toward a real time framework in cloudlet-based architecture

TSINGHUA SCIENCE AND TECHNOLOGYISSNl l1007-0214 l l07 /10 l lpp80-88Volume 21, Number 1, February 2016

Toward a Real-Time Framework in Cloudlet-Based Architecture

O. Kotevska�, A. Lbath, and S. Bouzefrane

Abstract: In this study, we present a framework based on a prediction model that facilitates user access to a

number of services in a smart living environment. Users must be able to access all available services continuously

equipped with mobile devices or smart objects without being impacted by technical constraints such as performance

or memory issues, regardless of their physical location and mobility. To achieve this goal, we propose the use of

cloudlet-based architecture that serves as distributed cloud resources with specific ranges of influence and a real-

time processing framework that tracks events and preferences of the end consumers, predicts their requirements,

and recommends services to optimize resource utilization and service response time.

Key words: smart city; cloudlet; prediction; recommendations; framework; real-time

1 Introduction

The concept of smart environments evolves from thedefinition of ubiquitous computing, which, according toMark Weiser, promotes the ideas of “a physical worldthat is richly and invisibly interwoven with sensors,actuators, displays, and computational elements,embedded seamlessly in the everyday objects of ourlives, and connected through a continuous network”[1].This is a general idea from which the smart city concepthas evolved. The vision of the future smart city isof a city with a pervasive overlay of Information andCommunication Technologies (ICT) connecting things,organizations, and people (e.g., sensors that connectcars to transportation management centers that analyzeday-to-day traffic flow data and provide solutions towhat-if scenarios in case of events or accidents[2]); thisis known as the Internet of Things (IoT).

The IoT connects various objects, such as people,data, smart objects, and processes in networks

�O. Kotevska and A. Lbath are with the CNRS, LIG,University Grenoble Alpes, Grenoble F-38000, France. E-mail: fkotevksa.olivera, [email protected].� S. Bouzefrane is with the CEDRIC Lab, Conservatoire

National des Arts et Metiers CNAM, Paris 75003, France. E-mail: [email protected].�To whom correspondence should be addressed.

Manuscript received: 2015-09-15; accepted: 2015-10-13.

of billions or even trillions of connections. Theseconnections create vast amounts of data, some of whichwe have never been accessible before[3]. Because theseobjects have embedded intelligence, the device itselfcan filter out relevant information and even applyanalytics to enable real-time decision making. Thanksto these enabling capabilities, the IoT will completelychange the type of services that are offered and alsohow they are offered[2]. Services will be availableto interact with these objects over the Internet, andquery their state and any information associated withthem. These new services provided over the Internetare collectively known as the Internet of Services(IoS)[3]. The main goal of IoS is to present everythingon the Internet as a service, including softwareapplications, platforms for developing and deliveringthese applications, and underlying infrastructures. Inthis scenario, cloud technology plays an importantrole in enabling IoS deployment, because it comprisesdifferent provisioning models for on-demand accessto applications, platforms on which developers canbuild services and applications, and elastic computinginfrastructures.

The main goal of cloud computing is to makea better use of distributed resources, combine themto achieve higher throughput, and solve large-scalecomputation problems. This means that a mobile devicecan execute a resource intensive application on a distant

www.redpel.com +917620593389


O. Kotevska et al.: Toward a Real-Time Framework in Cloudlet-Based Architecture 81

high-performance server or computational cluster andsupport thin client user interactions with the applicationover the Internet. The mobile device functions as a thinclient, with all complex computations being conductedin a nearby cloudlet, which is a trusted, resource-richcomputer or cluster of computers that is well-connectedto the Internet and available for use by nearby mobiledevices[4]. Figure 1 shows the general overview ofcloudlet-based architecture.

At present, the use, prioritization, and selection ofresources is a major concern in the design of solutionsbased on emerging technologies. The focus of this studyis to create a framework based on cloudlet architecturefor a dynamic and real-time predictive model for mobileuser potential requirements.

