Beyond Grid Portals and Towards SaaS

A New Access Model for Computational Grids, in the dMRI Brain Context

Tarik Zakaria Benmerar, Mouloud Kachouane, Fatima Oulebsir-Boumghar

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

Keywords:

Grid Computing, Grid Portals, Cloud Computing, SaaS, OAuth, MyProxy, Grid-Cloud Interoperability, dMRI

Brain, Web Worker, Canvas2D, WebGL, MarionetteJS, BackboneJS.

Abstract:

Acigna-G is an ongoing research project to develop a new hybrid Grid SaaS architecture. CloudMRI is our

proof-of-concept Acigna-G based SaaS Service for the Brain dMRI ﬁeld. The main objective of such architec-

ture is to provide local (Browser) and remote (Grid) intensive computational capabilities completly abstracted

to the SaaS user. The result is a combination of an in Browser Rendering and Computation engine, interop-

erable REST-SOAP Grid Services, and interoperable web-grid authentication mechanisms. Such architecture

can allow new types of SaaS Services, speciﬁcally for the dMRI Brain ﬁeld.

1 INTRODUCTION

At the higher level of the Cloud Computing archi-

tecture, SaaS Services provide ubiquitous access to

hosted software stack. These softwares provide an in-

tuitive and an interactive access to user’s Cloud Data.

From the computational standpoint, we have ob-

served that the current SaaS Services do not pro-

vide the intensive computational capabilities, while

remaining highly interactive and intuitive, with rich

2D/3D rendering.

Also, the grid computational capabilities have not

been used in conjunction with SaaS Cloud Services,

to abstract the underlying job submission mecha-

nisms.

The current grid portals use somewhat intuitive

web interfaces, but still follow the job submission

paradigm, where the user is provided with job sub-

mission forms and basic visualization tools.

Through the ongoing Acigna-G project and plat-

form, we have investigated a hybrid SaaS Cloud Grid

architecture. The SaaS Cloud Service provides a

rich interactive and intuitive web interface, with in

browser 2D/3D rendering and basic computations.

The intensive computational part of the service are

delegated to the grid infrastructure, while the user is

nearly unaware of this fact.

This architecture has been developed under the

umbrella of the dMRI Brain ﬁeld. CloudMRI is the

resulting SaaS Service built with this architecture.

The present paper explains the importance of such

architecture. Its remaining parts are structured as fol-

lows. In Section 2, we introduce the dMRI Brain

Tractography and related softwares. In Section 3, we

explain the importance of a Hybrid SaaS Grid Archi-

tecture. In Section 4, we present the Acigna-G Plat-

form and some parts of the its built upon CloudMRI

SaaS Service. In Section 5, we present a comparative

study. In Section 6, we conclude the paper.

2 dMRI BRAIN TRACTOGRAPHY

With the Diffusion-weighted imaging (DWI), we can

investigate a tissue microstructure in vivo and nonin-

vasively. By combining MRI diffusion with imaging

DWI became possible, enabling the mapping of both

diffusion constants and diffusion anisotropy inside the

brain and revealing valuable information about axonal

architectures (Le Bihan and Breton, 1985).

Later, describing the degree of anisotropy and the

structural orientation information quantitatively be-

came possible with the introduction of Diffusion Ten-

sor model (Basser et al., 1994). With only six parame-

ters, this Diffusion Tensor Imaging (DTI) was able to

provide a simple and elegant way to model this com-

plex neuroanatomical information.

Different regions of the brain are connected

through axons, that form the brain white matter.

White matter tracts represent a set of axons that share

a similar destination. With tractography or ﬁber-

tracking algorithms, major tracts can be clearly ob-

211

Zakaria Benmerar T., Kachouane M. and Oulebsir-Boumghar F..

Beyond Grid Portals and Towards SaaS - A New Access Model for Computational Grids, in the dMRI Brain Context.

DOI: 10.5220/0004965302110216

In Proceedings of the 4th International Conference on Cloud Computing and Services Science (CLOSER-2014), pages 211-216

ISBN: 978-989-758-019-2

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

served by DTI with 2-3 mm image (Tournier et al.,

2011) (See Figure 1).

