HYBRID CLOUD ARCHITECTURE FOR SHORT MESSAGE

SERVICES

Yrjo Raivio

, Oleksiy Mazhelis

, Koushik Annapureddy

Ramasivakarthik Mallavarapu

and Pasi Tyrv¨ainen

Department of Computer Science and Engineering, Aalto University, Aalto, Finland

Department of Computer Science and Information Systems, University of Jyv¨askyl¨a, Jyv¨askyl¨a, Finland

Keywords:

Hybrid Cloud, Short Message Service, Cost Evaluation, Trafﬁc Characterization, Autoscaling.

Abstract:

Dedicated and expensive computing platforms are commonly applied to mobile network systems. This is

necessary, despite the economic burden, due to strict performance requirements in availability, latency and

throughput. However, cloud computing is changing the rules of the game by offering cost efﬁcient and high

performance computer systems. Pay-per-use principle is helping network administrators to scale the comput-

ing capacity on a need basis, reducing both capex and opex costs. Several networks can beneﬁt from this

advantage in wireless services, including both end user and internal back end services. The focus in this paper

is on the Short Message Service (SMS), which is one of the most successful and widespread end user services

after voice in mobile networks. The SMS Center (SMSC) is used as a test bed to optimize the usage of public

and private clouds in network operations, both in technology and business. This paper presents a hybrid cloud

architecture that enables an automatic up-and-down-scaling of the system, using dynamic resource provision-

ing and depending on the service load. In addition, a cost analysis to ﬁnd the optimal balance between public

and private clouds is described. Finally, the proposed solution is thoroughly evaluated, future research ideas

are highlighted and conclusions are drawn.

1 INTRODUCTION

It is a common belief (Murphy, 2010) that cloud com-

puting cannot yet offer the performance level that is

required in carrier grade mobile services. This is

partly true for real time services such as voice. On the

other hand, most other mobile operations can be exe-

cuted in clouds (Gabrielsson et al., 2010). Operators

already offer cloud computing services to their cus-

tomers in fact, but until now, very few of them have

transferred their own operations to cloud. However,

the situation is changing rapidly, and vendors in par-

ticular are evaluating opportunities in the cloud space.

The easiest options for applying clouds in the tele-

com industry can be found in back end systems such

as business support systems (BSS) (Raivio and Dave,

2011) and operational support systems (OSS), but

also the core services, including database (Paivarinta

and Raivio, 2011) or voice control systems (Venu-

gopal et al., 2011), can be implemented in clouds.

A Short Message Service (SMS) is a good ex-

ample of a service that can be run in a cloud. Text

messaging has been a globally very successful end

user service. Roaming agreements and standardized

technologies have guaranteed smooth interoperability

over operator borders, offering a good service expe-

rience for end users and economic success for oper-

ators. Although the SMS growth rate has stabilized

in the developed markets, the message volumes are

still increasing in the developing countries, and for

that reason system and cost optimizations are sought

after.

Telecommunication trafﬁc ﬂuctuates, and the

SMS trafﬁc is no exception. During daytime and ex-

ceptional events or incidents, messages are sent much

more frequently than in the night. In current SMS

systems, the capacity is designed on the basis of the

peak load and computation resources are wasted dur-

ing silent periods due to the dedicated nature of the

systems. Cloud computing can signiﬁcantly improve

the situation because virtual machines (VM) located

in the public cloud can be turned off when comput-

ing power is not anymore required and vice versa

(Khajeh-Hosseini et al., 2010). On the other hand, ex-

489

Raivio Y., Mazhelis O., Annapureddy K., Mallavarapu R. and Tyrväinen P..

HYBRID CLOUD ARCHITECTURE FOR SHORT MESSAGE SERVICES.

DOI: 10.5220/0003908104890500

In Proceedings of the 2nd International Conference on Cloud Computing and Services Science (CLOSER-2012), pages 489-500

ISBN: 978-989-8565-05-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

tensive usage of the public cloud can become expen-

sive. Furthermore, operators must support high avail-

ability and unconditional data security, and for those

reasons operators are forced to seek compromises.

One alternative to solve the described challenges

is called a hybrid cloud. The main idea here is to

utilize private and public clouds in parallel. In this

setup, the private cloud provides the base capacity,

while the public cloud can be used for optimal load

balancing. However, ﬁnding an optimal setup is not

an easy task. The solution should take into account

trafﬁc variations, and based on those, tune the system

to provide simultaneously high performance and low

cost. This paper focuses on studying the applicability

and feasibility of using the hybrid cloud in the context

of telecommunication services, particularly the SMS

case.

The paper is organized as follows. Section 2 intro-

duces the backgroundand related work of the research

along with a short review of SMS architecture, hybrid

cloud and dynamic resource provisioningmodel. Sec-

tion 3 describes in detail the proof of concept of the

SMS service in a cloud. Section 4 elaborates the cost

analysis, giving reasons for the selection of the op-

timal hybrid architecture. Next, Section 5 critically

evaluates the results using both reactive and proactive

resource provisioning models, and Section 6 summa-

rizes the work and proposes future research ideas. Fi-

nally, Section 7 offers paper’s conclusions.

2 BACKGROUND AND RELATED

WORK

This section explains the basics and related work of

the SMS architecture, SMS trafﬁc characteristics, the

hybrid cloud concept and also the autoscaling algo-

rithms utilized in the veriﬁcation of the results.

2.1 SMSC Architecture and Load

The key element of the text messaging system in mo-

bile networks is the Short Message Service Center

(SMSC). It is connected to the Mobile Switching Cen-

ter (MSC), the Visitor Location Register (VLR), the

Home Location Register (HLR), and further to the

application servers (AS) and the other SMSCs. Text

messages are ﬁrst sent to the closest MSC, which lo-

cates the home SMSC where the message is to be for-

warded. Depending on the content and address of the

message, it is sent either to the next SMSC, to the AS

or - if the receiver belongs to the same SMSC domain

- directly to the receiver. Basically SMSC acts as a

store-and-forward router. If the SMSC is overloaded,

the MSC will store the messages and will resend them

to the SMSC later. In the case when the message re-

ceiver suddenly is outside the network, the SMSC will

store the message and send it to the receiver the mo-

ment it joins the network again. The SMSC archi-

tecture is presented in Figure 1 (Mouly and Pautet,

1992).