The rest of the paper is organized as follows.Section 2 describes related works. Section 3 presentsour proposed framework for smart cities based oncloudlet architecture, while Section 4 illustrates theproposed solution under different scenarios. Section 5 isdedicated to a detailed case study and Section 6 containsthe conclusions and plans for future work.

2 Related Work

Recently, the concept of Smart Cities has becomeincreasingly attractive for research, academic, andindustry communities. There are factors that determinethe different approaches to solve some of the challengesposed by Smart Cities, e.g., how to handle and analyzeBig Data from social and physical sensors, how to makelife in Smart Cities easier for users, or how users canselect services that will be useful for them.

Large companies in the ICT sector, such as Fujitsu,Cisco, and IBM, are strongly involved in and have beensharing their vision for the Smart City[2, 5, 6]. Followingare some related works in this domain.

Similar to participatory sensing, Mobile CrowdSensing (MCS)[7] uses the power of citizens for

Fig. 1 General architecture.

large-scale sensing. It goes beyond participatorysensing by having implicit and explicit participationsand collects crowdsourced data from both mobilesensing and mobile social network services. Theauthors characterized the key features of MCS,including crowd-powered data collection, cross-spacedata mining, and low-quality data analysis. To facilitatethe development of MCS apps, they proposed aframework and discussed the efforts of balancinghuman and machine intelligence in MCS systems.

A semantic service composition architecture,method, and algorithm has been presented in thecontext of Web of Objects (WoO)[8]. Ontologicalrelations among objects, services, and rules aredescribed to create new services dynamically. Servicesare described in the knowledge-base of the ontologylayer, and rules are generated to perform reasoningtasks for automatic service discovery and processing.The system is capable of natural language processing,which allows users to determine their requirements forinspection and repair. The authors also implemented anobject assembler and a service composer for applicationto match services.

Anteater[9] is a service-oriented architecture fordata mining that relies on Web services to achieveextensibility and inter-operability. It offers simpleabstractions for users and supports computationallyintensive processing on large amounts of data throughmassive parallelism.

All these approaches provide similar solutions, i.e.,they process the received data offline. In this work,we expand this approach by adding a frameworkfor a real-time prediction model. We consider usercontent, profiles, and trends from social sensors (socialmedia) to create timely predictions and context-awarerecommendations using machine learning and datamining techniques.

3 Proposed Approach

In order to overcome the limitations of existingrecommendations, we propose a framework for a SmartEnvironment aimed at allowing unlimited connectivityfor end consumers to their context aware servicesthrough any accessible device as well as allowing real-time prediction of user requirements.

3.1 Actors in the system

The actors in the proposed system are cloud, cloudlet,



82 Tsinghua Science and Technology, February 2016, 21(1): 80-88

end consumer, and user, defined as follows:The cloud acts as a federated cloud for a defined set

of cloudlets. It offers better performance than cloudletsand can be rapidly provisioned and integrated with thecloudlets with minimal management effort.

A cloudlet is the middle tier in our three-tierarchitecture model. It uses less memory but offersless performance than the cloud, and its function isto act as a local cloud for some territory (this meanseach cloudlet has an area of influence). This approachdecreases requests and accesses to the federated cloud.

End consumers are devices receiving output resultsfrom the processing units. The definition includes smartobjects and Cyber-Physical Systems (CPS). Smartobjects are any type of user interface with the capabilityof receiving commands via voice or touch pad andreturning the response. Examples of Smart Objectsinclude mobile devices, Google Glass, smart watches,and smart cars. Cyber-physical systems are devicesthat have the capability to receive commands such asbillboards and alarm systems.

User is a person who uses smart objects for theiractivities.

3.2 System architecture

In our proposed framework, we consider a collectionof interconnected smart objects within a smart city.Based on the scope introduced in Section 1, weuse a reference system architecture[10] that connectssmart objects through wireless technology using cloudprinciples as in Fig. 2.