Figure 1: Deterministic VS Probabilistic tractography

(Tournier et al., 2011).

Differing in quality and performance, there are

mainly two types of tractography:

• Deterministic Tractography: This tractography

supposes that from a starting point or seed, there

exists exactly one tract that goes out from it (See

a in Figure 1). This algorithm is fairly fast (a few

minutes to less than a hour), but produces limited

results.

• Probablistic Tractography: This tractography

supposes that from a starting point or seed, there

can be many tracts that go out from it (See b in

Figure 1). This algorithm is heavy in processing

(several hours), but produces better results.

From the dMRI Brain Software standpoint, Med-

Inria (MedInria, 2014) provides an intuitive and an in-

teractive desktop application. Neverthless, it is fairly

limited as it provides only deterministic tractography,

due to the limited ressources of a PC.

On the other hand, FSL (FSL, 2014) provides ac-

cess to advanced probablistic tractography. But, ac-

cess to such algorithms are through a non intuitive

command line, as it is required for a grid deployment.

3 HYBRID SaaS GRID

ARCHITECTURE

From the dMRI Brain Tractography algorithms and

related softwares introduced in the previous section,

we argue that the existance of a hybrid SaaS Grid ar-

chitecture is very important for the dMRI Brain ﬁeld.

In fact, the current SaaS Services such as Google

Docs (Google, 2014), provide a visualization and a

manipulation tool to act on your Cloud data. These

services provide the same features as their desktop

counterparts, with the additionnal ubiquitous access

to Cloud Data hosted on the Cloud Provider infras-

tructure.

Neverthless, we have seen that the computational

axis of the software is nearly absent in the current ser-

vices. Hence, the Cloud Providers provide only data

hosting and manipulation, and no computational ca-

pabilities are provided through the software abstrac-

tion level.

In the Grid context, Grid Portals can provide all

the computational capabilities required. With a web

interface, the user can submit jobs and visualize the

end result inside a web browser.

Through the recent years, Grid Portals have tried

to cope with the evolution of the web, speciﬁcally the

web 2.0 trend (Pierce et al., 2006). RSS feeds, social

networks, wikis and visualization through interactive

maps, have been incorporated.

Neverthless, advanced interaction and 2D/3D ren-

dering that we ﬁnd in SaaS services haven’t been in-

tegrated. Also, the job submission paradigm is still

the only paradigm followed, in the sense that a Grid

Portal is all about job submission and result visualiza-

tion. No advanced software abstraction exists, where

the user is totally unaware about any job submission.

Through our Acigna-G initiative (Benmerar and

Oulebsir-Boumghar, 2011) (Benmerar and Oulebsir-

Boumghar, 2012), we have tried to provide such a

hybrid architecture after issues faced with a previous

grid plaform (Oukﬁf and Oulebsir-Boumghar, 2009).

Through iterative improvements, the architecture has

evolved from a platform for Linux shell-like interac-

tive web access to computational tasks, to the actual

SaaS service completly abstracted from the underly-

ing grid infrastructure.

CloudMRI is our proof of concept, where we use

the Acigna-G platform to deliver a dMRI Brain SaaS.

It allows to provide the same interactive experience

as MedInria, with the additional access to advanced

processings such as probabilistic tractography, imple-

mented in FSL.

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

212

4 THE ARCHITECTURE OF THE

ACIGNA-G PLATFORM AND

THE CLOUDMRI SaaS

SERVICE

In this section, we present the Acigna-G architecture

and its built upon CloudMRI SaaS Service.

The architecture is composed into four parts: In

Browser Computing and Rendering, Restful Grid Ser-

vices, Grid-Cloud Interoperable Authentication and

Computing Grid Services. We introduce each part in

the following subsections:

4.1 In Browser Computing and

Rendering

The CloudMRI web frontend has been built as a Sin-

gle Page Application (or SPA). Rather than decom-

posing the application into multiple pages, it is hosted

under one url that dynamically fetches the required

ressources using AJAX, then emulates passing from

one url to another. We usually ﬁnd this technology

in modern web mail clients, we use daily. Mari-

onetteJS (Bailey, 2014) and BackboneJS (Document-

Cloud, 2014) have been used as the underlying tech-

nology for this purpose. It was required to build the

software in this way, as we wanted a high interactivity

at the frontend.

