Harnessing Clouds for E-Learning: New Direcions Followed by UNED Agusín C. Caminero, Anonio Robles-Gómez, Salvador Ros, Robero Hernández, Rafael Pasor Dep. de Sisemas de Comunicación y Conrol Universidad Nacional de Educación a Disancia, UNED Madrid, Spain {accaminero, arobles, sros, robero, rpasor}@scc.uned.es, Nuria Oliva, Manuel Casro Dep. de Ingeniería Elécrica, Elecrónica y de Conrol Universidad Nacional de Educación a Disancia, UNED Madrid, Spain {noliva, mcasro}@ieec.uned.es Absrac In his paper, we presen our work on enhancing he echnological infrasrucure a Spanish Naional Universiy for Disance Educaion (UNED, Universidad Nacional de Educación a Disancia) wih cloud compuing principles. This includes he developmen of (1) virualized environmens o allow sudens o do pracical exercises easily in he field of neworks and communicaions; and (2) load forecasing echniques o improve on he usage of he echnological infrasrucures of UNED, so ha he Qualiy of Service (QoS) experienced by users is kep and power consumpion is minimized. Keywords- cloud compuing; e-learning; qualiy of service; power consumpion; load forecasing. I. INTRODUCTION Wih he adven of he cloud echnology [1], [2], more dynamic compuing plaforms are been developed. The cloud is a shif from he previous compuing archiecures in which compuers had saic sofware feaures, hus making users of such resources fi ino hose feaures. For example, if a shared compuer has a Linux operaing sysem insalled along wih some programs and libraries, users willing o run heir applicaions on i had o make sure ha heir applicaions could run on such sysem. Hence, he use of compuing sysems could be considered as compuer guided, since users had o fi heir applicaions o mee he feaures of he compuer. The cloud allows sysems o dynamically provide he compuing resources heir users need, reducing expenses, energy consumpion and improving on heir scalabiliy [3], [4], [5]. Hence, if users wan o run some applicaions in a cloud, i is he compuer which has o fi ino he needs of he users. In he example above, Virual Machines (VM) can be insaniaed dynamically o mee he users requiremens. The cloud sysem can hus be considered as a user guided sysem, since he compuing resource is adaped o he users needs. Furhermore, an appropriae cloud infrasrucure manager (such as OpenNebula [1], [6] or Eucalypus [7]) can provide on demand insaniaion, monioring, and live migraion of VMs. In his manner, faul olerance and scalabiliy are provided. Anoher imporan poin o keep in mind is he power consumpion of he compuers [8]. According o [9], daaceners now drive more in carbon emissions han boh Argenina and he Neherlands. So, cloud infrasrucures should be managed rying o reduce he power consumpion of he compuers, along wih keeping efficien processing and uilizaion of machines. In order o achieve hese, load forecasing mehods are needed [10] [8]. Furhermore, e-learning echnologies are aimed a providing learning capabiliies by means of echnological infrasrucures (e.g. sudens conneced o servers hrough he Inerne). For a large-sized disance universiy, such as he Spanish Naional Universiy for Disance Educaion [11] (UNED, Universidad Nacional de Educación a Disancia), wih more han 200,000 sudens, 1,500 lecurers, and 2,000 adminisraive saff, he cloud seems o be he righ paradigm o efficienly manage is echnological infrasrucure. This efficiency is especially needed by UNED since his is a disance universiy (no face-o-face classes are provided), hence i relies on is echnological faciliies o deal wih sudens, keep rack of hem, and provide hem wih he access o lecurers and academic resources. UNED is working on harnessing cloud echnology for is e-learning purposes, and his paper presens he direcions being followed. Therefore, wo direcions are presened in his paper. These are he developmen of (1) a virualized environmen o allow sudens o do pracical exercises easily in he field of neworks and communicaions; and (2) echniques o predic he load of he echnological infrasrucures of UNED, so ha users receive Qualiy of Service (QoS) while power consumpion of compuers is minimized. This paper is srucured as follows: Secion II provides relaed work in he fields of cloud compuing and e-learning. Secion III explains he specific necessiies of UNED. Secions IV presens he main conribuions of his paper, menioned above. Finally, Secion V concludes he paper and presens guidelines for fuure work. II. RELATED WORK The combinaion of cloud echnologies and e-learning has been scarcely explored. The mos promising works are reviewed in his secion. References [12], [10] propose an e- learning framework whose infrasrucure relies on cloud Page 412
