PERFORMANCE EVALUTION OF NOSQL CLOUD DATABASE

IN A TELECOM ENVIRONMENT

Rasmus Paivarinta

Ixonos Plc, P.O. Box 284, FI-00811 Helsinki, Finland

Yrjo Raivio

Department of Computer Science and Engineering, Aalto University, P.O. Box 15400, FI-00076 Aalto, Finland

Keywords: NoSQL, IaaS, Home location register, Performance, SLA.

Abstract: Although the cloud computing paradigm has emerged in several ICT areas, the telecommunication sector is

still mainly using dedicated computer units that are located in operators’ own premises. According to the

general understanding, cloud technologies still cannot guarantee carrier grade service level. However, the

situation is rapidly changing. First of all, the virtualization of computers eases the optimization of

computing resources. Infrastructure as a Service (IaaS) offers a complete computation platform, where

instances can be hosted locally, remotely or in a hybrid fashion. Secondly, NoSQL (Not only SQL)

databases are widely used in the internet services, such as Amazon and Google, but they are not yet applied

to telecom applications. This paper evaluates, whether cloud technologies can meet the carrier grade

requirements. IaaS cloud computing platforms and HBase NoSQL database system are used for

benchmarking. The main focus is on the performance measurements utilizing a well known home location

the main measurements are carried out in Amazon Elastic Compute Cloud (EC2). The discussion section

evaluates and compares the results with other similar research. Finally, the conclusions and proposals for the

next research steps are given.

1 INTRODUCTION

Telecommunications operators are used to running

their embedded computer systems on proprietary

platforms. Typically operators have not shared

infrastructures either, but have purchased their own

networks. However, this situation is slowly

changing. The first step has been taken by the

Mobile Virtual Network Operators (MVNO), who

have outsourced some part, or even the whole

network, to network vendors. MVNOs have also

utilized shared radio access networks (RAN) to

avoid high initial investment costs. Recently, due to

saturated revenues, cost pressures on operability and

introduction of flat network architectures, such as

Long Term Evolution (LTE), also dominant

operators have shown interest in network sharing

initiatives.

Cloud computing offers a new perspective on

mobile network optimization. Unlike the past,

mobile networks are based on commercial

computers equipped with the Linux operating

system. Parallel to this, CPU and data storage

performance are still developing almost

exponentially (Armbrust, et al., 2009). This

paradigm shift will open novel opportunities for

cloud technologies in the telecom sector

(Gabrielsson, et al., 2010).

It is probable that mobile networks will not

change from private and proprietary servers to

public, generic purpose computers in the short term,

but telecom networks definitely include areas where

cloud options can have a role. Especially mobile

application servers and backend support systems

might suit cloud computing well. The main drivers

for the successful introduction of cloud technologies

imply a large variation in traffic patterns or massive

data volumes. In addition, telecommunication

networks are normally designed based on the peak

load, meaning that during off-peak periods systems

333

Paivarinta R. and Raivio Y..

PERFORMANCE EVALUTION OF NOSQL CLOUD DATABASE IN A TELECOM ENVIRONMENT.

DOI: 10.5220/0003384703330342

In Proceedings of the 1st International Conference on Cloud Computing and Services Science (CLOSER-2011), pages 333-342

ISBN: 978-989-8425-52-2

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

have a lot of unused capacity. Cloud computing thus

offers a natural technology for resource sharing.

However, there are still concerns whether cloud

computing meets the carrier grade requirements

(Murphy, 2010). Service level agreements (SLA),

such as high availability (HA), latency and

transactions per second, are strict in several telecom

services. One of the most critical mobile network

functionalities is the home location register (HLR).

HLR is the core element of the mobile system, and

fore example, is responsible for subscriber

authentication and roaming functionalities. HLR

also incorporates a risk for single point of failure,

resulting in very high SLA requirements.

The paper evaluates, whether cloud technologies

can meet the strictest telecom SLA requirements.

HLR is used as a use case, although it is clear that

the HLR, being the crown jewel of the operator, will

not be the first functionality that operators would

outsource to the cloud. However, the HLR presents

exact SLA requirements, and also benchmark data

and tools are available from the existing systems.

HLR behavior in the cloud is studied by using two

cloud technologies. First of all, all computation is

placed into the Infrastructure as a Service (IaaS).