Figure 1: SMSC architecture.

The SMSC load varies depending on the time of

day. During weekdays the load is normally very pe-

riodic, but at the weekend the load curve slightly

changes. Figure 2 shows an example of the SMS load

curve during a full week, starting from Monday and

ending on Sunday. The data has been collected from

an SMSC located on the network of an European op-

erator. The curve does not include the absolute vol-

ume, but for this research the knowledge of data vol-

ume dynamics is enough. The message volume can be

extrapolated for various markets when the total mes-

sage volume of the market is known.

Mon Tue Wed Thu Fri Sat Sun

Message volume

Time

Figure 2: Trafﬁc distribution during one week.

It is notable that the example trace does not show

any exceptional trafﬁc peaks that might occur dur-

ing accidents, special events and certain dates, such

as New Year’s Eve. It is estimated that the mes-

sage peaks ultimately can be as much as up to ten-

fold compared to the usual base load (Zerfos et al.,

2006). However, due to competitive communication

methods, text messaging peaks are not so common

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

490

anymore as they used to be before the introduction of

social networks and mobile email.

For evaluating the proof of concept and cost, the

real numbers for the load curve shown in Figure 2

should be added. Using the scale 0 to 6, the average of

the volume during the whole week is 3.07. According

to ITU-T

it was estimated that in 2010 6 trillion SMS

messages were sent. That number equals a continuous

load of slightly less than 200 000 messages per sec-

ond. On the other hand, an average mobile user sends

approximately 1000 messages per year

, which fact

means that a market with 10 million subscribers typi-

cally sends 10 billion SMS messages in a year. If the

numbers are applied to our example, we may set up

the value 1 to equal a rate of 100 messages per sec-

ond, which yields on average a rate of 307 messages

per second during a week and about 10 billion mes-

sages per year.

2.2 Hybrid Cloud

In the current SMSC systems, operators use dedicated

hardware and software solutions. The capacity of the

systems is tailored according to the maximum load,

but peak loads are usually excluded from the calcula-

tions. In the case of a sudden peak occurring, SMSCs

become overloaded and service quality deteriorates.

It means that the message delivery time will become

longer due to retransmissions between the MSCs and

SMSCs. Scaling the SMSC capacity on the basis of a

rare peak load does not make sense either, because it

would mean high and wasted capex costs.

Cloud computing can help to design a system

that automatically scales up and down depending on

the load. Public clouds support autoscaling func-

tions, and the pay-per-use principle will reduce the

capex costs. Given this background, a logical solution

would be to set up the whole system in a public cloud.

However, in the long run the public cloud CPU hourly

cost is more expensive than the same hour in the pri-

vate cloud (Walker, 2009). It is also mathematically

proved that the hybrid cloud is the most economical

solution upon variable trafﬁc load (Weinman, 2011).

The basic idea with a hybrid solution is simple.

A service provider invests in a private cloud that can

manage the performance critical applications, while

a public cloud can run background processes (Haj-

jat et al., 2010). However, our research focuses on

a dynamic hybrid cloud approach. In case a peak

load occurs, all excess trafﬁc is forwarded to VMs

http://www.itu.int/ITU-D/ict/material/

FactsFigures2010.pdf

http://epp.eurostat.ec.europa.eu/statistics explained/

index.php/Telecommunication statistics

located in the public cloud, while the VMs in the

private cloud deal with the base load. When the to-

tal load again drops below the capacity of the private

cloud, the public VMs are shut down. The system also

has to have an autoscaling functionality that can make

correct decisions on the load balancing between pri-

vate and public clouds, and which can also control the

VMs in the public cloud (Moreno-Vozmediano et al.,

2009).

The costs of a hybrid cloud are comprised of the

costs of the private and the public portions of the

cloud. Private cloud costs are incurred due to acquir-

ing, deploying and operating private cloud hardware

and software, and therefore it depends on the volume

of computing capacity acquired for the private cloud.

Public cloud capacity is charged on the basis of the

actual usage, and hence the public cloud cost depends

on the length of time when the public cloud capacity

is used.

Both the private and the public cloud cost de-

pend on how the demand for computing capacity is

divided between the pre-acquired infrastructure (pri-

vate cloud) and the consumed on-demand infrastruc-

ture (public cloud). Several approaches are avail-

able for cost-optimally distributing the load within

cloud (Martens and Teuteberg, 2011) and within a hy-

brid cloud, e.g. by optimally allocating computing

tasks to cloud resources at each time slot (Strebel and

Stage, 2010), or by determining the ﬁxed cost-optimal

threshold for the private cloud resources, and allocat-

ing the excess load to the public cloud (Mazhelis and

Tyrv¨ainen, 2011).

2.3 Dynamic Resource Provisioning

Performance unpredictability is one of the major chal-

lenges of cloud computing (Armbrust et al., 2010).

Autoscaling in the hybrid cloud is an important en-

abler for a high availability and a carrier grade per-

formance. Due the relatively long startup time of the

public cloud VMs, the system must launch new in-

stances well before they are required. However, de-

signing an automated resource controller is a complex

task. It is essential that we can accurately estimate the

workload handling capacity of an individual VM in

both private and public IaaS cloud environments. We

broadly categorize the resource controllers into reac-

tive and proactive models.

In addition to load monitoring, resource provi-

sioning can be based on cost-efﬁciency and some

other parameter, such as the deadline of application

execution (Mao and Humphrey, 2011; Van den Boss-

che et al., 2010), or performance (Chang et al., 2010).

Compared to our research setup, the previous ex-

HYBRIDCLOUDARCHITECTUREFORSHORTMESSAGESERVICES

491

amples include a set of interconnected applications,

while our work concentrates on analyzing a single

application that includes several components. The re-

sult is slightly different if the optimization is reviewed

from a service provider’s (SP) point of view. One pa-

per (Mazzucco et al., 2010) evaluates how the differ-

ent scaling strategies bring the best proﬁt for the SP,

and at the same time minimizes the energy consump-

tion. The energy part could be applied to the user of

the service, as well.