compuing. The archiecure of such sysem has several componens aimed a he efficien provision and managemen of he e-learning services. Among oher ineresing feaures, his sysem is able o pre-schedule resources for he ho conens and applicaions before hey are acually needed, o safeguard he performance in concurren access, bu no deails have been found wih regard o how his is achieved. Reference [13] focuses on he reservaion of VMs o sudens for an specific ime frame. In [14] CloudIA, a framework which provides on-demand creaion and configuring of VMs, is presened. Finally, reference [15] inroduces a proposal for personal and virual learning. This proposal ineracs wih services ha rely on he cloud, such as YouTube or GoogleDocs. As menioned above, references [10] and [8] agree ha load forecasing mehods are needed o efficienly manage compuing resources in a cloud environmen. In [16], i is shown ha alhough load exhibis complex properies, i is sill consisenly predicable from pas behavior. In [17], an evaluaion of various linear ime series models for predicion of CPU loads in he fuure is presened. For his work, a echnique based on hisorical daa is used, since i has been demonsraed o provide beer resuls compared o linear funcions [18], [19]. III. THE SPANISH NATIONAL UNIVERSITY FOR DISTANCE EDUCATION The Spanish Naional Universiy for Disance Educaion (UNED, Universidad Nacional de Educación a Disancia) [11] is an off-campus universiy which is devoed o provide universiy raining o sudens who canno aend on-campus classes. For insance, sudens having igh imeable because of oher commimens (e.g. work or family commimens), convics and hospial paiens, whose aendance o on-campus classes is oally unfeasible, may harness his Universiy. UNED provides such sudens wih he plaform o receive educaion a home wihou he ime-consuming and someimes money-consuming effor requesed by on-campus universiies. Courses a UNED are managed a he cenral offices in Madrid, where lecurers aend sudens from anywhere by means of he echnological infrasrucure. The echnological infrasrucure of UNED [20] allows sudens o virually mee each oher and lecurers and simulae he ineracions ha ake place in radiional face-o-face classes. In figures, UNED is he larges universiy in Spain in erms of number of sudens (more han 200,000) and number of lecurers (more han 1,500). UNED employs around 6,900 local uors, who are spread over 72 associaed ceners all over Spain. Furhermore, UNED also provides universiy raining o more han 2,100 sudens in more han 11 counries. Recall ha ineracions beween sudens and universiy are by means of he echnological infrasrucures of he universiy, ha is, sudens download maerial and upload essays and oher works hrough he web sie, read and send emails, and read and pos messages in he forums. I can be seen ha he echnological infrasrucure of UNED is he key o he proper funcionaliy of he universiy, since all he ineracions beween sudens and he universiy saff are conduced hrough i. If he infrasrucure does no perform smoohingly, his would seriously affec he service received by sudens and saff, making he learning/eaching process more complicaed. Considering he size of he universiy in number of sudens and saff, he use of scalable soluions is essenial o keep he infrasrucure working efficienly, and his is where he use of cloud echnologies becomes ineresing. Nex secion presens he direcions being followed by UNED o harness cloud echnologies. IV. UNED S CLOUD INFRASTRUCTURE In a similar fashion o he effors reviewed in he previous secion, UNED is planning o deploy an infrasrucure relying on cloud echnology. This infrasrucure is inended o improve he way how sudens use he universiy resources while minimizing he power coss of such infrasrucure. This infrasrucure is aimed a wo poins. Firs, virualized environmens will be developed o allow sudens o do pracical exercises easily in he field of neworks and communicaions. Second, load forecasing echniques are being developed o improve on he usage of he echnological infrasrucures of UNED, so ha he QoS experienced by users is kep and power consumpion