IaaS can be applied locally, remotely and in both

ways, referring to private, public and hybrid clouds,

respectively. Secondly, the HLR benchmark tool is

implemented in the HBase cloud database, which is

based on NoSQL technology.

The paper is organized as follows. Firstly, the

background data for the applied cloud technologies,

namely IaaS and NoSQL, and also the benchmark

tool, are presented. This is followed by the

measurement system description. The results are

shown using various setups, but the main emphasis

is on the public cloud environment. The main

criteria are latency and transactions per second.

Next, the results are discussed, critically reviewed,

and also a short business comparison is given.

Finally, the conclusions and the future research

proposals are made.

2 BACKGROUND

2.1 IaaS

An IaaS provides the most natural approach for the

research. The existing telecom network elements,

using the Linux operating system, can be easily

ported as such into the IaaS platforms. Compared to

the Platform as a Service (PaaS) or Software as a

Service (SaaS) alternatives, IaaS offers the best

flexibility for its users. Unlike IaaS, PaaS service

providers, such as Google App Engine or Microsoft

Azure, require that the software is tailored for the

associated platform. On the other hand, SaaS

provides a complete service that does not allow

running your own code. In addition, IaaS supports a

large selection of open source software solutions that

are compatible with the commercial IaaS market

leader, Amazon Elastic Compute Cloud (EC2)

(Amazon, 2011). By selecting the IaaS approach,

the users can avoid the vendor or system lock-in, a

feature that is much appreciated by the operators.

From commercial, public IaaS cloud vendors

EC2, being a market leader, was a natural choice.

On the private cloud side, the selection process was

a lot more difficult. There are several alternatives in

open source software IaaS platforms. The most well

known, EC2 compatible, projects are called

Eucalyptus (Eucalyptus, 2011), OpenNebula

(OpenNebula, 2011) and OpenStack (OpenStack,

2011). As one of the more mature projects,

Eucalyptus was selected for the private cloud

platform, but for future research, OpenNebula and

OpenStack are worth closer consideration.

Interoperability and backward compatibility of

the software are essential features. Amazon EC2 and

Eucalyptus provide an attractive duopoly, where

software can be ported with minor efforts from one

entity to another. The good interoperability basically

enables two different scenarios. First of all,

companies may develop their product using their

own cloud, and at once a stable phase has been

achieved, the software can be commercialized using

a public cloud. The second possibility is to utilize a

hybrid model, where private and public clouds

complement each other, enabling load balancing

functionalities. This is a valid scenario also in

telecom applications, where the traffic peaks are a

common challenge.

2.1.1 Amazon EC2

Amazon EC2 also provides several features that are

important for the research carried out. First of all,

EC2 offers a large variation of Linux distributions

and instance types from small instances up to high

performance computing (HPC) clusters. Secondly,

the basic SLA guarantee, 99.95 percent, can be

increased by using several, parallel, availability

zones. The use of parallel zones is complemented by

elastic IP addresses and monitoring services that

support the implementation of HA targets. Thirdly,

the hybrid clouds are backed by Virtual Private

Clouds (VPC), securing the connections from local

CLOSER 2011 - International Conference on Cloud Computing and Services Science

334

clouds to remote public clouds. Finally, EC2 pricing

structure is very flexible and enables various

business models. For research purposes, EC2 is an

affordable choice, because the pricing is based on

the active computing hours.

2.1.2 Eucalyptus

Eucalyptus uses a set of services running on various

terminals on different or same networks to manage

and coordinate the whole system. A Cloud

Controller (CLC) acts as a command terminal,

which defines the cloud identity and resources

available to it. It is the main service, which needs to

be running prior to all other services of the cloud to

function. A Cluster Controller (CC) provides

management services for a set of clusters, controlled

by a Node Controller (NC). It manages a set of end

nodes where Eucalyptus Machine Images (EMIs)

can run and support any user application. A Walrus

storage service is a near clone of the EC2 Simple

Storage System (S3). It provides a similar interface

for storage and can use the same tools as are

available for EC2 to manage the storage. Figure 1

illustrates the Eucalyptus architecture. (Nurmi,

Wolski and Grzegorczyk, 2009)

Figure 1: Eucalyptus architecture.