2.3.1 Reactive Model

In the reactive model the resource controller con-

stantly monitors the incoming message trafﬁc and

scales up the resources once a substantial increase in

the workload is detected. This approach works well

in theory, but in practice, especially during unpre-

dictable or dynamically varying workload peaks, it

fails to maintain the required Quality of Experience

(QoE). The reactive model suffers from the fact that

the resource controller will start the process of re-

source provisioning only when an increase in the load

is detected, even though it is known that a VM re-

quires at least 2 to 3 minutes to reach an operational

state. During the startup time we observed a steep in-

crease in the message processing time, a fact which

will create a heavy loss in QoE.

2.3.2 Proactive Model

Dynamic resource allocation schemes that adapt to

varying workloads will involve three phases at a min-

imum: extensive monitoring of incoming message

trafﬁc, workload modeling and subsequent decision

making on scaling resources. Proactive models, espe-

cially, need to characterize the incoming workload,

predict the next few steps and act according to the

prediction preemptively. Fortunately there are several

predictive models in use in various ﬁelds like econo-

metrics, machine learning and neural networks.

Research on using proactive models for autoscal-

ing in cloud computing is a hot topic, and interest-

ing results have been recently published. For exam-

ple, linear regression is applied to solve optimal cost

parameters for deploying business software to hybrid

cloud (Strebel and Stage, 2010). Probabilistic model

checker called PRISM evaluates the performance of

VM live migration (Kikuchi and Matsumoto, 2011).

The PRESS algorithm (Gong et al., 2010) uses sig-

nal processing and statistical methods to identify re-

peating patterns and discrete-time Markov chain to

predict load in the near future. Markov models suit

well to analyze http trafﬁc that is often heavy-tail

type (Pacheco-Sanchez et al., 2011). Finally, sta-

tistical methods such as autoregressive moving aver-

age (ARMA) (Roy et al., 2011) supported by Fourier

Transfer analysis (Zhang et al., 2009) are highly efﬁ-

cient in modeling stationary time series data.

The SMS trafﬁc fulﬁlls the stationary characteris-

tics, and for that reason we chose the ARMA model

(Box and Jenkins, 1990) for predicting the future

workload based on the recent history of the workload

pattern. The resource controller forecasts incoming

message trafﬁc every minute, and it is assumed to be

centralized and not distributed like in (Yazir et al.,

2010; Ardagna et al., 2011). We used a ﬁrst order

ARMA ﬁlter to model and predict the incoming mes-

sage trafﬁc rates. This ﬁlter is represented by the fol-

lowing equations:

X(t) = µ+ α× X(t − 1) + β × ε(t− 1) + ε(t) (1)

µ = Mean× (1− α) (2)

where

• µ is a constant;

• α and β are ARMA parameters, having values

0.90 and 0.15 respectively; and

• ε represents white noise.

In the experiments on Amazon EC2 Large that lo-

cated in the EU region, we observed that the time re-

quired to provision a VM varies between 80 to 90 sec-

onds. Taking this observation into account, we have

deﬁned our prediction interval as 120 seconds. The

results with the workload prediction have been highly

successful. Figure 3 shows the predicted workload

compared with the actual workload for one weekday,

using Figure 2 as input material. The results were

good, presenting in average 4.5% prediction error.

Figure 3: Workload prediction for one day.

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

492

3 PROOF OF CONCEPT

This section describes the experimental system and

algorithms of the resource controller in detail.

3.1 Experimental System

An experimental setup was created to review the func-

tionality of the hybrid cloud. In order to simulate the

features of the SMSC in a VM, an SMSC simulator

was bundled as a black box into a VM image. The

SMSC functionality was offered by a simulator called

SMPPSim

, which uses the Short Message Peer-to-

Peer (SMPP) protocol for communication. The vari-

able message load was created by OpenSMPP API

Next, the threshold capacity value for a VM was mea-

sured. The measurement results using one EC2 Large

instance are shown in Figure 4. The results indi-

cate that the time required for processing a message

increases dramatically for a workload above 100 re-

quests per second. Each time there was a substantial

increase in the workload, QoE worsened due to the

ineffective reactive resource provisioning approach.

0.5

1.5

2.5

60 80 100 120 140

Processing time per message (s)

Workload (Request/s)

SMS request processing time

Figure 4: SMS message processing time with EC2 Large.

The optimal hybrid solution can be calculated

when the capacity and CPU hourly costs of the private

and public clouds are known. In the setup, Amazon

EC2

was chosen for the public cloud provider due

to the rich feature set, while the private cloud com-

ponent was taken from OpenNebula

, which offers a

superior openness as well as good integration capa-

bilities with several public clouds. The private cloud

includes a Load Balancer that is implemented using

an open source software called HAProxy

. The Load

http://www.seleniumsoftware.com/

http://sourceforge.net/projects/smstools/

http://aws.amazon.com/ec2/

http://opennebula.org/

http://haproxy.1wt.eu/

Balancer can distribute the SMPP messages to the un-

derlying VMs located both in the private cloud and

public cloud.

The Resource Controller includes two blocks: the

Monitor software based on the TCPSTAT

, and an op-

tional Forecast functionality. The usage of the Fore-

cast block is described in the following subsection.

The Monitor reports periodically the load detected by

the Load Balancer to the Resource Allocator that is

implemented on the XML-RPC

client. Based on the

received information, the Resource Allocator, help

from the Virtual Infrastructure Manager, scales up or

down the resources available to the system. A virtual

bridge inside the private cloud enables communica-

tion between the SMSCs, while on the other hand, the

Load Balancer provides a gateway to the public net-

work. Figure 5 presents the architecture for the exper-

imental setup.

Figure 5: Hybrid cloud architecture for SMSC.

Initially VMs are launched in the private cloud un-

til the overall maximum capacity is achieved. When

that happens, the front end, implemented on the

OpenNebula platform, starts new VMs in the public

cloud, thus forming the hybrid cloud. If there is a

need to scale the resources down, the front end shuts

down the VMs in the public cloud. Before the SMSC

is shutdown in the EC2, the undelivered messages are

sent forward or saved to the private cloud. Due to

privacy issues and government regulations, it is sug-

gested that VMs are launched only in regions that

meet the requirements of the authorities (Kaufman,

2009).