To meet the requirement of real-time interactiveresponse (i.e., low latency) for the services that requireit, instead of relying on a distant cloud (as measured bythe latency), we might be able to address the limitationsof a mobile device’s resource using a nearby, resource-rich cloudlet. If no nearby cloudlet is available, the

Fig. 2 System architecture.

mobile device can gracefully degrade to a fallbackmode that involves a distant cloud or, in the worstcase, rely solely on its own resources providing limitedfunctionality (with full functionality and performancereturning when the device subsequently discovers anavailable nearby cloudlet).

This system architecture is divided into four levels ofcommunication between the actors in our system.� The first level is the communication between

user(s) and smart object(s) by employing a userinterface for interaction.� The second level comprises the communication

between all types of end consumers (smart objects,CPS, and physical sensors) with various cloudlets.This level uses LAN and Wi-Fi technologies forcommunication.� The third level of communication happens

between cloudlets and the Federated cloud andbetween cloudlets. As we see, only cloudlets cancommunicate directly with the cloud. This leveluses Internet protocols for communication.� The fourth level of communication takes place

between “social sensors” and the federated cloud,using Internet protocols. The federated cloudcontains the processing unit, data managementsystems and mechanisms, analytic engines, andservice deployment and composition.

Cloudlets, originally motivated by narrow end-to-end latency, can play a much broader foundationrole in mobile computing. A cloudlet can be viewedas a surrogate or proxy of the real cloud, locatedas the middle tier of a three-tier hierarchy: mobiledevice, cloudlet, and cloud. To meet the need forreal-time interactive response with low latency, insteadof relying on a distant cloud, data is transferredto a nearby resource-rich cloudlet[4]. Cloudlets areinterconnected in a similar way to the topologies usedin wireless sensor networks, namely mesh, star, orhub. Each cloudlet contains a service repositoryand outgoing logic, an end consumer register, VirtualObjects (VO) representation, processing unit(s), and adata management mechanism.

In turn, smart objects contain the user interface forinteraction with users and available services.

Cloudlets, smart objects, and CPS devices areconsidered to be trust worthy, and the networkconnection between them is of sufficient bandwidthand low latency. All subjects that are includedand participate use the common platform to enable




interoperability among the service distributors.Finally, an important aspect of this architecture is that

from the perspective of the end consumer everything isconsidered as a cloud service. This means that usersdo not need to be aware of or notified whether they areconnected to different cloudlets, or where the servicesthat they are using are executed (in their mobile device,in a cloudlet or in the cloud). They just need to havethe opportunity to access and use their services withoutperformance or memory constrains or limitations ontime and place, and to have a quick response time.This whole process should be automatic, and completewithout user intervention.

3.3 Prediction model for user requirements

As mentioned, a smart city is a complex networkof people and things all interdependent, increasinglyinstrumented, and interconnected. In this system, thismeans that users will always be connected to a cloudlet,have interaction with IoT, social sensors, and usedifferent types of services. Therefore, there will besignificant amounts of data streams, which leads toquestions such as:� How to extract knowledge from data collected

from physical and social sensors?� How to identify the right services to be deployed

on the cloudlets based on recommendation?� How to predict service deployment on the

cloudlets in an optimal way?These questions are the guidelines from which

this smart framework for real-time processing isderived. By using new technologies that allow thecollection and analysis of data in real-time, we canmake rapid decisions and smart predictions. Thisproposed approach for a smart framework for real-timeprocessing provides intelligent integration, aggregation,and processing of real-time data streams based onclouds.

An important source of information for our proposedframework is social sensors where people share theiropinions, experiences, events, etc., on a daily basis.Analyzing this data can give us insight into current andfuture actions. Especially important is the detection ofcurrent situations on social sensors: what is happeningin some regions, or what are the people usually doingduring a particular time of the day or year? Is theresome unexpected event that people going in a givendirection need to be notified? Based on these questions,our framework provides answers by collecting social

sensor data streams, analyzing them in real-time, anddetecting “significant trends”. In this way, when anevent or situation occurs, all users affected by this eventwill be notified rapidly. This is especially importantin the case of natural disasters (flash flood, fire, etc.)or traffic jams which require preventive action bythe users. There will be knowledge-bases with knownpatterns classified into different categories (such assport events, art performances, safety, etc.), but itwill also be possible to identify patterns generated byunknown events or situations. Knowledge discovery inorder to obtain timely information is a very importantcomponent of this proposed framework. Figure 3 showsthe complete architecture of the smart frameworkfor real-time event processing based on cloudletarchitecture.