The architecture for setting up Eucalyptus can

vary according to user needs. In a basic testing

environment all four services can even reside on one

machine and work together as a proper cloud

deployment. However, there can be resource

limitations depending on the hardware available. A

better idea is to allocate one machine for CLC, CC

and Walrus, and deploy NCs and EMIs on different

computers.

2.1.3 Hybrid Cloud

Usually cloud services are implemented using public

or private IaaS concepts. However, companies can

also choose a third deployment strategy, called a

hybrid model. The basic idea is to utilize in parallel

both public and private models. The selection can be

done dynamically to match the current needs and to

minimize the costs. The hybrid model sets a difficult

selection challenge in the table. The technology

should dynamically be able to provide load

balancing between private and public clouds.

The hybrid model is not feasible for all

applications and functions, but it looks attractive for

services where the traffic load varies and the

variations can be predicted well in advance. For

example, a ticket sales service fulfils these criteria.

Telecom networks also suffer from high traffic

variations. Voice and text messaging services can

become congested during exceptional events and on

special dates. Figure 2 shows the text messaging

volume trace on New Year’s Eve (Zerfos, Meng,

Wong and Samanta, 2006, p. 267). The data shows

that during the midnight the peak load volume is ten-

fold compared to the average load. In a situation like

this, hybrid clouds can offer an economic alternative

for improving the end user experience during traffic

peaks. Although the latency times can increase, a

delayed service is a better option than no service at

all.

Figure 2: Text messaging traffic pattern at New Year

(Zerfos, Meng, Wong and Samanta, 2006, p. 267).

2.2 NoSQL

Distributed databases have been at the forefront of

cloud computing since the beginning, although the

term NoSQL was invented much later. It is an

umbrella term for a family of databases that

typically do not implement the SQL interface, but

are designed scale horizontally to support massive

data. Originally the need to create a new kind of

database stemmed from the data storage

requirements of the first globally scale internet

services. Soon, in addition to internal use at social

media sites and internet companies, NoSQL

solutions became available as services for all deve-

Cloud

Controller

(CLC)

Cluster

Controller

(CC)

Cluster

Controller

(CC)

Node Controller (NC)

Node

Controller

(NC)

Node

Controller

(NC)

Node

Controller

(NC)

Walrus

EMI

PERFORMANCE EVALUTION OF NOSQL CLOUD DATABASE IN A TELECOM ENVIRONMENT

335

lopers.

The main differences between a NoSQL and a

SQL, i.e. a Relational Database Management System

(RDBMS), in a data model are provided interfaces,

transaction guarantees and scalability. NoSQL

differs fundamentally from the SQL databases that

form the basis of telecom database systems.

Generally RDBMS is optimal for online transaction

processing (OLTP), and NoSQL for online analytics

processing (OLAP) (Abadi, 2009). While a SQL

database confirms ACID (atomicity, consistency,

isolation, durability) requirements, NoSQL

databases typically support BASE (Basically

Available, Soft state, Eventually consistent)

principles (Pritchett, 2008).

The modern history of the NoSQL movement as

an effort to store web scale data can be seen to have

begun in 2003 when Google published details on its

Google File System (GFS) (Ghemawat, Gobioff and

Leung, 2003). Later in 2006, the company published

an article describing Bigtable (Chang, et al., 2006), a

distributed storage system built on top of GFS.

Imitating Google's efforts, the Apache Software

Foundation (ASF) has developed open source

clones, called the HBase (Apache, 2011) and

Hadoop Distributed File System (HDFS) (Shvachko,

Kuang, Radia and Chansler, 2010).

HBase and HDFS were chosen for a closer

examination due to three reasons. First of all, HBase

supports consistent transactions when updating a

single row at a time. Secondly, it has a modular

design and proven basis, thanks to underlying HDFS

and ZooKeeper layers. Thirdly, HBase has active

community and support from strong internet

companies such as Yahoo. Yahoo has also

developed a benchmark tool for cloud storages,

including HBase (Cooper, et al., 2010).

2.3 TATP

The Telecommunication Application Transaction

Processing (TATP) benchmark aims to measure the

performance of a database under load which is

typical in telecommunication applications. In

particular, it is modelled after the type of queries

that are processed in HLR on a GSM network. The

benchmark tool is described in detail in the literature

(Strandell, 2003; TATP, 2011). TATP encompasses

seven different transactions of which three are reads

and four are writes. The description gives

probabilities at which each of the transactions is

executed in the client. Broadly, 80 percent are reads

and 20 percent are writes.