3.2 Algorithm

In order to improve the performance and increase the

high availability of the SMSCs, a framework is de-

http://www.frenchfries.net/paul/tcpstat/

http://www.xmlrpc.com/

HYBRIDCLOUDARCHITECTUREFORSHORTMESSAGESERVICES

493

signed by assuming the capacity of an SMSC simu-

lator. In this case, we know that an SMSC simulator

can process up to λ

inst

messages per second within a

given service time. If there is a ﬂow of more than

inst

, then there would be a delay introduced by the

simulators, thus increasing the service time for each

message. Algorithm 1 explains the implementation

details of the Resource Controller and Algorithm 2

describes the functionality of the Forecast module.

Algorithm 1 initially starts a ﬁxed number of VMs

in a private cloud. The monitoring management sys-

tem tracks the load for every k seconds. Depending

on the recent history of the workload, the Forecast

module predicts the future workload. The module

utilizes the ARMA prediction algorithm described in

Subsection 2.3.2. In the ﬁrst step we must stationarize

the non-stationary time series to stationary one. The

output is used as the input to the algorithm, and the

resulting value is de-stationarized to obtain the pre-

dicted workload for the next prediction interval. De-

pending on the result, the required and also optimum

number of VMs are deduced using the map function.

Table 1 describes the variables used in the algorithms.

Algorithm 1: Resource Controller.

1: start VM in C

2: VM

cur

← VM

3: while true do

4: monitor W

(n);

5: W

(n+ 1) ← Forecast(W

(n));

6: VM

← map(W

(n+ 1));

7: if VM

cur

< VM

then

8: scale up;

9: else if VM

cur

> VM

then

10: scale down;

11: end if

12: sleep k;

13: end while

Algorithm 2: Forecast.

1: W

(n) ← stationarize(W

(n));

2: µ = M × (1− α);

3: ε(n) = W

(n− 1) − β× ε(n− 1);

4: W

(n+ 1) = µ + α ×W

(n) + β× ε(n);

5: W

(n+ 1) ← destationarize(W

(n+ 1));

4 COST ANALYSIS

Several approaches are available for cost-optimally

distributing the load within a hybrid cloud, e.g. by op-

timally allocating computing tasks to cloud resources

Table 1: List of variables used in Algorithms 1 and 2.

Variables Description

(n) Actual workload till n

instant

(n) Actual workload at n

instant

(n) Stationarized workload at n

in-

stant

(n+ 1) Forecasted workload at (n+1)

instant

Private cloud

Public cloud

map Returns VM

cur

Number of running VMs

Fixed number of running VMs

in Private cloud

Optimum number of VMs

k Monitoring interval

(n) Stationarized actual workload

values

α AR coefﬁcient

β MA coefﬁcient

µ ARMA constant

ε White Noise

at each time slot (Strebel and Stage, 2010), or by de-

termining the ﬁxed cost-optimal threshold for the pri-

vate cloud resources, and allocating the excess load

to the public cloud (Mazhelis and Tyrv¨ainen, 2011).

In this section, we apply the last approach, which is

more lightweight and whose outcomes are easier to

interpret, in order to determine the cost-optimal divi-

sion of the SMSC’s computing capacity between the

private and the public portions of a hybrid cloud.

4.1 Seeking Cost-efﬁcient Division of

Load in Hybrid Cloud

The unit of the above division of load is a comput-

ing instance that corresponds to a processor core (in

the private cloud) or to an Amazon Large EC2 in-

stance (in the public cloud). For the sake of simpli-

fying the analysis, the cores are conﬁgured to approx-

imately match the performance characteristics of the

EC2 Large instances.

When the SMS request arrival rate grows, the in-

stances in the private portion of the hybrid cloud are

allocated ﬁrst. In the event that all of the available

private instances are already allocated, the public in-

stances are launched. Alternatively, when the SMS

request arrival rate decreases, the public instances are

released ﬁrst, after which the private instances are

freed. Note that the decommissioning of the unused

public instances is done at the end of the charging

period (i.e., in case of EC2, at the end of an hour)

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

494

to avoid extra charging if two or more bursts occur

within the same charging period.

Given the maximum SMS request arrival rate

max

, the maximum number of required instances can

be estimated as:

max

= ⌈

max

inst

⌉. (3)

Denote the number of acquired private instances

as n

. Then, the threshold portion of SMS requests

processed by the private cloud is

q =

max

. (4)

While the SMS request arrival does not exceed the

threshold

= qn

max

inst

(5)

these requests are processed by the private instances.

When the arrival rate grows beyond λ

, the excessive

requests are processed using public instances.

Thus, the costs of the private and the public por-

tions of the cloud depend on the threshold q. We will

therefore look for the value of q, minimizing the over-

all costs: q

min

= min

According to the literature (Mazhelis and

Tyrv¨ainen, 2011), the cost-optimal time of using the

public cloud τ

min

) can be estimated as

min

) =

+ p

k)T

+ p

k(u

+ ρ(1 + u

))

, (6)

where

• p

is the unit price of the privatecloud computing

capacity;

• p

is the unit price of the private cloud data trans-

fer;

• u

= p

(where p

is the unit price of the

public cloud computing capacity) is the premium

charged by the public cloud provider for the use

of its on-demand computing capacity;

• u

= p

(where p

is the unit price of

the public cloud data transfer) is is the premium

charged by the public cloud provider for the use

of its data communications capacity;

• k is the volume of the data transferred between the

hybrid cloud and the clients during the time a sin-

gle unit of the computing capacity is consumed;

• ρ is the ratio of the data transferred between the

private and the public portions of the cloud to the

data transferred between the public cloud and the

clients; and

• T is the overall time during which the costs are

estimated (e.g. the depreciation period).

Eq. 6 indicates that the cost-optimal time of using

the public cloud decreases i) as the premium charged

by the cloud provider (u

and u

) grows, and ii) as the

intensity of data communications (k and ρ) increases.

