FROM GRIDS TO CLOUD

The Pathway for Brain dMRI Cloud Services

Tarik Zakaria Benmerar and Fatima Oulebsir-Boumghar

ParIMed Team, LRPE, USTHB, BP32, El Alia, Bab-Ezzouar, Algiers, Algeria

Keywords:

Cloud Computing, Grid Computing, dMRI Brain, Cloud Services, SaaS, PaaS.

Abstract:

In this paper, we present the actual architecture of Acigna-G, our Cloud-oriented Grid Computing platform

and the ongoing deployment of a MAS algorithm for brain segmentation. Also, we discuss three important

improvements for this platform to allow the deployment of brain dMRI cloud services : HTTP/Restful oriented

computing services for the management of user’s service requests, application-level virtualization coupled

with distributed computing models, and separation of user request management and computing tasks execution

as found on actual PaaS Cloud Services. Such architecture would offer a convenient deployment and use of

brain dMRI PaaS/SaaS Cloud Services onto a computing grid.

1 INTRODUCTION

For more than a decade, Grid Computing has fueled

the scientiﬁc research by offering an access to a huge

number of federated computing ressources of differ-

ent organizations. Many experiments have been un-

dertaken, from oil industry to ﬁnance while passing

by Particles Physics and Medical Imagery, more par-

ticularly, in neuro-imagery and dMRI images process-

ing and analysis.

Neverthless, in such infrastructure the complexity

of deployment and development of grid applications

remain a present hurdle, and as pointed by Gabrielle

Allen et al. there exists a shortage of real grid users

(Allen et al., 2003). Also, as mentioned by Chris-

tian Vecchiola et al. (Vecchiola et al., 2009), some

issues are bureaucratic : as grids are shared world-

wide, a proposal should be submit by research groups

describing the type of research carried out. This

approach has lead to a competitive use of scientiﬁc

grids, and minor research projects couldn’t get access

to them.

Cloud Computing has permitted to avoid such is-

sues, with virtualized computing ressources offered

as an IaaS Cloud Service. Small research groups can

harness these resources, and pay only for usage. In

such approach, these groups manage also the underly-

ing computing infrastructure, and not only the design

of computing applications.

To avoid infrastructure management, PaaS Cloud

Services can be used for application deployment

only. They offer application-level services that

harness distributed computing common models like

Map/Reduce, without any prior expert knowledge.

Unfortunately, actually no PaaS cloud service offers

the same features for high performance computing,

more particularly for brain dMRI processing and anal-

ysis.

Acigna-G project is an ongoing effort to offer

a convenient deployment and use of brain dMRI

PaaS/SaaS Cloud Services onto a computing grid. In

this paper, we present the actual architecture of this

Cloud-oriented Grid Computing platform and the on-

going deployment of a MAS algorithm for brain seg-

mentation. Also, we discuss three important improve-

ments for this platform to allow the deployment of

brain dMRI Cloud Services : HTTP/Restful oriented

computing services for the management of user’s ser-

vice requests, application-level virtualization coupled

with distributed computing models, and separation of

user request management and computing tasks execu-

tion as found on actual PaaS Cloud Services.

The remaining parts of the paper are organised as

follows : In section 2 we present brain dMRI soft-

wares and features needed for a cloud service. In Sec-

tion 3, we discuss computing grids architectures, and

in Section 4, PaaS Cloud Services architecture. In

Section 5, we present the actual Acigna-G architec-

ture and the ongoing deployment of a MAS algorithm

for brain segmentation. In Section 6, we discuss the

improvements to achieve for the platform. Section 7

concludes the paper.

141

Benmerar T. and Oulebsir-Boumghar F..

FROM GRIDS TO CLOUD - The Pathway for Brain dMRI Cloud Services.

DOI: 10.5220/0003960401410146

In Proceedings of the 2nd International Conference on Cloud Computing and Services Science (CLOSER-2012), pages 141-146

ISBN: 978-989-8565-05-1

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

2 BRAIN dMRI CLOUD

SOFTWARES AND SERVICES

2.1 Introduction

In this section, we introduce the actual Brain dMRI

Softwares. After that, we discuss the important fea-

tures that are needed for a Cloud Service for Brain

dMRI.

2.2 Actual Brain dMRI Softwares

Medical Imagery softwares such as MedInria (Med-

Inria, 2012) and FSL (FSL, 2012) propose different

kinds of dMRI processing and visualization like dif-

fusion tensor ﬁeld estimation and visualization, and

white ﬁbers tractography. We should note that tools

such as FSL provide a low level command lines to

launch certain processings. This allows the deploy-

ment of grid tasks using FSL, and commands such as

fsl sub provide a parallelization for these processings.