The database industry has been dominated by

RDBMSs for several decades, and it still is.

Accordingly, TATP benchmark is heavily dependent

on SQL, and does not provide functionality to test

other kinds of database systems. However, we have

taken action and implemented a comparable

benchmark for HBase. The schema in TATP

consists of four inter-relational tables. When

modelling the schema for HBase, the tables were

denormalised and finished off with just one table.

Denormalisation is a popular approach when

designing data models for NoSQL databases.

3 MEASUREMENT SETUP

3.1 Environment

The test environment simulates a real mobile

network, where one HLR was loaded by one or

several Mobile Switching Centers (MSC). The

TATP benchmarking tool emulates the real

signalling traffic between the MSCs and HLR. See

Figure 3 for the model. The focus in the

measurements was on the SLA, latency and

transactions per second. HA measurements were

beyond the scope of the research. All measurements

were repeated a few times and the diagrams shown

are based on average results.

It is noteworthy that the telecom level HA

requirement, 99.999 percent, can be achieved by

using independent IaaS clusters. For example,

utilization of two different Amazon EC2 zones, both

with an HA value 99.95 percent, yields together an

HA value 99.9999 percent. Similar results can be

achieved by using hybrid models.

Figure 3: HLR test environment.

3.2 TATP Transactions

The TATP benchmark database representing HLR

includes four different tables called Subscriber,

Access Info, Special Facility and Call Forwarding.

HLR

MSC

Client N

MSC

Client 2

MSC

Client 1

…

CLOSER 2011 - International Conference on Cloud Computing and Services Science

336

Subscriber table includes the basic customer data,

while Access Info table describes the access type.

Special Facility table defines the services

subscribers are entitled to, and finally Call

Forwarding table reveals forwarding rules. MSC

Clients use seven different transactions that either

read or write data. The transaction distribution is

known from real networks. Table 1 summarizes the

transactions, their types, distribution and effected

tables (Strandell, 2003).

Table 1: TATP transactions (Strandell, 2003).

Transaction name Type % Tables

Get-Subscriber-

Data

Read 35 Subscriber

Get-New-

Destination

Special Facility

Call Forwarding

Get-Access-Data Read 35 Access Info

Update-

Subscriber-Data

Write 2

Subscriber

Special Facility

Update-Location Write 14 Subscriber

Insert-Call-

Forwarding

Write 2 Call Forwarding

Delete-Call-

Forwarding

Write 2 Call Forwarding

3.3 Initial Setup

The rewritten version of TATP was tested in several

small HBase clusters. The focus was in transactions

per second capability. Running the benchmark is

interesting, especially because the results can be

compared to existing reports on SQL database

performance. One such article (Gupta, 2006) reports

a throughput of approximately 5500 transactions per

second. The performance level was achieved for 200

000 subscribers using carrier-grade hardware from

the year 2006 and an in-memory database.

However, comparing measurement results

obtained from different benchmarks testing different

databases running on top of different infrastructures

head-to-head, is not particularly meaningful.

Therefore we use the results in the white papers only

to set up a base line so that we know, whether the

first results of running HBase in a HLR setting are

on the same scale with recent commercial HLR

databases.

In the initial measurements the environment was

the following. The HBase version 0.20.6 and

Hadoop 0.20.2 were run on a multitude of test

setups, which all were considerably smaller than

what HBase is designed for. Amazon Small EC2 had

1.7 GB memory on a Ubuntu Lucid 10.04 32 bit

server. Eucalyptus had also 1.7 GB memory on top

of a Ubuntu Lucid 64 bit desktop. Local

communication was based on a 100 Mbit/s LAN,

and the PCs were equipped with Intel Core 2 Duo

processors and 8 GB memory.

All setups consisted of a four virtual machine

(VM) instance cluster. One instance was a dedicated

master running the Hadoop Master Server and

HDFS NameNode, two instances were running the

HBase Region Server and HDFS DataNode

processes, and the fourth machine was running a

single HLR benchmark process and collecting the

results. Instance deployment on Eucalyptus is

presented in Figure 4. EC2 setup is similar.

Figure 4: Test environment with Eucalyptus.