As was shown in (Mazhelis and Tyrv¨ainen, 2011),

is a decreasing function of q, and therefore there

exists an inverse function q

min

= Q(τ

) that can be

used to obtain the value of q

min

. Thus, having found

the value of τ

, and using the knowledge of the SMS

request distribution to estimate Q(τ

), the value of

min

can be determined.

4.2 Cost Assessment

As discussed above, the hybrid cloud cost is the sum

of the costs of the private cloud and the public cloud

portions. The private cloud cost is independent of the

rate of requests at each instant; rather, the cost is de-

termined by the number of virtual cores allocated (i.e.

acquired or leased) for the purpose of serving as an

SMSC. The cost of the public cloud, on the other

hand, depends on the volume of consumed capacity

(CPU-hours, transferred data, etc.).

The cost of the private portion of the hybrid cloud

encompasses the costs of acquiring hardware and

software, as well as the costs of rack space, power,

allocated bandwidth, and the technical support costs.

These costs are estimated based on the pricing for vir-

tual servers offered by a hosting service enterprise

Nebula

. Speciﬁcally, the cost of a computing in-

stance running Linux Debian5 OS, 2 GB of memory

and a 120 GB hard disk supporting automated back-

ups and high availability is estimated to be C

= $233

per month.

The costs accounted for in the public portion of

the hybrid cloud include (based on the Amazon EC2

pricing (Ireland) for on-demand instances, effective

on November 7, 2011):

• Consumption-based cost of on-demand comput-

ing capacity (no reservation is assumed) C

$0.38 per hour per instance;

• Volume-based data transfer costs assuming V

1680B and V

OUT

= 1322B of data transmitted per

SMS request, respectively, into and out of the pub-

lic cloud: C

= $0 and (since volume is below

10TB per month) C

OUT

= $0.12per GB;

• Volume-based storage costs (200 bytes per re-

quest, $0.11 per GB);

• Volume-based costs of I/O operations (4 I/O per

request, $0.11 per million I/O requests);

http://www.nebula.ﬁ/

HYBRIDCLOUDARCHITECTUREFORSHORTMESSAGESERVICES

495

• Additional costs incurred due to the use of Elastic

IPs, Load Balancing and CloudWatch.

In the real SMSC system, successfully delivered

SMS messages are usually not stored. According

to (Zerfos et al., 2006), 73% of SMS messages are

delivered successfully within 10 seconds, implying

that 27% of the messages should be stored for re-

sending until they are successfully delivered. How-

ever, most of the messages are delivered successfully

within a few minutes, which means that the volume-

based storage and I/O operation requirements related

to the message deliveries are very minor. The same

comment applies to the storage costs of the VMs. The

size of the VM is on the level of 100 GBs, resulting in

only a price of $11 per VM in a month.

Likewise, the allocation of Elastic IPs with the

price of $0.01 per hour incurs a minor yearly cost

below $90 per IP. Load balancing or CloudWatch

monitoring services will not be needed because the

load balancing function was implemented outside of

the EC2 domain and because the monitoring software

was provided by the open source components. There-

fore, only the cost of on-demand computing capacity

and the data transfer costs will be taken into account

when estimating the public cloud costs, whereas the

other costs are assumed minor or absent and hence

are ignored.

Finally, the maximum SMS arrival rate to be han-

dled by an instance was estimated empirically by ex-

perimenting with the SMSC simulator. It was found

that, should the SMS arrival rate grow above 100 mes-

sages per second, the time of handling a message

would exceed the acceptable limit. Therefore, the

maximum SMS arrival rate per core is assigned the

value of λ

inst

= 100.

Based on the assumptions and arguments above,

the pricing parameters can be estimated as follows.

The price of private computing capacity is esti-

mated per core per hour as:

24× 30

= 0.32$/hour. (7)

The price of public computing capacity is esti-

mated based on the Amazon EC2 pricing as:

= 0.38$/hour. (8)

Furthermore, based on (Hamilton, 2010), we can

assign

= 0.04$/GB, (9)

and, based on Amazon EC2 pricing,

bcIN

= 0$/GB, (10)

bcOUT

= 0.12$/GB. (11)

Since in the case of the SMSC the incoming trafﬁc

is approximately equal to the outgoing,

= (p

bcIN

+ p

bcOUT

)/2 = 0.06$/GB. (12)

From the above, it follows that

= p

= 0.38/0.32 = 1.2; (13)

= p

= 0.06/0.04 = 1.5. (14)

Since the pricing of the computing capacity is

charged on an hourly basis, we shall estimate a value

for k that reﬂects the volume of transferred data (in

GB) per unit of computing capacity over a period of

an hour too, i.e. T

CPU

= 3600 sec. Furthermore, while

processing an SMS request, V

ext

= 1160 B of data is

transferred between the SMSC and the receiver, and

int

= 1842 B is transferred between the sender and

the SMSC. Thus,

k = λ

inst

CPU

ext

= 100 × 3600×

1160

1073741824

= 0.39GB per instance-hour, and (15)

ρ =

int

ext

1842

1160

= 1.59. (16)

Now, having estimated the pricing parameters, let

us go on to identify the cost-efﬁcient division of the

SMS demand. From eq. (6) we get:

min

) =

(0.32+ 0.04 × 0.39) × 24× 30

1.2× 0.32+ 0.04× 0.39(1.5+ 1.59× 2.5)

241.6

0.47

= 515. (17)

Thus, in order to achieve a cost-optimal divi-

sion of SMS demand out of 24 × 30 = 720 hours

in month, the public cloud instances should be used

during the “busiest” 515 hours (i.e. the hours with

the highest load). These public instances are em-

ployed to serve the load exceeding the threshold value

= qn

max

inst

, whereas during the remaining (720−

515 = 205) hours the private cloud capacity alone is

sufﬁcient to handle the load.