The information generated by the Pattern Discoveryunit is one of the parameters used for predicting useractions and creating service recommendations.

The Recommendation unit utilizes user profiles,geo-location history, and contextual information tounderstand what the user is interested in. When arelevant event takes place in the city (based on theinformation received from the Pattern Discovery unit),the user will receive recommendation notificationsabout it.

The Prediction unit uses the previously mentioneduser data and knowledge discovery information toforecast what will be the next user action, for instance,which path they will take if there is a traffic jam,which concert they will attend, when will they continuewatching an online movie, etc. Prediction actions areused to predict what the next user steps will be and todiscover how the user is deciding which action to take.

The idea of this proposed framework is to beindependent from communication technologies andprotocols. As previously mentioned, these units have

Fig. 3 Framework architecture for real-time processing.




a service-oriented access, and can be approached as asingle service.

These will be deployed to the cloudlets, and beaccessible to every user.

4 General Scenarios

In this section we describe cases that would benefit fromthis proposed framework and explain how the user caninteract and benefit from it.

4.1 Scenario 1: Event detection

Information collected from social sensors cansignificantly contribute to timely inform aboutdisruptive and preventive actions in people’sroutines. Social media is a source of informationfor situations and facts related to the users and theirsocial environment. When urgent situations (like anatural disaster) arise, this information needs to becollected, analyzed, and mined as soon as possible, inorder to inform the population in a timely manner. Ina smart city, the real-time event detection processingengine is listening for data streams gathered fromdifferent geo-locations in the city, and using thisinformation the processing unit can then provideinformation about what is going on in each part of thecity in real-time. If there is some social event like aconcert or an art event, people with compatible interests(based on the user’s context, profile, and history)will be informed, so they can attend the events orfollow the associated social media streams. Similarly,if some natural disaster occurs (like fire, earthquake,etc.), users in that area or heading in that direction(as gathered from their location and context) will betimely informed. This kind of event detection acts aspreventive action for user’s safety. Figure 4 depicts theprocess step actions.

For this scenario, as already described, streams ofdata are collected in real-time from different areasand sources, then filtered, cleaned, and normalizedresulting in a standard form ready for analysis. Ifthe data shows a known pattern, it will be processedimmediately. However, if it doesn’t, it will be processedby unknown pattern discovery algorithms that attemptto extract patterns. This process is integrated witha learning system that takes into account previouslyunknown events. Output data received from this phaseis ready for semantic analysis which makes sense of thereceived data stream. The semantic analysis phase isvery important in this step because its output is used to

Fig. 4 Complex event processing.

determine if some event happened or not in a particulararea. In the case of a positive output, users affected orinterested in it will be informed. A concrete example ofthis process would be real-time streams of Twitter datafrom New York City. The streams are pre-processedand forwarded to the processing phase. If this phaseidentifies a fire at 7th Avenue and 52nd Street, theprocess of notifying users in or heading to that areawill be triggered. This information is later used by therecommendation and prediction units.

4.2 Scenario 2: Predicting user actions and contentrecommendation

In the smart city environment, where users are accessingmultiple services simultaneously, it is very important topredict their actions. The system can be pre-populatedwith information collected by other means, such ashabits, preferences, context, locations, service usagehistory, and event detection data. Building on this, ourproposal uses these sources of information to predict theuser’s actions. This functionality is part of the real-timeevent detection unit, shown in Fig. 5. The user profileinformation such as age, education, hobbies, favoritebooks, artists, movies, and habits are saved in a databasesystem with respect to privacy. This information isused to construct a baseline for determining what theuser preferences are. This baseline is complemented byother information sources, such as service registration(containing all the existing services and their meta-data), or the previously mentioned social sensors. Basedon this information (the user’s movements, time of year,day, or week), the prediction unit uses data miningalgorithms to predict the next activity or event thatthe user may want to take part in. For example, ifa given user in the New York City area frequentlyreads news about tennis and the event detection unit