In the hybrid setup the HBase Region Servers

and HDFS DataNodes were split into both EC2 and

Eucalyptus, while the HBase Master, HDFS

NameNode and benchmark client were running on a

local Dell Optiplex 960 desktop. A noteworthy

result itself is that we were able to run HBase and

HDFS with default settings on Amazon Small EC2

instances without problems. Figure 5 presents the

hybrid cloud setup.

Figure 5: Hybrid cloud setup.

3.4 Final Setup

The final measurement setup was decided to be

based on Amazon EC2 only. It was already

beforehand clear that a hybrid IaaS architecture is

Eucalyptus

Controller

Node

Controller1

Small VM

HBase Master

HDFS NameNode

Small VM

HBase Region Srv

HDFS DataNode

Node

Controller2

Small VM

HBase Region Srv

HDFS DataNode

Small VM

Benchmark Client

Small VM

HBase Region Srv

HDFS DataNode

Dell Optiplex 960

HBase Master

HDFS NameNode

Benchmark Client

Small EC2 VM

HBase Region Srv

HDFS DataNode

Amazon

Eucalyptus

PERFORMANCE EVALUTION OF NOSQL CLOUD DATABASE IN A TELECOM ENVIRONMENT

337

not optimal for a centralized database system.

Furthermore, during the research process it was

found that Eucalyptus is not the best platform for a

hybrid cloud. The main reason is the lack of

necessary management tools for remote instances. In

contrast, OpenNebula and OpenStack support these

functionalities, and for that reason those IaaS

alternatives should be researched more in the hybrid

context. On the other hand, private cloud

measurement results were mainly limited only by the

local hardware applied, having a small difference to

the usual HLR environment.

In the later experiments, we investigated the

effect of load, replication, database size and node

failure on performance by running HBase on a

cluster of six Large EC2 instances as shown in

Figure 6. As the characteristics of EC2 instance

types are sometimes modified, it is purposeful to

specify here that the large instances used in the

experimentation were virtual machines with 7.5 GB

of memory and two virtual cores with two EC2

compute units, each running a 64-bit Ubuntu Server

version 10.04. In addition to one master and four

slave nodes, one large instance was hosting the

benchmark clients.

Figure 6: Final setup with Amazon Large EC2s.

4 MEASUREMENT RESULTS

4.1 Initial Measurements

The initial test results are shown in Figure 7. All

transactions per second measurements were made

with a database size of 200 000 subscribers. As

described above, the comparison of different IaaS

platforms with each other is not useful, but on the

other hand, the results can be used for getting the big

picture of the system. Compared to the four year old

carrier grade numbers, the results were encouraging.

The Amazon Large EC2 cluster achieved roughly 15

percent performance of the carrier grade system.

Even the local, Eucalyptus based IaaS cluster

managed to produce reasonable results. Plain

workstation and legacy cluster measurements were

made to gather experiences of running the bench-

mark setup in different environments.

The performance of a hybrid setup, consisting of

Small Eucalyptus and EC2 instances, was better than

with a Small EC2. The results prove that a hybrid

IaaS can be made, and that the throughput is roughly

the average of the building blocks. The first

experiments also revealed that the bottleneck in the

measurements was the single benchmark client. In

the main measurements this bottleneck was removed

by using several parallel clients. In the real networks

one HLR is connected to several MSCs, too.

Figure 7: Throughput with various systems.

4.2 Main Measurements

The experiments analyzed here and presented in

Figures 8 and 9 illustrate a load curve typical for I/O

heavy systems. The throughput improves to a certain

limit when client processes, i.e. load increases, are

added, but after the limit latency grows dramatically.

The point of maximal curvature is known as the

knee. Our benchmark collects the latency

distribution for each of the seven transaction types

of TATP separately. Therefore, the response time

values shown in the figures are the 95th percentile

values perceived by the worst performing client

process from the heaviest transaction type.