Using eq. (3) and an assumption that the max-

imum message rate is 500 messages/second, the

total number of instances is estimated as n

max

⌈500/100⌉ = 5. Furthermore, based on the load

distribution rearranged to be monotonically non-

decreasing (see Figure 6), we get the λ

correspond-

ing to 205 hours:

= 190requests per second,

from which it follows that

= q

min

max

= λ

/λ

inst

= 190/100 ≈ 2, (18)

and hence

min

= n

max

= 2/5 = 0.4. (19)

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

496

0 0.2 0.4 0.6 0.8 1

2000

4000

6000

8000

10000

12000

14000

Threshold of private cloud (q)

Cost, $

Costs of public cloud

CPU

Trf.In

Trf.Out

0 0.2 0.4 0.6 0.8 1

2000

4000

6000

8000

10000

12000

14000

Threshold of private cloud (q)

Cost, $

Costs of private cloud

Private total

0 0.2 0.4 0.6 0.8 1

2000

4000

6000

8000

10000

12000

14000

Threshold of private cloud (q)

Cost, $

Total costs of public and private cloud

Public

Private

Total

Figure 7: The yearly costs of the public portion (left) and the private portion (middle) of the hybrid cloud, as well as the

overall hybrid cloud costs (right) plotted as a function of the private cloud threshold q.

0 200 400 600 800

100

150

200

250

300

350

400

450

500

Rearranged request distribution

Time (t) in hours

Requests per second (λ)

Figure 6: SMS request distribution, rearranged to be mono-

tonically non-decreasing.

Thus, the cost-optimal private portion of the hy-

brid cloud is q

min

= 0.4, i.e. n

= 2 private computing

instances should be acquired or leased for the purpose

of SMSC. Whenever the SMS load is greater than the

private cloud can efﬁciently handle, an instance from

the public cloud needs to be launched.

This cost-optimal SMS load division is also vis-

ible in Figure 7, where the costs of the private and

the public portions of the hybrid cloud, as well as

the overall hybrid cloud costs are shown for differ-

ent values of q. Note that the costs in Figure 7 are

shown for a period of one year. The fully public cloud

(q = 0) and the fully private cloud (q = 1) incur the

costs of $13 396 and $13 980 per year, respectively,

whereas the yearly cost of the cost-optimal load divi-

sion is $12 190. As can be seen, the overall hybrid

cloud costs C is a (not strictly) convex function, and

its minimum occurs when q

min

= 0.4, thus matching

the results of the above estimations.

5 RESULT ANALYSIS

According to Section 4 the cost optimal hybrid cloud

should have a composition where 40% of all trafﬁc

should be executed by a private cloud and 60% by a

public cloud. Using our previous example, we note

that originally we should have 2 private instances, but

we must be prepared to update the system with 3 pub-

lic cloud instances to cover the maximum load of 5

VMs or 500 requests per second. Table 2 presents

number of required VMs for each load range deﬁned

by the map function presented in Algorithm 1. In our

case the functionality of the map module is very sim-

ple and there is no need for a complex analysis as we

should do in a general case (Mishra and Sahoo, 2011).

Table 2: VM capacity boundary values.

Workload (Req/s) Number of VMs Type

[0,200) 2 private

[200, 300) 3 hybrid

[300, 400) 4 hybrid

[400, 500) 5 hybrid

To evaluate the results from the performance point

of view, the dynamic resource provisioning models

were applied to test the average message processing

time during one day. Figure 8 presents the review re-

sults using both the reactive and proactive models. As

shown in Figure 3 and Table 2, new resource alloca-

tions are needed three times per day when the exam-

ple numbers are applied. Due to the startup time of

a new public cloud instance, the reactive model suf-

fers from an increased processing delay always when

a new public cloud instance is launched, while the

proactive model offers a stable performance. The re-

sults slightly improve when the number of instances

increases because at the same time the average load

per VM decreases. A lower load in VM results in a

HYBRIDCLOUDARCHITECTUREFORSHORTMESSAGESERVICES

497

better performance which fact is veriﬁed by Figure 4.

Our approach concentrates in the direct computa-

tion costs, and ignores several other factors (Martens

and Teuteberg, 2011). In addition, the research is only

dedicated to a black box approach and does not look

at the dependencies between the sub modules (Mao

and Humphrey, 2011; Van den Bossche et al., 2010).

On the other hand, the presented algorithm is novel,

and unlike the previous research (Chang et al., 2010;

Yazir et al., 2010), the special requirements of the hy-

brid cloud in the telecommunication context are high-

lighted in the algorithm.

Figure 8: Review of hybrid cloud using reactive and proac-

tive models.

6 DISCUSSION

The text messaging service was selected as the use

case for good reasons. SMS is a widely spread ser-

vice, and it can be well modeled and simulated. How-

ever, the SMS use case is not the most illustrative

from the viewpoint of the hybrid cloud. This is due to

the fact that the messaging trafﬁc is periodic and the

absolute differences in volume between minimum,

base and peak loads are not huge. On the other hand,

extreme peak workloads are very rare. In addition, the

actual message size is moderate, accumulating a very

small transmission load. The computation power re-

quired to manage an average SMS market segment is

not large either. The advantages of the hybrid cloud

will not be clearly visible in case of low trafﬁc volume

or variation. The message size as such is not very crit-

ical parameter (Hill and Humphrey, 2009).

The results can be applied to other services as

well. In the mobile sector, the Multimedia Messag-

ing Service (MMS) would be a logical extension to

the use case. Unfortunately the MMS is still in the

early market phase, and it is not yet well supported by

the open source software community, creating chal-

lenges for the implementation of proof of concepts.

Other interesting application examples are the global

video on demand (VoD) (Li et al., 2011) and video-

conference services that also require transcoding fea-

tures (Cervi˜no et al., 2011). Volumes vary a lot in

video service trafﬁc (Li et al., 2011), and transmission

load is signiﬁcant. These facts enable a hybrid cloud

architecture that can be geographically distributed.

The cost analysis has certain limitations. First

of all, the cost analysis of the Amazon EC2 did not

include prices of the reserved instances or the spot

prices. Secondly, the model does not take into ac-

count the message retransmissions between MSCs

and SMSC. According to the measurements (Zerfos

et al., 2006), about 5% of the sent messages will never

reach the target, and one or more retransmissions are

needed in 22% of the cases. However, the perfor-

mance of the latest SMSC systems can be expected

to be a lot higher. Recently an operator reported a

peak capacity of 19 000 messages per second

was

delivered without congestion.

