total Capacity

As explained before, partitioning total hardware ca-

pacity to higher number of virtual machines provides

beneﬁts of cloud computing but reduces overall uti-

lization. This reduction can be especially signiﬁcant

when different application types are not combined ap-

propriately and a certain resource of a VM runs much

shorter than the others. In this experiment, we aim to

observe utilization performance of our algorithm in

comparison to greedy approach given various number

of VMs sharing a ﬁxed hardware capacity.

4.3.1 Experimental Setup

min

heuristic which is one of the best performing

and simplest strategies, is compared to the greedy

round-robin algorithm using 3 to 12 VMs. The VMs

share a total capacity of 1.000 units per resource.

Similar to the ﬁrst experiment, part I of the algorithm

is executed 30.000 times for both strategies to calcu-

late average count of applications that are assigned.

Executions are stopped when a migration is necessary.

10%

11%

12%

13%

14%

15%

3 4 5 6 7 8 9 10 11 12

Improvement Rate

Number of Assigned Applications

VM Count

GR HR Improvement

Figure 1: Results of the constant capacity analysis of the

heuristic approaches (HR) in comparison to gready (GR).

4.3.2 Results and Discussion

Chart in Figure 1 displays the number of applications

that are assigned by heuristic (SP

min

) and greedy (RR)

approaches to the same constant total capacity that is

partitioned to varying number of VMs. It also shows

the rate of improvement provided by the heuristic

strategy on its secondary y-axis.

As expected, both strategies assign less applica-

tions as the number of VMs increase, however the

rate of decrease is different. RR starts with 60, 0 ap-

plications and encounters a decrease to 50, 7 applica-

tions with a slope of −1, 03. SP

min

, on the other hand,

starts with slightly more applications, 62, 6, and ends

around 56, 9 with a less steep slope around −0, 63.

Other heuristics are also tested but the results are

omitted since their performance is nearly as good as

min

. That’s why we can consider this experiment

as a comparison of heuristic and greedy approaches

rather than SP

min

and RR. Experimental results in-

dicate that using a more intelligent heuristic instead

of the greedy algorithm slows down the decrease in

resource utilization while partitioning increases. Ac-

cordingly, improvement provided by heuristics gets

more signiﬁcant. This is especially important when

we consider that beneﬁts of cloud computing becomes

more evident with higher number of VMs.

5 CONCLUSION

In this paper we propose a method for assigning ap-

plications with multi-dimensional resource requests

to virtual machines with ﬁnite capacity. The method

employs heuristics and mixed integer programming to

optimally solve the application placement problem.

Our goal is to allow SaaS providers to host maxi-

mum number of applications with their resources at

the same time increasing their service quality.

While optimization algorithms such as our MIP

ImprovingResourceUtilizationinCloudEnvironmentsusingApplicationPlacementHeuristics

533

formulation guarantee to assign maximum number of

applications, they decrease QoS in two ways: (1)

They are extremely time consuming for high number

of instances so they delay the decision making pro-

cess while the request is pending, (2) Their decisions

cause migrations of applications which require tem-

porary downtimes.

Suggested heuristics approximate to the optimal

placement before the optimization is inevitable. They

do so by combining workloads together to keep uti-

lization of resources as even as possible. Previously

suggested unevenness metric, skewness, fails to yield

good combinations while greedy algorithms such as

round-robin do not consider evenness at all.

Heuristics are evaluated via a simple simulator

that supports random generation of demands and

maintains VM capacities. Our experiments demon-

strate that heuristics provide the following improve-

ments to QoS in comparison to greedy approximation.

1. They make the optimal placement and almost

fully utilize VMs up to 10, 8% of cases. This

is four times more than the rate provided by the

greedy algorithm. In such cases, service quality

is preserved since the execution of MIP algorithm

and application migrations are not required.

2. In the rest of the cases, they delay the requirement

for MIP algorithm up to 12, 1% applications. This

means more applications can be accepted without

migrations and any effect on service quality.

3. When the execution of MIP algorithm is manda-

tory, their placements require up to 34, 5% less ap-

plication migrations causing less applications to

suspend and less harm to service quality.

As future work, we aim to reproduce our results

on a state of the art cloud computing simulator and to

test our heuristics on other scenarios. These include

other random distributions of resource requests and

dynamic resource demands.

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 Council

of Turkey (TUBITAK) – TEYDEB project number

7130115.

REFERENCES

Armbrust, M., Fox, A., Grifﬁth, R., Joseph, A. D., Katz,

R., Konwinski, A., Lee, G., Patterson, D., Rabkin, A.,

Stoica, I., and Zaharia, M. (2010). A view of cloud

computing. Commun. ACM, 53(4):50–58.

Buyya, R., Broberg, J., and Goscinski, A. M. (2011). Cloud

computing: Principles and paradigms, volume 87.

Wiley. com.

Endo, P., de Almeida Palhares, A., Pereira, N., Goncalves,

G., Sadok, D., Kelner, J., Melander, B., and Mangs, J.-

E. (2011). Resource allocation for distributed cloud:

concepts and research challenges. Network, IEEE,

25(4):42–46.

Espadas, J., Molina, A., Jim´eNez, G., Molina, M.,

Ram´ıRez, R., and Concha, D. (2013). A tenant-based

resource allocation model for scaling software-as-a-

service applications over cloud computing infrastruc-

tures. Future Gener. Comput. Syst., 29(1):273–286.

Papagianni, C., Leivadeas, A., Papavassiliou, S., Maglaris,

V., Cervello-Pastor, C., and Monje, A. (2013). On the

optimal allocation of virtual resources in cloud com-

puting networks. Computers, IEEE Transactions on,

62(6):1060–1071.

Tang, C., Steinder, M., Spreitzer, M., and Paciﬁci, G.

(2007). A scalable application placement controller

for enterprise data centers. In Proceedings of the 16th

international conference on World Wide Web, WWW

’07, pages 331–340, New York, NY, USA. ACM.

Urgaonkar, B., Shenoy, P., Chandra, A., and Goyal, P.

(2005). Dynamic provisioning of multi-tier internet

applications. In Autonomic Computing, 2005. ICAC

2005. Proceedings. Second International Conference

on, pages 217–228.

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.

Wu, L., Garg, S., and Buyya, R. (2011). Sla-based resource

allocation for software as a service provider (saas) in

cloud computing environments. In Cluster, Cloud and

Grid Computing (CCGrid), 2011 11th IEEE/ACM In-

ternational Symposium on, pages 195–204.

Xiao, Z., Song, W., and Chen, Q. (2013). Dynamic resource

allocation using virtual machines for cloud computing

environment. Parallel and Distributed Systems, IEEE

Transactions on, 24(6):1107–1117.

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.

Zhang, Q., Cheng, L., and Boutaba, R. (2010). Cloud com-

puting: state-of-the-art and research challenges. Jour-

nal of Internet Services and Applications, 1(1):7–18.

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

534