Replication is a standard way of achieving

durability of data in NoSQL databases. Figure 8

shows the results of the experiment where the goal

was to assess the effect of replication on

performance. The table was populated with 200 000

subscribers. First of all, the throughput results show

that 16 client processes are close to the knee, e.g. a

point where the results turn worse. Secondly, we

notice that the replication factor does not have a

major impact on the throughput. Also the response

time holds almost steady independent of the

replication factor. We assume that the performance

penalty of replication is virtually nonexistent,

because even if writes become heavier, reads are

Large EC2 VM

HBase Region Srv

HDFS DataNode

Large EC2 VM

Benchmark Clients

Large EC2 VM

HBase Master

HDFS NameNode

Large EC2 VM

HBase Region Srv

HDFS DataNode

Large EC2 VM

HBase Region Srv

HDFS DataNode

Large EC2 VM

HBase Region Srv

HDFS DataNode

CLOSER 2011 - International Conference on Cloud Computing and Services Science

338

scaled across the replicas, which balances the results

in a read heavy benchmark.

Figure 8: Impact of replication factor and number of

clients to throughput and latency.

Figure 9 presents the effect of the amount of

subscribers in the database on performance. The

results were gathered when the replication factor

was set to three. As expected, the performance

gradually decreases as the table size grows. Looking

at the results from 16 concurrent client processes,

increasing the table size from one to five million

subscribers decreases the throughput 32 percent and

lengthens the response time 36 percent.

Figure 9: Impact of database size and number of clients to

throughput and latency.

To verify that the HDFS replication gives

protection from node failures, we studied the effect

of killing one slave node in the middle of a

benchmark run. The measurement was done with a

database size of 1 million subscribers and replication

factor two. Figure 10 shows the effect of one failing

node 10 seconds after the launch of the run as

perceived by four client processes. The throughput

values are gathered once a second for each client and

the results are stacked in the presentation. In this

sample the distributed database quickly recovers

from the failure and continues serving clients within

two seconds. The perceived recovery time in the

experiment would be too much for real-time

telecommunications applications, but it could be

improved by tuning the parameters related to

timeout mechanisms.

Figure 10: Recovery time from a node failure.

4.3 Measurements vs. Requirements

In order to give an idea of the load generated by the

modified version of TATP, the performance testing

tool bundled with HBase was also run on the test

setup of six large EC2 instances. We run the

performance test on the master node using 16 client

threads and disabled MapReduce for it. By default,

the test populates a table with one million rows of 1

kB each. In our experiment the random Read test

took 1492 seconds, which leads to a throughput of

5.36 Mbit/s per client thread, and an aggregated

throughput of 85.8 Mbit/s.

The 3GPP has defined the HLR performance

requirements in their general specification (3GPP,

2009). According to that each subscriber produces

on average 1.8 mobility related and 0.4 call handling

related transactions per hour. Together this yields

2.2 transactions per hour per subscriber. With this

information and the total number of subscribers, a

requirement curve for transactions per second can be

defined. Pulling together the requirements for HLR

performance and the measurements from the HBase

benchmark leads us to the conclusion that up to 4

million subscribers could in theory be supported by

six large EC2 instances. Measurement results and

comparable requirements are shown in Figure 11.

PERFORMANCE EVALUTION OF NOSQL CLOUD DATABASE IN A TELECOM ENVIRONMENT

339

Figure 11: Number of subscribers and throughput vs.

requirements.

In the experiment each subscriber added 3.7 kB

to the database size leading to a total of 18.5 GB for

5 million subscribers. This is still in the area that can

be handled in main memory, and therefore existing

HLR solutions can support such deployments using

an in-memory database. Similarly to most NoSQL

databases, HBase does not support transactions,

which span multiple rows, but on the other hand

HBase guarantees that a single row remains

consistent at all times. In an HLR all transactions

read or update a single subscriber, and therefore the

database was modelled so that all data related to a

single subscriber is on one row.

5 DISCUSSION

Although the HLR is not a primary candidate for the

cloud, the results give evidence that some other

mobile network elements could be placed there.

Application servers, such as SMS Center (SMSC),

IP Multimedia Subsystem (IMS) and Service

Delivery Platform (SDP), are examples of those.

Backend processes can provide an even better

solution area (Hajjat, et al., 2010). Billing, customer

care and maintenance systems create a lot of data

that could be computed by cloud infrastructures. A

general purpose cloud can also be provided by

mobile network vendors, who might use their large

customer base to benefit from the statistical

multiplexing. The same approach can work with

operators, who operate in several countries and

continentals. It can be expected that in future mobile

networks, such as Long Term Evolution (LTE),

operators will compete and cooperate at the same

time, leading to network sharing initiatives.