Providing a Cloud Service for such processing is

important for broad access, and for providing a ﬂexi-

ble software upgrades.

2.3 A Brain dMRI Cloud Service

A Brain dMRI Cloud Service should deliver the same

features provided by the mentionned softwares, but

accessible using a web browser to provide a conve-

nient and broad access. In fact, three types of users

will interact with such cloud : Researchers will pro-

vide the state of the art optimised and parallelized al-

gorithms, developers will provide the back-end PaaS

applications and front-end SaaS applications, and ﬁ-

nally the clinicians who will authenticate and use the

provided SaaS for medical diagnosis.

We should note that the client web browser inter-

acts with the cloud through a Restful/HTTP interface.

Also, all the back-end infrastructure is composed of

grid ressources that span different organisations.

3 TASK-BASED AND

SERVICE-BASED GRID

ARCHITECTURE

3.1 Introduction

In the grid computing world, two strategies exist for

the infrastructure deployment : Task based strategy

and service based strategy (Glatard et al., 2008). In

this section, we present both strategies and the rele-

vant pros and cons for latter comparisons.

3.2 Task-based Architecture

In the task based strategy or global computing, com-

puting tasks to be executed are deﬁned by the user

by specifying the executable code ﬁle, the input data

ﬁles, and the command line parameters to invoke the

execution. This strategy is known for using batch sys-

tems for task execution. The best known Grid Middle-

ware adopting this strategy is Globus (Foster, 2006).

Although these systems allow the deployment of a

rich set of grid applications, they are known for being

highly centralized, and alternative P2P Grid platforms

such as JaceP2P (VUILLEMIN, 2008) have been pro-

posed for problems having interdependant comput-

ing units. Also, prior knowledge of the Grid API is

mandatory.

3.3 Service-based Architecture

In the service based strategy or meta computing, the

application codes are wrapped into standard inter-

faces, and only the invocation interface is known.

DIET (Amar et al., 2008) adopts this strategy, and en-

ables the deploment of NES (Network Enabled Server

System) (Matsuoka et al., 2000), using an RPC-style

(Remote Procedure Call). It offers access to a set

of servers offering speciﬁc computing services, for

a speciﬁc problem set, using a web browser, a com-

piled program or a PSE (Problem Solving Environ-

ment) like Matlab or Scilab.

This strategy is a simpler alternative, as input pa-

rameters are the only mendatory elements needed for

service invocation. Neverthless, the client is con-

strained to available services, and custom grid appli-

cation deployment is impossible. Also, the user must

be familiar with Grid API for service interaction us-

ing a custom code. Furthermore, for several depen-

dant service invocations, data are sent back and forth,

resulting to unnecessary communications. Recent re-

searches have proposed solutions such as data persis-

tence and redistribution to solve this issue (Desprez

and Jeannot, 2004).

4 PaaS CLOUD SERVICES

ARCHITECTURE

4.1 Introduction

Microsoft Windows Azure (Chappell, 2012) and

Google App Engine (Google, 2012) are among the

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

142

earliest and the major PaaS Cloud Providers. We re-

view in this section some archicture points that can be

used for the construction of a PaaS Cloud Service for

a brain dMRI.

4.2 Application-level Virtualization

Actually, PaaS applications are run in a virtualized

sandboxed environment, for a secure execution. The

sandboxing is ensured by the VM executing the ap-

plication. It restricts the application execution to a

limited set of standard libraries and PaaS application-

level services (datastore, mail, memcache etc.).

Neverthless, as this approach restricts access to

the underying OS, binary programs written in lan-

guages such as C/C++ or Fortran are not supported.

4.3 Common Distributed Computing

Model

Today, Distributed Computing Models such as Map

Reduce (Dean and Ghemawat, 2004) or Memcache

(Dormando, 2012) allow us building scalable web ap-

plications. Web giants such as Facebook and Google

actually use them for their web portals and applica-

tions, and have opensourced some of them for broad

use.

From a PaaS standpoint, these models are pro-

vided as application-level services. No detailed im-

plementation knowledge is required, and access is

done in a convenient way.

4.4 Separation of User Request

Management and Background

Tasks Management

Some user requests to a PaaS web application cannot

be handled in a timely fashion. These requests require

intensive tasks to be executed.

To solve such issue, PaaS platforms provide a way

to place a task in a queue, and be executed in the back-

ground. In this case, the web application recieves the

request and delegates it to a background task, so that

it can deliver a response in a short time.