Fig. 5 Data stream processing.

identifies that people are attending the US Open TennisChampionship, the system determines that the userwill be interested in attending. Recommendations areespecially important in this Smart City scenario where alarge amount of very different services exist as it may bevery difficult for end users to find appropriate services,so mechanisms are needed to help users find suitablecloud services. The Recommendation unit is also partof the real-time processing unit, shown in Fig. 5.

It uses the same source of information as thePrediction unit (user profile data and informationobtained from the knowledge unit), but also takesinto consideration the user’s context before providingrecommendations and tuning the output scenario.Context-aware recommendations are designed to reactto a user’s context without their intervention. It consistsof two parts: sensing a context scenario and adaptingthe system to a changing scenario by providing desiredservices for the user. As an example let’s say that onweekdays the user requires lunch recommendations forplaces with quick service as the time for the lunch breakis limited, but on weekends the suggestions providedshould be for family friendly restaurants.

4.3 Scenario 3: Service continuity, distribution,and accessibility on cloudlets

Adding continuous data providers as services providesvalue-added information, with various levels ofgranularity, that best suit the users’ requirementsand interests: concert events, sport events, transportproviders, local and regional planning authorities, etc.Users are offered the option to pause the servicesthen continue to use them later, resuming from theprevious state. While this characteristic is more usefulfor services that provide video or music streaming orfile downloads, it can also help users to monitor theprogress of events, such as utility restoration after a

severe thunderstorm or traffic congestion on certainroads so that alternative routes can be planned (e.g.,someone going to the airport on a Sunday).

Services are distributed among cloudlets dependingon the frequency of usage, the region specification,users’ preferences, and/or time of year. Some of theservices are more attractive for some region(s) of thecity during a particular time of the year. For example,if there is a sporting event, such as the aforementionedUS Tennis Open in New York at the end of Augustand beginning of September, during this period thereis a possibility that more users will use services relatedto the event (ranging from information about the eventto traffic reports). When the event is over, some ofthe services will be no longer distributed on all thecloudlets, but the users will still be able to find thoseservices.

Service accessibility, as shown in Fig. 6, meansallowing access to all the services when the userrequires them without constraint. Accessibility is alsoconcerned with the mechanism of finding the rightservices with ease. As there are multiple servicesavailable, users need to be provided with searchmechanisms that use a simple interface and timelyresults. By doing this, our system acknowledges anduses the inherent granularity of the architecture tooptimize the interaction with the end user, allowing,among others, approximate searches, spatial synonyms,and complex information retrieval, thereby surpassingthe limitations of traditional, query-based systems.

5 A Case Study: Cloud Connected Vehicles

As stated in Ref. [11], among the researchwork regarding smart cities, we took note ofsmart transportation that relies on sophisticatedinfrastructures to overcome situations like trafficcongestion’s, emergencies, and accidents. In this paper,

Fig. 6 Service accessibility.




we consider a vehicular infrastructure built on mobilecloud computing in order to achieve more efficienttransportation.

Connected vehicles can be viewed as entities thathave not only access to the cloud or cloudlet but arealso equipped with sensors and actuators capable ofcollecting data and actuating some vehicle components.As shown in Fig. 7, each cloudlet covers an area definedby a limited distance traveled by vehicles.

In addition to the data collected by vehicle sensorsbeing used for machine-to-machine purposes, such asmaintaining a minimal distance between the vehiclesin order to avoid accidents, some of this data (suchas speed, location, etc.) can be sent by the vehiclesto the cloudlet for it to build a view of the localtraffic in real time. These data and others gatheredby the cloud generate a global view of traffic in orderto provide recommendations to drivers regarding theweather, congested routes, etc. In addition to the dataprovided by physical sensors, helpful when makingrecommendations to the drivers, other data may besupplied by social sensors such as providing relevantinformation on events related to the covered area.