The HLR benchmarking measurements produced

a lot of information about the cloud computing

opportunities in the telecom sector. The main lesson

is that even the strict telecom SLA requirements can

be achieved with both public and private clouds. The

initial measurements also revealed that to match the

existing RDBMS solutions, NoSQL databases have

to fully utilize horizontal scalability. In addition,

configuration parameters must be properly tuned,

enterprise class infrastructure must be used and

several client processes must be deployed.

IaaS, both private and public versions, operated

according to expectations in the measurements. Due

to the short history of private clouds, they are still

developing. Amazon EC2, on the other hand, is

already a mature product. The hybrid cloud was a

side track in this research. It became evident that the

hybrid cloud does not suit well to a centralized

database system. In addition, the hybrid setup must

be carefully designed to overcome configuration,

management and load balancing challenges.

For certain applications a hybrid cloud can be an

interesting option to optimize the dimensioning for

peak loads (Moreno-Vozmediano, Montero and

Llorente, 2009). However, the database solutions

should be centrally located backed by 2N or N+1

redundancy algorithms. Database distribution will

increase latency times and create unnecessary

functional complications as well. For example,

security and regulation challenges would become

high.

The financial comparison of Amazon EC2, a

private commodity server cluster and a carrier grade

server would be interesting, but is not within the

scope of this paper. The EC2 pricing structure is the

most versatile, including also spot prices (Mattess,

Vecchiola and Buyya, 2010). In a nutshell, EC2

prices are competitive with private clusters,

especially if three-year term fixed prices and lower

hour prices are utilized. The price scale of carrier

grade services is large, which makes a reliable

comparison almost impossible. It is also worth

mentioning that, unlike in clouds (Greenberg,

Hamilton, Maltz and Patel, 2009; Walker, Brisken

and Romney, 2010), the weight of computing power

and storage is marginal in the HLR price formula.

However, in the application servers computing costs

are becoming ever more dominant.

6 CONCLUSIONS

We have introduced research on how cloud

computing performance meets the SLA requirements

of mobile networks. The home location register

(HLR) was chosen as an example for benchmarking

CLOSER 2011 - International Conference on Cloud Computing and Services Science

340

measurements. The HLR benchmark tool, originally

developed for the SQL databases, was ported into

the NoSQL, HBase specific environment. The

software instances were deployed on private, public

and hybrid Infrastructure as a Service (IaaS)

platforms. The measurement results indicate that

cloud technologies can achieve the mobile network

latency and transactions per second requirements.

Also telecom high availability (HA) targets can be

met by using parallel computing zones. It is

recommended that future studies should evaluate

whether cloud technologies can be applied to mobile

application servers and backend processes. Also the

Long Term Evolution (LTE) will provide interesting

research opportunities on network sharing between

operators. Finally, the hybrid clouds deserve

attention in managing traffic peaks.

ACKNOWLEDGEMENTS

The work is supported by Tekes (the Finnish

Funding Agency for Technology and Innovation,

www.tekes.fi) as a part of the Cloud Software

Program (www.cloudsoftwareprogram.org) of Tivit

(Strategic Centre for Science, Technology and

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

REFERENCES

3GPP, 2009. TS 43.005, version 9.0.0 release 9 -

Technical performance objectives.

Abadi, D. J., 2009. Data management in the cloud:

limitations and opportunities. IEEE Data Engineering

Bulletin, 32(1), March 2009.

Amazon, 2011. Amazon Elastic Compute Cloud (EC2).

(online) Available at: <http://aws.amazon.com/ec2/>

(Accessed: 24 February 2011).

Apache, 2011. The Apache Software Foundation – Hbase.

(online), Available at: <http://hbase.apache.org/>

(Accessed: 24 February 2011).

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

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

Rabkin, A. and Stoica, I., 2009. Above the Clouds: A

Berkeley view of cloud computing. Technical Report

No. UCB/EECS-2009-28, Electrical Engineering and

Computer Sciences, University of California at

Berkeley, 10 February 2009.

Chang, F., Dean, J., Ghemawat, S., Hsieh W. C., Wallach,

D. A., Burrows, M., Chandra, T., Fikes, A. and

Gruber, R.E., 2006. Bigtable: a distributed storage

system for structured data. In Proceedings of the 7th

Symposium on Operating Systems Design and

Implementation (OSDI ’06), Seatle, WA, USA 6-8

November 2006, USENIX Association.