The frontend uses the Acigna-G MyThreadJS

Javascript library, to manage computations, rendering

and data sources. These different axis of the library

are explained in the following paragraphs:

Computational Management

The Compute Manager Component provides a central

entry point to local and remote job execution. The

remote jobs are delegated to the Restful Grid Service

presented in 4.2.

Locally, the Compute Manager manages threads

that execute either in the local browser thread, or

in a separated thread using Web Workers standard

(Mozilla, 2014b).

The local threads are managed in a way that avoid

blocking user browser interaction. They are built with

interruption points speciﬁed by the thread developer,

that the compute manager uses to schedule threads

execution.

Currently, our CloudMRI frontend uses particu-

larly the Computer Manager for parsing NIfti (NIMH,

2014) ﬁle format, through a threaded class called

ParserNII. This processing can be heavyweight, and

the component helped avoid any threading issues.

Rendering Management

The Render Manager Component constitutes a cen-

tral point for efﬁcient scheduled rendering. Rather

than immediatly rendering after a user interaction, it

schedules it at a speciﬁc time interval to avoid costly

reprocessing which can be the case of 2D/3D ren-

dering using Canvas2D (Mozilla, 2014a) or WebGL

(Khronos, 2014) standards respectively.

Currently, the CloudMRI fronted uses the Ren-

der Manager to render MRI Slices with a component

called MRISliceWidget, illustrated in Figure 2.

Figure 2: MRI Slice Widget.

Data Source Management

The Source Manager component in myThreadJS li-

brary provides a central entry point to local and re-

mote data fetching, caching, persisting and locking to

the rest of the application. Two types of data are man-

aged by this component: local data and remote data.

Local data are data hold on the user ﬁle system,

that stay persistent across sessions. Remote data can

be considered as cloud data. The Indexed Database

API (W3C, 2014) is used for persisting local data, and

for caching remote data.

The data sources are referenced across the ap-

plication with a point separated string (e.g: ’pa-

tient.mri’). The point separation is used to differ-

entiate data namespaces, or for building a hierarchy

of data sources processings (e.g ’patient.mri.parsed’

has been generated from the ’patient.mri’ processing).

Such hierarchy is used to avoid later reprocessing of

data.

To avoid that two threads process the same data,

locking data sources mechanisms exist. For example,

if the ﬁrst thread is processing ’patient.mri’ to gener-

ate ’patient.mri.parsed’, the later is marked as locked.

The second thread waits for the availablity of the later

data source, after the ﬁrst thread processed it and un-

locked it.

The CloudMRI frontend uses the Source Manager

for the user’s Brain dMRI Images persistance.

BeyondGridPortalsandTowardsSaaS-ANewAccessModelforComputationalGrids,inthedMRIBrainContext

213

4.2 Restful Grid Service

To submit a job from the SaaS Front-end, we have de-

signed two types of Restful services. They are inter-

mediaries between the SaaS Front-end and the Com-

puting Grid Services. Their role is to avoid the con-

struction of a heavy weight XML SOAP envelope at

the browser side, by using JSON data format.

Restful BES Service

Restful BES Service is a REST equivalent to the

OGSA-BES (OGF, 2008) standard. This service is

based on the proposed solution in (Andreozzi and

Marzolla, 2009), but uses JSON for the JSDL doc-

ument (OGF, 2005). The JSON format construction

follows particular rules, that are illustrated in Figure

3 for sample examples.

Figure 3: Conversion examples between JSDL JSON and

JSDL XML, for sample fragments.

Task Speciﬁcation Restful Service

The Task Speciﬁcation Restful Service avoids the

construction of a whole JSDL document, by just re-

quiring a task name and parameters. The Restfull Ser-

vice retrieves the corresponding JSDL document and

inserts the speciﬁed parameters. This service can be

important when the SaaS service has a predeﬁned set

of jobs that can be submitted.