Mobile devices are also useful entities that enabledrivers not only to get data from the cloudlet orcloud but also to externalize services and data on thecloudlet. Because of the displacement of vehiclesmoving from one cloudlet coverage to another, datamust be processed in real time. Hence, the data, aswell as the tasks that process it, must meet temporalconstraints. In fact, data are collected periodically andeach must be consumed/processed by a task beforegenerating new data, otherwise the preceding collecteddata becomes obsolete.

In order to manage the services and data of thevehicular infrastructure in real-time, our scenario isbased on the following features.

5.1 Data collection by physical sensors

The data collected by the physical sensors are processed

Fig. 7 Vehicle infrastructures.

by application tasks executed within smart objects suchas user mobiles or vehicle devices. Because of theirlimited resources, smart objects are mono-processordevices that host a real-time kernel to periodicallyexecute application tasks because of periodic datasensing. These tasks are characterized according totemporal parameters as defined by Ref. [12]. In fact,each periodic task Ti in real-time systems is definedwith four parameters (ri ; Ci ; Pi ; Di ) where:� ri is the arrival time of Ti .� Ci is the execution time of Ti , i.e., the time used to

process the collected data.� Pi is called the period, and it is the time interval

that separates two successive data collections,hence two instances of Ti occur.� Di is the hard deadline of Ti : dDi D riCDi is the

absolute deadline, a time after which the executionof Ti is obsolete. Generally, Di is equal to Pi .

Each smart object that collects distinct types of sensordata can schedule its corresponding tasks using a fixed-priority scheduling algorithm called Rate Monotonic, inwhich the priority is calculated using the tasks’ periods.Hence, the higher the priority of a task, the lower is itsperiod.

According to Ref. [12], each configuration of n

periodic tasks, with Pi D Di .Vi /, is schedulable usingRate Monotonic, if the following condition is satisfied:

nX1

Ci

Pi

6 n.21n � 1/ (1)

In other words, to process data in real-time, thecorresponding tasks must meet the temporal constraintsdictated by Eq. (1). Another algorithm proved to beoptimal is the Earliest Deadline First (EDF) whichschedules real-time tasks according to their absolutedeadline. In this case, the highest priority is assignedto the task with the nearest deadline. Reference [12]proved that a configuration of n periodic tasks, withPi D Di .Vi /, is schedulable using EDF, if and only if,the utilization factor U satisfies the following condition:

U D

nX1

Ci

Pi

6 1 (2)

5.2 Data aggregation in the cloudlet

Because of the movement of vehicles, data must beprocessed in real-time to provide recommendations inreal-time, such as suggesting an efficient direction for avehicle regarding its desired destination.

To supply such recommendations to the driver, eithertasks or data can be offloaded from vehicles to thecloudlet in order to be aggregated with other data from




other vehicles, thereby building a local view of thecovered area in terms of vehicle-to-vehicle interactionsand moving-vehicle states. Data issued by the in-vehiclesensors can be enriched with social data provided by thecloud. To manage these distinct data types, the cloudletneeds to host an ordinary guest operating systems suchas Windows or Linux, as well as a real-time guestOS. The real-time VM hosts a Real-Time OperatingSystem (RTOS) to handle real-time tasks while otherVMs are dedicated to general-purpose services suchas those related to social networks. Hence, tasksthat process the sensing data must be scheduled andexecuted by the real-time VM to allow the tasks to meettheir timing constraints. The real-time VM handlesreal-time applications by executing the application tasksaccording to a real-time strategy such as EDF describedabove. However, the deadline cannot be guaranteed tobe met unless the cloudlet hosts a real-time hypervisor.The use of this type of hypervisor is motivated bythe fact it schedules VMs according to a real-timescheduling algorithm based on a fixed priority. Thehighest priority is assigned to the VM that hosts a real-time guest OS while lower priorities are assigned to theother VMs. Hence, an ordinary VM is launched by thereal-time hypervisor only if there are no real-time tasksto execute within the real-time VM.