5 Acigna-G ARCHITECTURE AS

A SOLUTION

5.1 Introduction

Previous works on Acigna-G (Benmerar and

Oulebsir-Boumghar, 2011) introduced an application-

level virtualization layer inspired by actual PaaS

platforms, and a particular architecture named

multi-level services architecture. The initial objectif

was to provide an experimental cloud service for the

submission and use of grid applications without prior

expertise of grid protocols.

Three levels of hierarchy exist represented by

the Master Server, the local server, and the com-

pute ressource. The master server manages all the

grid ressources and tasks, the local server manages

ressources and tasks at a given site. The compute

ressource manages the execution of tasks at its level.

A given application is decomposed into application

instances or tasks, each one executing different codes,

sandboxed using Plash (Seabon, 2012), and interact-

ing with each other through the virtualization layer.

For a convenient interaction with tasks, we provide a

virtual web terminal as illustrated in Figure 1.

Figure 1: Virtual web terminal (Benmerar and Oulebsir-

Boumghar, 2011).

In the remaining parts of this section, we introduce

the multi-level services architecture, and we present

the ongoing deployment of a MAS algorithm for brain

segmentation.

5.2 Multi-level Services Architecture

Multi-level Services Architecture is the main archi-

tecture contribution of Acigna-G. The Multi-level

Services (see Figure 2) architecture implements web

services which are at different hierarchical levels

represented by the Compute Ressources, the Local

Servers and the Master Server. Each service can call

services implemented on nodes (i.e Master Server,

Local Servers or Compute Ressources) located at the

same hierarchical level with the same parent, on the

parent node or on the child nodes.

This architecture allows a ﬂexible services calls

propagation with interresting consequences : Firstly,

some service calls originated from applications can

avoid upward propagation of calls, to decrease loads

on Local and Master Server. Secondly, the possibil-

FROMGRIDSTOCLOUD-ThePathwayforBraindMRICloudServices

143

Figure 2: Acigna-G Multi-level Services architecture.

ity of P2P inter-nodes communication allows an efﬁ-

cient deployment of applications with interdependant

computing units. Finally, this architecture is a ﬂexible

way to interpret application instance service calls into

an interaction with the grid infrastructure.

5.3 Deploying a MAS Algorithm for

Brain Segmentation

Haroun et al. have proposed a new hybrid segmen-

tation algorithm (Haroun et al., 2005). It is a MAS

algorithm combining different methods such as K-

Means, FCM and contextual methods such as Re-

gion growing and produces better results than non-

hybrid ones. A Python implementation of this algo-

rithm (Laguel, 2010) using the SPADE environment

(Spade, 2012) have been deployed onto the BrainVisa

platform (BrainVISA, 2012).

Currently, there is an ongoing deployment of this

algorithm onto Acigna-G. Before that, we are cur-

rently adapting the python application to the platform

by following two main steps (see Figure 3) :

1. Creating pseudo agents and many spade servers.

The communication between pseudo agents is

done with socket.

2. Using socketd application service for pseudo

agents communication. Socketd allows socket use

for communication between two instances even if

executed in different compute ressources. It gives

an illusion to both of them that they are on the

same ressource.

To conclude this section, we should note that ad-

ditional efforts are required to integrate any existing

platform such as Spade to a virtualized multi tenant

environment, as they have been originally designed

for non-virtualized single tenant one.

Figure 3: Adapting the MAS algorithm to Acigna-G.

6 Acigna-G IMPROVEMENTS TO

ACHIEVE

6.1 Introduction

In the present section, we discus the future improve-

ments of our Acigna-G platform, towards building

a SaaS/PaaS Brain dMRI Cloud Service. We dis-

cuss three important improvements : HTTP/Restful

oriented computing services for the management of

user’s service requests, application-level virtualiza-

tion coupled with distributed computing models, and

separation of user request management and comput-

ing tasks execution as found on actual PaaS Cloud

Services.

We also discuss the use of a convenient program-

ming language for the development of PaaS applica-

tions, and describe brieﬂy the integration of workﬂow

management in Acigna-G.

6.2 Managing User Service Requests

We have seen that PaaS platforms separate back-

ground tasks management from user requests man-

agement. By analogy and from a grid computing

standpoint, task management would refer to the grid

application execution management. HTTP/Restful

user request management actually doesn’t exist in a

grid environment.

To improve our platform, we propose adding

HTTP/Restful cloud services support, separated from

the task management. These cloud services would al-

low external non-grid users to be authenticated and in-

voke certain computing services. Hence, we have two

type of users : Grid users, that have access to comput-

ing ressources and service users, that have only ac-

cess to the different HTTP/Restful services. The user

authentication is ensured at the Restful/HTTP Master

Service level.