4.3 Grid-Cloud Interoperable

Authentication

In the proposed architecture, we have started from the

hypothesis that the Cloud Provider is an independant

third-party, separated from the Computing Grid. The

SaaS Provider should authenticate itself to the grid,

before job submission. We have then supposed two

different scenarios: Grid User and Non-Grid User, ex-

plained in the following paragraphs.

Grid User

In such scenario, the SaaS Service provides compu-

tational capabilities of the grid infrastructure the user

has access to. The Service Provider should then have

access to the user credentials to authenticate itself.

These credentials are usually stored in a MyProxy

Credential Management Service and protected with a

Providing such information (login and password)

to the Service Provider is not secure, as it could lead

to user’s credentials abuse by the provider. A solu-

tion proposed in (Basney et al., 2011) for grid portals,

avoids enveiling the login and password. The archi-

tecture uses an OAuth service in front of the MyProxy

service, as depicted in Figure 4.

Figure 4: Science Gateways Using OAuth with MyProxy

(Basney et al., 2011).

Using OAuth tokens, the user provides access to

his MyProxy account, without enveiling his login and

password. At a later time, these tokens can be revoked

and the third-party will no longer have access to the

MyProxy account.

Without change, such solution can be used in our

architecture as a means to provide the required com-

putational capabilities to grid users.

Non-Grid User

In the case of a non-grid user, other mechanisms are

required for the Service Provider to be allowed to sub-

mit jobs to the underlying grid infrastructure.

Dynamic Accounts mechanism (Insley et al.,

2009) used in the grid portal context, can be inte-

grated in our architecture. With Dynamic Accounts,

the computing grid can allocate dynamically a pool of

user accounts to the grid portal. These accounts can

be mapped to the real users at runtime.

In the SaaS context, the Service Provider could

buy a pool of dynamic accounts from the computing

grid. It could then use them to provide the required

computational capabilities to the SaaS users.

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

214

4.4 Computing Grid Services

At the Grid level of our architecture, a set of Python-

based grid services are deployed. These services have

been developped using the ZSI (ZSI, 2013) Python

library.

The ZSI library is a python library for web ser-

vices deployment and use. We have successfully de-

ployed a basic WS-BES Grid Service that accepts a

JSDL template for the job submitted.

But different extensions such as JSDL SPMD

(Savva, 2005) are required for parallel tasks deploy-

ment. We are currently working on such extensions,

to provide a more complete capabilities at the grid ser-

vice level.

We have privilegied the use of Python, as we

have found it easier for deployment than for exam-

ple Java based services. There is for example no

need for precompilation before deployment. Further-

more, Python-based webservices exhibit comparable

performances to Java-based ones (Santos and Koblitz,

2005).

Neverthless, the deployment required a bit of

patching on the source code, and we are planning to

support the open source project through source code

contributions.

HTCondor (HTCondor, 2014) is used as the job

submission system at the grid site level. Though other

job submissions systems can be used, an adaptor plu-

gin should be developped for each one, at this mo-

ment.

5 COMPARATIVE STUDY

NEWT (Cholia et al., 2010) is a RESTFul Service

provided by NERSC for building High Performance

Computing Web Applications. It provides an easy

API for authentication, persistance management and

job submission. It uses the lightweight JSON data for-

mat for data-exchange between the client application

and service.

Neverthless, the service is tightly-coupled to the

underlying NERSC infrastracture. When submitting a

job, a speciﬁc machine queue should be speciﬁed. In

some cases, the user should even specify the required

batch system script (e.g for PBS).

Also, the application examples using this plat-

form, still follows the job submission paradigm we

have presented. They don’t provide an interactive

software environment, abstracted from the infrastruc-

ture detail.

The BES-REST (Andreozzi and Marzolla, 2009)

exhibits the required service abstraction by providing

a Restful interface equivalent to an established Grid

Job Submission Standard (WS-BES).