Consequently, there are two scheduling levels:� The first level concerns the real-time hypervisor

that schedules VMs and determines the privilegesof the real-time VM (that hosts RToS) among theothers. The scheduling strategy of the hypervisorcan be based on a fixed-priority scheduler.� The second level is related to the scheduling

algorithm provided by RTOS that can use EDFto schedule tasks. According to the criticality ofthe manipulated data, we can distinguish twocategories of real time tasks:Hard real-time tasks that must process data on timebecause a missed deadline could trigger a systemdisaster such as an accident.Firm real-time tasks that can miss their deadlineinfrequently, but degrade the system’s quality ofservice.

Regarding this classification, we can consider allthe tasks that deal with in-vehicle sensing data ashard tasks while tasks related to social sensors can beviewed as firm tasks. The hard tasks can be eitherexecuted periodically by the vehicle’s smart objectsor externalized to the nearest cloudlet. Whicheverexecution environment is used, these tasks must behandled using a real-time scheduling strategy. In

contrast, the firm real-time tasks can be definedaccording to an aperiodic task model which processessocial data provided by the cloud to the cloudlet.

Each aperiodic task Ai is modeled thanks to a triple(ri ; Ci ; Di ) where ri is the time that Ai is ready toexecute, Ci is the execution time of Ai , and Di is itsdeadline.

Aperiodic tasks handled by the cloudlet can eitherbe executed within the real-time VM (RTOS) duringprocessor slack time or within an ordinary VM.To verify if RTOS can execute the periodic andthe aperiodic tasks while meeting their temporalconstraints, the utilization rate of the processor must belower than 1 as in Eq. (3).

U D

nX1

Ci

Pi

6 1; 8Ti (3)

where Ti is a periodic or aperiodic task.

5.3 Data mining and learning methods on the cloud

The cloud can gather data from distinct cloudletsregarding vehicle movement and other events likecongestion and accidents to design a global view ofthe traffic. This view is also enriched with other dataobtained from social networks. Learning methods, suchas context recognition, can be used to compare currentsituations with previous ones and help the system bemore adaptive and reactive.

5.4 Recommendations

The system, based on its distributed infrastructureas well as its cooperative tasks for getting accurateinformation on traffic, can issue some directives todrivers in terms of the most efficient direction totake or a travel route to avoid social events around arecommended itinerary, etc.

6 Conclusions and Future Work

In the Smart City context, integrating information fromdifferent sources is a necessity in real-time to makesmarter decisions for individual citizens and for the cityas a whole.

In this paper, we propose an approach for aframework, based on prediction models, that providesintegration, aggregation, and processing of real-time tracking data based on cloudlet architectureand cloud principles. By adopting a service-orientedapproach when collecting, integrating, storing, andintelligently analyzing social and user data, we achievebetter granularity, accessibility, recommendation, andprediction of user requirements. In addition, our




approach provides service continuity capabilities thatallow the system to provide value added informationthat suits the users’ requirements at any time and place.

The next steps involve the implementation ofthe proposed framework by realizing the scenariospresented here and using existing standards in the field.In order to evaluate this implementation, we will builda prototype to conduct user performance requirements,using multiple metrics that measure request-responsetime, performance, processing time, and other relevantmeasures.

Acknowledgements

This work was supported by the National Institute ofStandards and Technologies (NIST), and conducted withina collaboration under Information Technology Laboratory,Advanced Network Technologies Division (ANTD) andthe Universities of Grenoble and CNAM. Our specialthanks go to Dr. Abdella Battou, ANTD division chieffor his support and advises.

References

[1] M. Weiser, R. Gold, and J. S. Brown, The origins ofubiquitous computing research at PARC in the late 1980s,IBM Systems Journal, vol. 38, no. 4, pp. 693–696, 1999.