As for master, local and application computing

services found in our Multi-level Services Architec-

ture, there would exist a whole hierarchy of Rest-

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

144

ful/HTTP services. For any invocation requiring a

task management, it should be delegated to a com-

puting service as illustrated in Figure 4.

Figure 4: HTTP/Restful Services.

In the context of a brain dMRI cloud service, non-

grid users are doctors that can submit a dMRI image

to a Restful Service using an external web interface.

The Restful Service would submit a new task to the

grid by invoking a grid-level service.

6.3 Application-level Virtualization and

Common Parallel/Distributed

Computing Models

We have observed that PaaS common distributed

models, conveniently accessible as application-level

services permit users the built of scalable web appli-

cations without detailed implementation knowledge.

The actual multi-level services architecture allows a

virtualized interaction between different application

instances, but it still gives a low-level control of the

application to the grid user.

Keeping the same architecture but making the

virtualized environment more abstract, we envi-

sion some application services as Parallel/Distributed

computing services that can be invoked from the ap-

plication. From the grid computing standpoint, this

can be seen as equivalent to a service based strategy

where both server and client codes are on the same in-

frastructure. The latter point allows the resolution of

dependant service invocation issue we have seen with

such strategy.

If we restrict the application to use only these ser-

vices, the load on a given node is due to services

code execution rather than applications code execu-

tion, leading to a better management of the infrastruc-

ture. For example, depending on the available nodes

and their computing loads, a given service can be ex-

ecuted on a different node, or be split into a differ-

ent number tasks, to load balance the multiple nodes.

As the execution detail is left to the service, it can be

optimised for nodes containing GPUs (Graphic Pro-

cessing Units) or CPUs (Central Processing Units),

depending on the task.

In the brain dMRI cloud service context, these

computing services can be seen as common process-

ing tasks such as Diffusion Tensor Field estimation,

parallelized and optimized for a grid execution.

6.4 Workﬂow Management

As part of the future improvments, a workﬂow can

be invoked at the application service level. This

will help implementing complexe services constitued

from smaller unit tasks. Where each one, can be a

service invocation or an executable task.

In the common parallel/distributed models con-

text, some models can be implemented as workﬂows

for an even better conveniency at the application ser-

vice level.

6.5 Convenient Programming

Languages

Use of common parallel/distributed models for

application-level services in this new architecture, has

led us to propose the support of speciﬁc VM powered

programming languages, as the real computing load

is on the service execution, rather than on the appli-

cation itself. Hence, the user has a high-level control

on the application execution, and a convenient pro-

gramming would make the use of application-level

services easier. For such purpose, we have choosen

the Python programming language, for different rea-

sons (Caia et al., 2005) :

• Python is a very convenient programming. MIT

school has proposed it in an introductory course

to programming.

• Python can be optimised for scientiﬁc comput-

ing and parallel programming, thanks to modules

such as numpy.

• In fact, for maximum optimisation, speciﬁc parts

that are commonly used in scientiﬁc computing

are implemented and compiled in C and can be

called from python, thanks to the C-API interface.

7 CONCLUSIONS

Additional efforts and improvements are needed to

adapt grids for providing PaaS/SaaS brain dMRI

Cloud Services. This is reﬂected in our experimen-

tal and ongoing Acigna-G project, and as a ﬁrst step

FROMGRIDSTOCLOUD-ThePathwayforBraindMRICloudServices

145

towards such goal the introduction of the Multi-level

Services Architecture, that was proposed in a previous

work is necessary, as it provides for the applications,

a more convenient access to their grid environment,

in a virtualized manner. Actually, there is an ongo-

ing deployment of a MAS algorithm implemented in

Python/Spade, and we can argue that one of the dif-

ﬁculties of such paradigm shift is that existing plat-

forms such as Spade need additional efforts to adapt

them to a virtualized multi tenant environment.

Deep analysis of the current PaaS platforms ar-

chitectures gave us interesting insights for the future

improvements of the platform to achieve our objectif.

Firstly, a whole new hierarchy of HTTP/Restful ser-

vices will be built, that are equivalent in architecture

to the existing Acigna-G computing services, but for

non-grid users service requests management. It will

allow to such users a convenient access to different

computing services without direct grid ressources ac-

cess. And secondly, we have proposed the integration

of parallel/distributed models such as optimized and

parallelized Diffusion Tensor Field estimation for the

platform. With such an architectural point introduced,

we’ll avoid the existing platform adaptation issue, and

have a better infrastructure management as the real

computing load is on the service execution.

REFERENCES

Allen, G., Davis, K., Dolkas, K. N., Doulamis, N. D.,

Goodale, T., Kielmann, T., Merzky, A., Nabrzyski, J.,

Pukacki, J., Radke, T., Russell, M., Seidel, E., Shalf,

J., and Taylor, I. (2003). Enabling applications on the

grid a gridlab overview.