This work has been the basis of our Restful

BES Service, with the difference that we use the

lightweight JSON data-format rather than XML for

the JSDL job speciﬁcation. This was required for an

efﬁcient job submission from the browser.

Cantor et al. (Cantor-Rivera and Peters, 2010)

have proposed a web application architecture for re-

mote MRI rendering, using WebGL. The architecture

doesn’t cover any local or remote computional capa-

bilities.

XTK (XTK, 2014) is a WebGL toolkit, for Scien-

tiﬁc Visualization. With XTK, medical images for-

mats can be rendered in the browser, such as NIfTI

format.

Compared to our work, XTK neglects threading

issues that we have solved with the MyThreadJS li-

brary. This constitutes a problem in a heavy software

environment.

In (Rings and Grabowski, 2012), the authors in-

vestigate the integration of Cloud and Grid infras-

tructures. The architecture tries to integrate computa-

tional ressources situated on an IaaS Cloud or a Com-

putational Grid. It differs signiﬁcatly from our work,

as we have looked into SaaS (rather than IaaS) Clouds

and Grids integration.

In (Church et al., 2012), the authors investigate the

deployment of an HPC environment in a SaaS Cloud.

All the computational ressources are provided by an

underlying IaaS provider. The SaaS Service provides

a web interface to submit computational jobs in the

Cloud Infrastructure. Unlike our architecture, the ser-

vice follows the job submission paradigm and don’t

manage local computations.

6 CONCLUSION

Current SaaS Services can evolve to provide compu-

tational capabilities, beyond the data capabilities ac-

tually provided. On the other hand, Grid Portals can

evolve to provide an interactive software environment

completly abstracted from the enderlying grid infras-

tructure, that would be easier and intuitive to use.

The proposed architecture tries to ﬁll in these ob-

served gaps. Particulary, we have emphasied the need

of such architecture in the Brain dMRI Field.

Currently, much of the functionalities of the

myThreadJS library have been developed and tested.

The results have been very satisfying with respect

to performance and interface responsiveness, and a

indepth comparaison with the XTK library is under

preparation. We have also developed and tested the

BeyondGridPortalsandTowardsSaaS-ANewAccessModelforComputationalGrids,inthedMRIBrainContext

215

submission of a grid job into HTCondor through a ba-

sic Restful BES and WS-BES Grid Service.

Other parts of Acigna-G are still under develop-

ment. We have also began the development of the

CloudMRI SaaS Service with the ﬁrst basic user in-

terfaces being setup.

In subsequent papers, we will provide detailed re-

views of the architecture and its performances. This

is necessary as we are heading away from a develop-

ping platform to a testing, and then a production one.

This will be characterised by rich feedbacks from the

concerned users, particulary our research colleagues.

REFERENCES

Andreozzi, S. and Marzolla, M. (2009). A restful approach

to the ogsa basic execution service speciﬁcation. In

Proceedings of the fourth International Conference on

Internet and Web Applications and Services ICIW ’09.

Bailey, D. (2014). Marionettejs website. Retrieved January

13, 2014. http://www.marionettejs.com.

Basney, J., Dooley, R., Gaynor, J., Marru, S., and Pierce,

M. (2011). Distributed web security for science gate-

ways. In Proceedings of the 2011 ACM workshop on

Gateway computing environments GCE ’11.

Basser, P., Mattiello, J., and Le Bihan, D. (1994). Mri dif-

fusion tensor spectroscopy and imaging. Biophysical

Journal.

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

ward a cloud architecture for medical grid applica-

tions : The acigna-g project. In Proceedings of the

10th International Symposium on Programming and

Languages ISPS ’2011.

Benmerar, T. Z. and Oulebsir-Boumghar, F. (2012). From

grids to cloud : The pathway for brain dmri cloud ser-

vices. In Proceedings of the 2nd International Con-

ference on Cloud Computing and Services Science

CLOSER ’2012.

Cantor-Rivera, D. and Peters, T. (2010). Perva-

sive medical imaging applications : current chal-

lenges and possible alternatives. Retrieved Jan-

uary 13, 2014. http://imaging.robarts.ca/ dcantor/wp-

content/uploads/2010/07/eHealth.pdf.