[2] R. Y. Clarke, Smart cities and the internet of everything:The foundation for delivering next-generation citizenservices, IDC Government Insights #GI243955, 2013.

[3] L. Heuser, C. Alsdorf, and D. Woods, eds. International

Research Forum 2008. Evolved Technology Press, 2009.[4] M. Satyanarayanan, G. Lewis, E. Morris, S. Simanta,

J. Boleng, and K. Ha, The role of cloudlets in hostileenvironments, IEEE Pervasive Comput., vol. 12, no. 4, pp.40–49, 2013.

[5] H. Tamai, Fujitsu’s approach to smart cities, vol. 50, no. 2,pp. 3–10, 2014.

[6] S. Dirks, C. Gurdgiev, and M. Keeling, Smarter cities forsmarter growth, IBM Global Business Services, 2010.

[7] B. Guo, Z. Yu, X. Zhou, and D. Zhang, From participatorysensing to mobile crowd sensing, in 2014 IEEEInternational Conference on Pervasive Computing andCommunication Workshops (PERCOM WORKSHOPS),2014, pp. 593–598.

[8] S. S. Ara, Z. U. Shamszaman, and I. Chong, Web-of-objects based user-centric semantic service compositionmethodology in the internet of things, InternationalJournal of Distributed Sensor Networks, vol. 2014, p.482873, 2014.

[9] D. Guedes, W. Meira, and R. Ferreira, Anteater: A service-oriented architecture for high-performance data mining,IEEE Internet Computing, vol. 10, pp. 36–43, 2006.

[10] M. Satyanarayanan, P. Bahl, R. Cceres, and N. Davies, Thecase for VM-based cloudlets in mobile computing, IEEEPervasive Computing, vol. 8, no. 4, pp. 14–23, 2009.

[11] G. Dimitrakopoulos, Intelligent transportation systemsbased on internet-connected vehicles: FundamentalResearch areas and challenges, in the 11th IEEEInternational Conference on ITS Telecommunications,2011, pp. 145–151.

[12] C. L. Liu and J. W. Layland, Scheduling algorithmsfor multiprogramming in a hard real-time environment,Journal of the ACM, vol. 20, no. 1, pp. 46–61, 1973.

O. Kotevska is a PhD student at JosephFourier University (UJF) in Grenoble,France and Guest Researcher at ITL LabNIST, Washington DC Metro, USA. Shereceived the MSc and BSc degrees fromthe University “Ss Cyril and Methodius”in Skopje, Macedonia in 2013 and 2008,respectively. Her research is in the field

of text mining, machine learning, cyber physical systems, andsoftware design.

S. Bouzefrane is an associate professorand has the accreditation to conductresearch (HDR) at the ConservatoireNational des Arts et Metiers of Paris.She received her PhD degree in computerscience from the University of Poitiers(France) in 1998. After four years at theUniversity of Le Havre (France), she joined

in 2002 the CEDRIC Lab of CNAM where she worked on real-time and embedded systems. She is the co-author of many books(Operating Systems, Smart Cards, and Identity ManagementSystems). Her current research areas cover service architecture

and security in mobile cloud computing, internet of things, andvirtualization.

Ahmed Lbath is a full professor ofcomputer science at University ofGrenoble1 (MRIM/LIG Laboratory),and he is also conducting research incollaboration with ITL Lab NIST inWashington DC metro area where hecarried out research activities as visitingprofessor. He is the IUT Deputy Director,

former Head of CNS Department, and former Director of R&Din a French company. He received his PhD degree in computerscience from the University of Lyon and hold an “HabilitationDiriger des Recherches”. He is currently acting as projectmanager coordinating scientific collaborations in the domain ofCyber Physical Systems and smart cities. His research interestsinclude cyber physical human systems, smart cities, mobilecloud computing, recommendation systems, web services, GIS,and software design. He published several patents, papers inbooks, journals, and conferences and he regularly serves asco-chair and/or member of several committees of Internationalconferences, journals, and research programs.



toward a real time framework in cloudlet-based architecture

Education