is minimized. These wo poins will be presened in his secion. A. Neworks and communicaions. This poin is aimed a improving he way how he sudens of nework subjecs carry ou pracical exercises. Tha is, for he nework courses in he Engineer Educaion, sudens mus insall and configure a number of neworking services such as DHCP, DNS, Acive Direcory, LDAP, or web servers, along wih he configuraion of he compuers acing as cliens of hose services. This is done under boh Windows and Linux operaing sysems. For he servers, Windows Server and Ubunu Server are used. For he cliens, Windows XP and Ubunu are used. Before, sudens had o insall virual machines (VM) in heir own compuers. The way how o se up he virual nework opology linking all he VMs is error prone. Besides, running all he VMs a he same ime in he same compuer leads o problems regarding he power of he suden s compuer (shorage of memory and CPU power o run all he VMs). The evaluaion process was also difficul since sudens had o submi heir configured VMs o he universiy lecurers so ha hey were appropriaely esed. Now, all he VM are kep a he universiy cluser of compuers, where hey are execued. So, sudens jus need an inerne connecion o carry ou he pracical exercises, hey can configure heir VMs according o he exercises and sore he modified VMs in he cluser. Figure 1 presens he archiecure under developmen a UNED. In his archiecure, all he infrasrucure is execued in he universiy machines (a cluser made of a number of compuers). This cluser is managed by OpenNebula [1], [6] Page 413
Fig 2. Variaion of he load of he echnological infrasrucure of UNED over ime. o hos he VMs of new conneced sudens; (3) his archiecure reduces he power consumpion of he cluser of machines, since machines can be shu down in he absence of connecions, and urned on when needed. Fig 1. VMs srucure. which is enrused wih he VM sarup, monioring, and migraion. One of he machines in he cluser is he fron end, which runs he OpenNebula services. The oher machines are he worker nodes, which provide he physical resources o VMs. All he machines in he cluser are inerconneced and share a file sysem, so ha all of hem have easy access o all he VM images. In he worker nodes, he virual infrasrucure for each suden is deployed. As he figure depics, each suden has a virual nework made of four VMs inerconneced by means of a virual swich. This virual nework is creaed by means of VirualBox [21]. Auhors have chosen his configuraion because VirualBox has been radiionally used o creae virual neworks in he sudens compuers, hus he upgrade o he archiecure presened in his paper does no represen a major change for sudens. Work on sudying oher ways of implemening he virual nework is among he fuure work. In he case ha all he worker nodes are busy, raher han degrading he QoS of he already conneced sudens or rejecing sudens willing o sar a new session, our archiecure will conac a public cloud (such as Amazon EC2 [4]) and will creae he sudens VMs here. This is done by means of he OpenNebula s public cloud inerface. This way, our archiecure provides scalabiliy, regardless of how many sudens are conneced a any given momen. This archiecure is being developed a UNED, and will provide he following advanages: (1) his archiecure moves compuaional complexiy away from sudens, who jus need an inerne connecion o carry ou he nework exercises; (2) his archiecure provides scalabiliy, since when UNED s machines are loaded a public cloud provider will be used B. Load forecasing. Along wih he archiecure presened in he previous secion, we are working on how o predic in advance he load ha UNED s echnological infrasrucure will have. Forecasing he load of he echnological infrasrucures is a key issue in cloud compuing in order o perform more efficien resource provisioning [8]. To his end, we are developing a load predicion echnique based on Exponenial Smoohing. Exponenial Smoohing (ES) [22] is a simple predicion mehod based on hisorical and curren daa ha works very good in pracice [23]. ES is a procedure for coninually revising a forecas in he ligh of more recen experience. In his way, ES assigns exponenially decreasing weighs as he observaions ge older. In his poin, recen observaions are given relaively more weigh in forecasing ha he older ones. There are several ypes of ES s. In his work a riple exponenial smoohing is used, which is also named Hol- Winers [22]. Wih his kind of ES he rend and seasonaliy of daa are aken ino accoun for he predicions. Trend refers o he long erm paerns of daa (e.g. sales of a oy may show a linear upward rend if hey increase by 1 million Euros a year), whils seasonaliy is defined o be he endency of ime-series daa o exhibi behavior ha repeas iself every L periods (e.g. sales of a oy may increase near Chrismas and decrease aferwards, every year). The reason for choosing an ES which akes ino accoun he rend and seasonaliy is because our daa may presen boh behaviors. For example, in one day log of CPU load, here is a seasonaliy of having less CPU load percenage a nigh hours. There is also a 4 monhs seasonaliy, since he CPU load increases as exams ge closer (e.g. sudens access he online courses more frequenly) and decreases aferwards. Moreover, here is a one year seasonaliy, since he same CPU variaions happen every year, and his is depiced in Figure 2 (keep in mind ha exams periods in Spain are in Sepember, February, and June). Furhermore, number of sudens in our Universiy increases every year, which reflecs a rend. The forecasing mehod used is presened in Equaion 1. A he end of a ime period, and being x he observed value of he ime series a ime (in our case, he CPU usage), hen f +m is he forecased value for m periods ahead, T is he rend of Page 414
he ime series, L is he deseasonalized level, and S is he seasonal componen. deadline deadline f+ m = + m= init m= init ( L + T ( m 1 ) ES = ) S L * (1) m L x α * + (1 α)*( L 1 + T 1) (2) S = L T = * ( L L 1 ) + (1 β ) * T 1 β (3) S x γ S (4) = * + (1 γ )* L The deseasonalized level L is calculaed following he Equaion 2. The rend of he ime series T is he smoohed difference beween wo successive esimaions of he deseasonalized level as Equaion 3 depics. Finally, he seasonal componen (S ) is calculaed using he Equaion 4. This expression means only a combinaion of he mos recenly observed seasonal facors given by he demand x divided by he deseasonalized series level esimae L and he previous bes seasonal facor esimae for his ime period. Thus, seasonaliy indicaes how much his period ypically deviaes from he period (in our case 4 monhs) average. So, a leas one full season of daa is required for compuaion of seasonaliy. In he equaions, α, β and γ are consans ha mus be esimaed in such a way ha he mean square error is minimized. I is imporan o se a correc value for hem o predic he behavior of resources and nework as accurae as possible. In our work, he R program [23] is used for calculaing hese parameers wih he aim of making he predicions as accurae as possible. Thanks o he presened load predicion echnique, more efficien resource provisioning can be performed, which leads o beer uilizaion of resources and lower power consumpion. V. CONCLUSIONS AND FUTURE WORK Cloud echnologies allow he efficien use of echnological infrasrucures, which is of grea ineres for e-learning. Thanks o clouds, scalable, faul-oleran and green compuing infrasrucures can be developed. This paper presens he effors being carried ou by UNED (he larges universiy in Spain) o harness cloud echnologies for e-learning purposes. This can be summarized as he developmen of (1) virualized environmens o allow sudens o do pracical exercises easily in he field of neworks and communicaions, and (2) load forecasing echniques o improve on he usage of he echnological infrasrucures of UNED, so ha he QoS experienced by users is kep and power consumpion is minimized. Thanks o his, ease of use, scalabiliy, and reduced power consumpion are achieved. l We are currenly working on implemening he echniques presened in his paper. For ease of implemenaion, he load forecasing echnique will be firs implemened and esed in a simulaed environmen using CloudSim [25] simulaor. Regarding he virualized environmen in he field of neworks and communicaions, work on adding a selfevaluaion module is anoher guideline for fuure work. ACKNOWLEDGMENT The auhors would like o acknowledge European Union Leonardo Projec 142788-2008-BG-LEONARDO-LMP, mpss mobile Performance Suppor for Vocaional Educaion and Training Projec, and Spanish Minisry of Science and Innovaion for he Projec TIN2008-06083-C03/TSI s-labs -- Inegración de Servicios Abieros para Laboraorios Remoos y Viruales Disribuidos. We also hank Communiy of Madrid for he suppor of E-Madrid Nework of Excellence S2009 TIC-1650. REFERENCES [1] B. Soomayor, R. S. Monero, I. M. Llorene, and I. Foser, Virual infrasrucure managemen in privae and hybrid clouds, Inerne Compuing, vol. 13, no. 5, pp. 14 22, Sep 2009. [2] R. Buyya, C. S. Yeo, and S. Venugopal, Marke-oriened Cloud compuing: Vision, hype, and realiy for delivering IT services as compuing uiliies, in Proc. of he Inl. Conference on High Performance Compuing and Communicaions (HPCC), Dalian, China, 2008. [3] Public Daa Ses on Amazon Web Services, Web page a hp://aws.amazon.com/publicdaases/. Dae of las access: 14h Sepember, 2010. [4] Amazon Elasic Compue Cloud, Web page a hp://aws.amazon.com/ec2/. Dae of las access: 14h Sepember, 2010. [5] Amazon Simple Sorage Service, Web page a hp://aws.amazon.com/s3/. Dae of las access: 14h Sepember, 2010. [6] OpenNebula, Web page a hp://www.opennebula.org/. Dae of las access: 14h Sepember, 2010. [7] Eucalypus, Web page a hp://hp://www.eucalypus.com/. Dae of las access: 14h Sepember, 2010. [8] R. Buyya, A. Beloglazov, and J. Abawajy, Energy-efficien managemen of daa cener resources for cloud compuing: A vision, archiecural elemens, and open challenges, in Proc. of he Inl. Conference on Parallel and Disribued Processing Techniques and Applicaions (PDPTA), Las Vegas, USA, 2010. [9] J. M. Kaplan, W. Fores, and N. Kindler, Revoluionizing daa cener energy efficiency, MacKinsey, Tech. Rep., 2008. [10] B. Dong, Q. Zheng, J. Yang, H. Li, and M. Qiao, An E-learning ecosysem based on cloud compuing infrasrucure, in Proc of Inl. Conference on Advanced Learning Technologies (ICALT), Riga, Lavia, 2009. [11] Universidad Nacional de Educación a Disancia (UNED), Web page a hp://www.uned.es/. Dae of las access: 14h Sepember, 2010. [12] B. Dong, Q. Zheng, M. Qiao, J. Shu, and J. Yang, Bluesky cloud framework: An E-learning framework embracing cloud compuing, in Proc. of he Firs Inl. Conference on Cloud Compuing (CloudCom), Beijing, China, 2009. [13] M. Vouk, S. Averi, M. Bugaev, A. Kurh, A. Peeler, H. Shaffer, E. Sills, S. Sein, and J. Thompson, Powered by VCL using virual compuing laboraory (VCL) echnology o power cloud compuing, in Proc. of he 2nd Inl. Conference on he Virual Compuing Iniiaive (ICVCI), Research Triangle Park, Norh Carolina, USA, 2008. [14] A. Sulisio, C. Reich, and F. Doelizscher, Cloud infrasrucure & applicaions - CloudIA, in Proc. of he Firs Inl. Conference on Cloud Compuing (CloudCom), Beijing, China, 2009. Page 415
[15] M. Al-Zoube, E-learning on he cloud, Inl. Arab Journal of e- Technology, vol. 1, no. 2, pp. 58 64, 2009. [16] P. A. Dinda, The saisical properies of hos load, Scienific Programming, vol. 7, no. 3-4, pp. 211 229, 1999. [17] H. Jin, X. Shi, W. Qiang, and D. Zou, An adapive mea-scheduler for daa-inensive applicaions, Inl. Journal of Grid and Uiliy Compuing, vol. 1, no. 1, pp. 32 37, 2005. [18] L. Tomás, A. Caminero, B. Caminero, and C. Carrión, Using nework informaion o perform mea-scheduling in advance in Grids, in Proc. of he 16h Inernaional Conference on Parallel Compuing (Euro-Par), Ischia, Ialy, 2010. [19] L. Tomás, A. Caminero, C. Carrión, and B. Caminero, Mea- Scheduling in Advance using Red-Black Trees in Heerogeneous Grids, in Proc. Of he High Performance Grid Compuing Workshop (HPGC), hold joinly wih he Inl. Parallel & Disribued Processing Symposium (IPDPS), Alana, USA, 2010. [20] S. Ros, R. Hernández, T. Read, R. Pasor, M. Casro, M. Rodríguez- Aracho, and A. Robles-Gómez, The UNED s ineroperable virual campus service managemen archiecure, in Proc. of he Froniers in Educaion Conference (FIE), Washingon D.C., USA, 2010 [21] VirualBox, Web page a hp://www.virualbox.org/. Dae of las access: 14h Sepember, 2010. [22] P. S. Kalekar, Time series Forecasing using Hol-Winers Exponenial Smoohing, Kanwal Rekhi School of Informaion Technology, Tech. Rep., 2004. [23] M. Dobber, R. van der Mei, and G. Koole, A predicion mehod for job runimes on shared processors: Survey, saisical analysis and new avenues, Performance Evaluaion, vol. 64, no. 7-8, pp. 755 781, 2007. [24] The R Foundaion, Web page a hp://www.r-projec.org/. Dae of las access: 14h Sepember, 2010. [25] R. N. Calheiros, R. Ranjan, A. Beloglazov, C. A. F. D. Rose, and R. Buyya, CloudSim: A oolki for modeling and simulaion of cloud compuing environmens and evaluaion of resource provisioning algorihms, Sofware: Pracice and Experience (in press, acceped on June 14, 2010). Page 416