Furthermore, the costs of the load balancer have

not been taken into account since these costs are con-

stant. The cost of data transfer in the private portion

of the hybrid cloud was assumed to be constant, too,

independently of the occupied bandwidth, whereas a

bandwidth-based charging was assumed in (Mazhelis

and Tyrv¨ainen, 2011). This may make the estimate of

cost-optimal load division somewhat incorrect; how-

ever, since the data transfer costs in the private cloud

are relatively low, also the inaccuracy of the estima-

tion can be ignored.

The SMS arrival rate is assumed to be monitored

on a per-minute basis, and the public cloud instances

are launched or released depending on whether the ar-

rival rate exceedsa certain threshold. However, public

cloud computing instances are charged on an hourly

basis, and hence, even if the period of high load -

when the public instance is needed - lasts for a couple

of minutes only, the whole hour will be charged. As a

result, the evaluation of the cost-optimal load division

somewhat underestimated the cost of public cloud in-

stances; however, as the SMS distribution function

rarely has sharp peaks, the underestimation can also

be ignored.

Telecommunication operators can seek further re-

ductions in costs by introducing a private cloud that

can be shared between several operators. It can be

also foreseen that infrastructure vendors will start to

offer public, telecom speciﬁc clouds to their cus-

tomers. The telecom clouds can be optimized for real

time applications and their SLA requirements. Green

http://www.itu.int/ITU-D/ict/newslog/CategoryView,

category,SMS.aspx

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

498

cloud computing is another interesting research path

that should be analyzed in the hybrid context (Maz-

zucco et al., 2010). These research ideas are left for

future studies.

Finally, the usage of dynamic resource provision-

ing models in cloud computing deserves more re-

search. The algorithms were not in the focus of this

paper, but the predictive ARMA model was chosen to

test the feasibility of our hybrid cloud solution. In the

SMSC use case the performance of the ARMA algo-

rithm was outstanding with an average error of 4.5%,

but it remains to be seen how well it would operate

with a random load that also includes a varying num-

ber of trafﬁc peaks.

7 CONCLUSIONS

We have presented a novel Short Message Service

Center (SMSC) architecture that is implemented in

a hybrid cloud using Amazon EC2 as a public and

OpenNebula as a private cloud. The hybrid cloud en-

ables an efﬁcient usage of computation resources de-

pending on the load curve. The research is supported

with an experimental setup that is implemented using

an SMSC simulator. The proof of concept is comple-

mented with a dynamic resource provisioning system

that can utilize either reactive or proactive algorithm.

The measurement data, gathered from the proof of

concept, is further applied to verify the key param-

eters needed for the cost analysis section. Finally, the

cost analysis proves that a hybrid cloud is a feasible

alternative to support services where the trafﬁc peaks

are frequent.

However, due to the trafﬁc characteristics of the

SMS service, little beneﬁt can be obtained from the

hybrid approach in this particular use case. The SMS

trafﬁc varies relatively little, and furthermore, the

message complexity and volume are on a minor scale.

On the other hand, the SMS was only used here as a

test case, and better results can be achieved with more

dynamic trafﬁc types, such as multimedia and video.

In addition, the experimental system proves that a dy-

namic resource controller that uses prediction tech-

niques, can signiﬁcantly improve the system perfor-

mance and in turn lower the costs. The research indi-

cates that the dynamic hybrid cloud is an interesting

alternative to static, either private or public, clouds.

The optimal architecture can ultimately be decided af-

ter a careful analysis of the trafﬁc pattern.

ACKNOWLEDGEMENTS

We thank the anonymous reviewers for their valu-

able comments. The work is supported by Tekes

(the Finnish Funding Agency for Technology and In-

novation, www.tekes.ﬁ) as a part of the Cloud Soft-

ware Program (www.cloudsoftwareprogram.org) of

Tivit (Strategic Centre for Science, Technology and

Innovation in the Field of ICT, www.tivit.ﬁ).

REFERENCES

Ardagna, D., Casolari, S., and Panicucci, B. (2011). Flex-

ible distributed capacity allocation and load redirect

algorithms for cloud systems. In Proc. 4th Interna-

tional Conference on Cloud Computing, Cloud 2011,

pages 163–170. IEEE.

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. Communications of the ACM, 53(4):50–

58.

Box, G. E. P. and Jenkins, G. (1990). Time Series Analysis,

Forecasting and Control. Holden-Day, Incorporated.

Cervi˜no, J., Escribano, F., Rodr´ıguez, P., Trajkovska, I., and

Salvach´ua, J. (2011). Videoconference capacity leas-

ing on hybrid clouds. In Proc. 4th International Con-

ference on Cloud Computing, Cloud 2011, pages 340–

347. IEEE.

Chang, F., Ren, J., and Viswanathan, R. (2010). Optimal re-

source allocation in clouds. In Proc. 3rd International

Conference on Cloud Computing, Cloud 2010, pages

418–425. IEEE.

Gabrielsson, J., Hubertsson, O., M´as, I., and Skog, R.

(2010). Cloud computing in telecommunications. Er-

icsson Review, 1:29–33.

Gong, Z., Gu, X., and Wilkes, J. (2010). PRESS: PRedictive

Elastic ReSource Scaling for cloud systems. In Proc.

6th International Conference on Network and Service

Management, CNSM 2010, pages 9–16. IEEE.

Hajjat, M., Sun, X., Sung, Y.-W. E., Maltz, D., Rao, S.,

Sripanidkulchai, K., and Tawarmalani, M. (2010).

Cloudward bound: planning for beneﬁcial migration

of enterprise applications to the cloud. In Proc. ACM

SIGCOMM 2010, SIGCOMM ’10, pages 243–254.

ACM.

Hamilton, J. (2010). Cloud computing economies of scale.

Keynote at AWS Genomics & Cloud Computing

Workshop, Seattle, WA, 08.06.2010, available from

http://www.mvdirona.com/jrh/TalksAndPapers/James

Hamilton GenomicsCloud20100608.pdf.