Cholia, S., Skinner, D., and Boverhof, J. (2010). Newt: A

restful service for building high performance comput-

ing web applications. In Proceedings of the Gateway

Computing Environments Workshop (GCE).

Church, P., Wong, A., Brock, M., and Goscinski, A. (2012).

Toward exposing and accessing hpc applications in

a saas cloud. In Proceedings of the 2012 IEEE

Nineteenth International Conference on Web Services

ICWS ’2012.

DocumentCloud (2014). Backbonejs website. Retrieved

January 13, 2014. http://www.backbonejs.org.

FSL (2014). Fsl website. Retrieved January 12, 2014.

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

Google (2014). Google docs. Retrieved January 10, 2014.

http://docs.google.com.

HTCondor (2014). Htcondor website. Retrieved January

13, 2014. http://research.cs.wisc.edu/htcondor/.

Insley, J. A., Leggett, T., and Papka, M. E. (2009). Us-

ing dynamic accounts to enable access to advanced re-

sources through science gateways. In Proceedings of

the 5th Grid Computing Environments Workshop GCE

’09.

Khronos (2014). Webgl website. Retrieved January 13,

2014. http://www.khronos.org/webgl/.

Le Bihan, D. and Breton, E. (1985). Imagerie de dif-

fusion in-vivo par r

esonance magn

etique nucl

eaire.

Comptes-Rendus de l’Acad

emie des Sciences.

MedInria (2014). Mediniria website. Retrieved January 12,

2014. http://med.iniria.fr.

Mozilla (2014a). Canvas. Retrieved January 13, 2014.

http://developer.mozilla.org/fr/docs/Web/HTML/.

Mozilla (2014b). Using web workers. Retrieved

January 22, 2014. http://developers.mozilla.org/en-

US/docs/DOM/Using web workers.

NIMH (2014). Nifti format. Retrieved January 13, 2014.

http://nifti.nimh.nih.gov/nifti-1/.

OGF (2005). Job submission description language

(jsdl) speciﬁcation. Retrieved January 10, 2014.

http://www.ogf.org/documents/GFD.56.pdf.

OGF (2008). Ogsa basic execution service (ogsa-bes) ver-

sion 1.0 speciﬁcation. Retrieved January 10, 2014.

http://www.ogf.org/documents/GFD.108.pdf.

Oukﬁf, K. and Oulebsir-Boumghar, F. (2009). Pr

traitement parall

ele de qsplat sous grille de calcul. In

Journ

ees Scientiﬁques sur l’Informatique et ses Appli-

cations.

Pierce, M. E., Fox, G., Yuan, H., and Deng, Y. (2006). Cy-

berinfrastructure and Web 2.0., volume High Perfor-

mance Computing and Grids in Action. IOS Press.

Rings, T. and Grabowski, J. (2012). Pragmatic integration

of cloud and grid computing infrastructures. In Pro-

ceedings of the 2012 IEEE ﬁfth International Confer-

ence on Cloud Computing CLOUD ’2012.

Santos, N. and Koblitz, B. (2005). Metadata services on

the grid. Nuclear Instruments and Methods in Physics

Research.

Savva, A. (2005). Jsdl spmd application exten-

sion, version 1.0 (draft 008). Retrieved January

10, 2014. http://www.ogf.org/Public

Comment Docs/

Documents/May-2007/draft-ogf-jsdl-spmd-008.pdf.

Tournier, J.-D., Mori, S., and Leemans, A. (2011). Diffu-

sion tensor imaging and beyond. Magnetic Resonance

in Medicine, 65:1532–1556.

W3C (2014). Indexed database api. Retrieved January 13,

2014. http://www.w3.org/TR/IndexedDB/.

XTK (2014). Xtk website. Retrieved January 13, 2014.

http://www.goxtk.com.

ZSI (2013). Zsi website. Retrieved January 10, 2014.

http://pywebsvcs.sourceforge.net/zsi.html.

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

216