Cooper, B. F., Silberstein, A., Tam, E., Ramakrishnan, R.

and Sears, R., 2010. Benchmarking cloud serving

systems with YCSB. In Proceedings of 1

ACM

Symposium on Cloud Computing (SoCC '10),

Indianapolis, Indiana, USA 10-11 June 2010.

Eucalyptus, 2011. The Open Source Cloud Platform.

(online) Available at: <http://open.eucalyptus.com/>

(Accessed: 24 February 2011).

Gabrielsson, J., Hubertsson, O., Más, I. and Skog, R.,

2010. Cloud computing in telecommunications.

Ericsson Review, vol. 1, pp. 29-33.

Ghemawat, S., Gobioff, H. and Leung, S.-T., 2003. The

Google file system. In Proceedings of the 19th ACM

Symposium on Operating Systems Principles

(SOSP’03), New York, NY, USA 19-22 October 2003,

ACM.

Greenberg, A., Hamilton, J., Maltz, D. A. and Patel, P.,

2009. The Cost of a Cloud: Research Problems in Data

Center Networks. ACM SIGCOMM Computer

Communication Review, Volume 39, Number 1,

January 2009.

Gupta, N., 2006. Enabling High Performance HLR

Solutions. Netra Systems and Networking, Sun

Microsystems, Inc., version 2, 17 April 17 2006.

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

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

Cloudward bound: planning for benefiticial migration

of enterprise applications to the cloud. In Proceedings

of the 2010 ACM SIGCOMM, New Delhi, India 30

August - 3 September 2010.

Mattess, M., Vecchiola, C. and Buyya, R., 2010.

Managing peak loads by leasing cloud infrastructure

services from a spot market. In Proceedings of the

12th IEEE International Conference on High

Performance Computing and Communications (IEEE

HPCC 2010), Melbourne, Australia 1-3 September

2010.

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

I.M., 2009. Elastic management of cluster-based

services in the cloud. In Proceedings of the 1st

workshop on Automated control for datacenters and

clouds (ACDC '09), Barcelona, Spain 19 June 2009.

Murphy, M., 2010. Telco cloud. (Presentation) Cloud Asia

Singapore, 5 May 2010.

Nurmi, D., Wolski, R. and Grzegorczyk, C., 2009. The

Eucalyptus open-source cloud-computing system. In

Proceedings of the 9th IEEE/ACM International

Symposium on Cluster Computing and the Grid

(CCGrid 2009), Shanghai, China 18-21 May 2009.

OpenNebula, 2011. The Open Source Toolkit for Cloud

Computing. (online) Available at:

<http://www.opennebula.org/start> (Accessed: 24

February 2011).

OpenStack, 2011. (online) Available at:

<http://openstack.org/> (Accessed: 24 February 2011).

Pritchett, D., 2008. Base: an acid alternative. ACM Queue,

vol 6, no. 3, pp. 48-55.

Shvachko, K., Kuang, H., Radia, S. and Chansler, R.,

2010. The Hadoop Distributed File System. In

Proceedings of IEEE 26

Symposium on Mass Storage

PERFORMANCE EVALUTION OF NOSQL CLOUD DATABASE IN A TELECOM ENVIRONMENT

341

Systems and Technologies (MSST), Incline Village,

NV, USA 3-7 May 2010, IEEE.

Strandell, T., 2003. Open source database systems:

systems study, performance and scalability. Master’s

thesis, University of Helsinki, Faculty of Science,

Department of Computer Science and Nokia Research

Center.

TATP, 2011. Telecom Application Transaction Processing

Benchmark. (online) Available at:

<http://tatpbenchmark.sourceforge.net/> (Accessed:

24 February 2011).

Walker, E., Brisken, W. and Romney, J., 2010. To lease or

not to lease from storage clouds. Computer, IEEE,

April 2010, pp. 44-50.

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

2006. A study of the Short Message Service of a

nationwide cellular network. In Proceedings of

Internet Measurement Conference 2006 (IMC ’06),

Rio de Janeiro, Brazil 25-27 October 2006.

CLOSER 2011 - International Conference on Cloud Computing and Services Science

342