Hill, Z. and Humphrey, M. (2009). A quantitative anal-

ysis of high performance computing with Amazon’s

EC2 infrastructure: The death of the local cluster? In

Proc. 2009 10th IEEE/ACM International Conference

on Grid Computing, GRID, pages 26–33. IEEE.

Kaufman, L. M. (2009). Data security in the world of cloud

computing. IEEE Security and Privacy, 7:61–64.

HYBRIDCLOUDARCHITECTUREFORSHORTMESSAGESERVICES

499

Khajeh-Hosseini, A., Greenwood, D., Smith, J. W., and

Sommerville, I. (2010). The cloud adoption toolkit:

Supporting cloud adoption decisions in the enterprise.

CoRR, abs/1008.1900.

Kikuchi, S. and Matsumoto, Y. (2011). Performance mod-

eling of concurrent live migration operations in cloud

computing systems using PRISM probabilistic model

checker. In Proc. 4th International Conference on

Cloud Computing, Cloud 2011, pages 49–56. IEEE.

Li, H., Zhong, L., Liu, J., Li, B., and Xu, K. (2011). Cost-

effective partial migration of VoD services to con-

tent clouds. In Proc. 4th International Conference on

Cloud Computing, Cloud 2011, pages 203–210. IEEE.

Mao, M. and Humphrey, M. (2011). Auto-scaling to min-

imize cost and meet application deadlines in cloud

workﬂows. In Proc. 2011 International Conference

for High Performance Computing, Networking, Stor-

age and Analysis, SC ’11, pages 49:1–49:12. ACM.

Martens, B. and Teuteberg, F. (2011). Decision-making in

cloud computing environments: A cost and risk based

approach. Information Systems Frontiers, pages 1–23.

Mazhelis, O. and Tyrv¨ainen, P. (2011). Role of data com-

munications in hybrid cloud costs. In Proc. 37th EU-

ROMICRO Conference on Software Engineering and

Advanced Applications, pages 138–145. IEEE.

Mazzucco, M., Dyachuk, D., and Deters, R. (2010). Max-

imizing cloud providers’ revenues via energy aware

allocation policies. In Proc. 3rd International Confer-

ence on Cloud Computing, Cloud 2010, pages 131–

138. IEEE.

Mishra, M. and Sahoo, A. (2011). On theory of VM place-

ment: Anomalies in existing methodologies and their

mitigation using a novel vector based approach. In

Proc. 4th International Conference on Cloud Comput-

ing, Cloud 2011, pages 275–282. IEEE.

Moreno-Vozmediano, R., Montero, R. S., and Llorente,

I. M. (2009). Elastic management of cluster-based

services in the cloud. In Proc. 1st workshop on Auto-

mated control for datacenters and clouds, ACDC ’09,

pages 19–24. ACM.

Mouly, M. and Pautet, M.-B. (1992). The GSM System for

Mobile Communications. Telecom Publishing.

Murphy, M. (2010). Platform-as-a-Service for Telcos.

Keynote at Cloud Asia, Singapore, 05.05.2010, avail-

able from http://cloudasia.ngp.org.sg/2010/programs

track2.php.

Pacheco-Sanchez, S., Casale, G., Scotney, B. W., McClean,

S. I., Parr, G. P., and Dawson, S. (2011). Markovian

workload characterization for QoS prediction in the

cloud. In Proc. 4th International Conference on Cloud

Computing, Cloud 2011, pages 147–154. IEEE.

Paivarinta, R. and Raivio, Y. (2011). Performance evalua-

tion of NoSQL cloud database in a telecom environ-

ment. In Proc. 1st International Conference on Cloud

Computing and Services Science, Closer 2011, pages

333–342. SciTePress.

Raivio, Y. and Dave, R. (2011). Cloud computing in mobile

networks - case MVNO. In Proc. 15th International

Conference on Intelligence, ICIN 2011, pages 253–

258. IEEE.

Roy, N., Dubey, A., and Gokhale, A. (2011). Efﬁcient

autoscaling in the cloud using predictive models for

workload forecasting. In Proc. 4th IEEE International

Conference on Cloud Computing, Cloud 2011. IEEE.

Strebel, J. and Stage, A. (2010). An economic deci-

sion model for business software application deploy-

ment on hybrid cloud environments. In Schumann,

M., Kolbe, L. M., Breitner, M. H., and Frerichs, A.,

editors, Multikonferenz Wirtschaftsinformatik 2010,

pages 195–206. Universit¨atsverlag G¨ottingen.

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

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

for deadline constrained workloads. In Proc. 3rd In-

ternational Conference on Cloud Computing, Cloud

2010, pages 228–235. IEEE.

Venugopal, S., Li, H., and Ray, P. (2011). Auto-scaling

emergency call centres using cloud resources to han-

dle disasters. In Proc. 19th International Workshop

on Quality of Service, IWQoS ’11, pages 34:1–34:9.

IEEE Press.

Walker, E. (2009). The real cost of a CPU hour. Computer,

42(4):35–41.

Weinman, J. (2011). Mathematical proof of the inevitability

of cloud computing. Online report, 08.01.2011, avail-

able from http://www.joeweinman.com/Resources/

Joe Weinman Inevitability Of Cloud.pdf.

Yazir, Y. O., Matthews, C., Farahbod, R., Neville, S., Gui-

touni, A., Ganti, S., and Coady, Y. (2010). Dynamic

resource allocation in computing clouds using dis-

tributed multiple criteria decision analysis. In Proc.

3rd International Conference on Cloud Computing,

Cloud 2010, pages 91–98. IEEE.

Zerfos, P., Meng, X., Wong, S. H., Samanta, V., and Lu,

S. (2006). A study of the short message service of a

nationwide cellular network. In Proc. 6th ACM SIG-

COMM conference on Internet measurement, IMC

’06, pages 263–268. ACM.

Zhang, H., Jiang, G., Yoshihira, K., Chen, H., and Saxena,

A. (2009). Intelligent workload factoring for a hybrid

cloud computing model. In Proc. 2009 Congress on

Services - I, pages 701–708. IEEE.

CLOSER2012-2ndInternationalConferenceonCloudComputingandServicesScience

500