Amar, A., Bolze, R., Boix, E., Caniou, Y., Caron, E.,

Chouhan, P. K., Combes, P., Dahan, S., Daila, H.,

Delfabro, B., Frauenkron, P., Hoesch, G., Isnard, B.,

Jan, M., L’Excellent, J.-Y., Mahec, G. L., Christophe,

P., Cyrille, P., Alan, S., C

edric, T., and Antoine, V.

(2008). Diet user’s manual. inria, ens-lyon, ucbl.

Retrieved January 27, 2012. http://graal.ens-lyon.fr/

DIET/download/doc/UsersManualDiet2.4.pdf.

Benmerar, T. Z. and Oulebsir-Boumghar, F. (2011). To-

ward a cloud architecture for medical grid applications

: The acigna-g project. In Proceedings of the 10st

International Symposium on Programming and Lan-

guages ISPS ’2011.

BrainVISA, T. (2012). Brainvisa ofﬁcial website. Retrieved

January 27, 2012. http://www.brainvisa.info.

Caia, X., Langtangen, H. P., and Moea, H. (2005). On the

performance of the python programming language for

serial and parallel scientiﬁc computations. Scientiﬁc

Programming, 13:31–56.

Chappell, D. (2012). Introducing Windows Azure.

David Chappell and Associates. Retrieved Jan-

uary 27, 2012. http://www.davidchappell.com/

OnBeingIndependent– –Chappell.pdf.

Dean, J. and Ghemawat, S. (2004). Mapreduce: Simpliﬁed

data processing on large clusters. In Proceedings of

the OSDI’04: Sixth Symposium on Operating System

Design and Implementation.

Desprez, F. and Jeannot, E. (2004). Improving the gridrpc

model with data persistence and redistribution. In

Proceedings of the Third International Symposium

on Parallel and Distributed Computing/Third Interna-

tional Workshop on Algorithms, Models and Tools for

Parallel Computing on Heterogeneous Networks (IS-

PDC/HeteroPar’04).

Dormando (2012). What is google app engine ? Retrieved

January 27, 2012. http://memcached.org.

Foster, I. (2006). Globus toolkit version 4: Software for

service-oriented systems. Journal of Computer Sci-

ence and Technology, 21(4):513–520.

FSL (2012). Fsl website. Retrieved January 27, 2012. http://

www.fmrib.ox.ac.uk/fsl/.

Glatard, T., Montagnat, J., Lingrand, D., and Pennec, X.

(2008). Flexible and efﬁcient workﬂow deployment of

data-intensive applications on grids with moteur. In-

ternational Journal of High Performance Computing

Applications.

Google (2012). What is google app engine ? Retrieved

January 27, 2012. http://code.google.com/appengine/

docs/whatisgoogleappengine.html.

Haroun, R., Oulebsir-Boumghar, F., Hassas, S., and

Hamami, L. (2005). A massive multi agents system

for brain mri segmentation.

Laguel, H. (2010). D

eploiement sur une plateforme de visu-

alisation 3D, d’un algorithme coop

eratif pour la seg-

mentation d’images IRM, autour d’un syst

eme multi-

agents. Computer Sciences P. F. E., 12 Oct. 2010, di-

rected by F. Oulebsir-Boumghar., FEI, USTHB Alger.

USTHB.

Matsuoka, S., Nakada, H., Sato, M., and Sekiguchi, S.

(2000). Design issues of network enabled server sys-

tems for the grid. grid forum, advanced programming

models working group whitepaper. volume 1971,

pages 4–17.

MedInria (2012). Mediniria website. Retrieved January 27,

2012. http://med.iniria.fr.

Seabon, M. (2012). Plash’s sandbox environnment.

Retrieved January 27, 2012. http://plash.beasts.org/

environment.html.

Spade (2012). Spade2 - smart python agent environment.

Retrieved January 27, 2012. http://code.google.com/

p/spade2/.

Vecchiola, C., Pandey, S., and Buyya, R. (2009). High-

performance cloud computing: A view of scientiﬁc

applications. In ISPAN 09: Proceedings of the 2009

10th International Symposium on Pervasive Systems,

Algorithms, and Networks, pages 4–16.

VUILLEMIN, P. (2008). Calcul it

eratif asynchrone sur in-

frastructure pair

apair : la plateforme JaceP2P. PhD

thesis, Universit

e de Franche-Comt

e UFR Sciences et

Techniques Laboratoire d’Informatique de l’Universit

de Franche-Comt

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

146