Performance optimization of edge computinghomeland security support applications

Marco GribaudoDipartimento di

Elettronica, Informaticae Bioingegneria

Politecnico di Milanovia Ponzio 51

20133, Milano, Italy

Mauro IaconoDipartimento di

Matematica e FisicaUniversita degli Studi

della Campania”L. Vanvitelli”viale Lincoln 5

81100 Caserta, Italy

Agnieszka JakobikInstitute of Computer

ScienceCracow University of

TechnologyWarszawska st 24,

31-155 Cracow, Poland

Joanna Ko lodziejResearch and Academic

Computer Network(NASK)

Kolska st 12, 01-045Warsaw, Poland

KEYWORDS

Edge computing; Cloud computing; Performanceevaluation; Emergency response; Queuing networks;Genetic algorithms.

ABSTRACT

Critical distributed applications have strict require-ments over performance parameters, that may affectlife of users. This is a limitation that may prevent theexploitation of cost effective solutions such as CloudComputing (CC) based architectures: in fact, the qual-ity of the connection with the CC facility and the lackof control on cloud resources may limit the overall per-formances of an application and may cause outages.A way to overcome the problem, and disclose the ad-vantages of CC to critical applications, is provided byEdge Computing (EC). EC adds local support to CC,allowing a better distribution of application tasks ac-cording to their timeliness requirements. In this paperwe present an innovative Special Weapons And Tac-tics (SWAT) support application, designed to empowereffective operations in wide scenarios, that leveragesEC to join CC elasticity and local immediateness, andwe exploit Queuing Networks (QN) and Genetic Algo-rithms (GA) to design and optimize the system param-eters for an effective workload distribution.

I. INTRODUCTION

Critical systems are systems characterized by strictrequirements (in general, on dependability, perfor-mances, or also security attributes), to avoid conse-quences that may be dramatic in terms of economicloss or threats to human life. In order to satisfy suchrequirements, the design process of the system shouldcarefully be planned, and the use of models to supportthe design of performance parameters during the lifecycle of the system may provide a significant help indesign choices.

In this paper, the focus is on performance model-ing of a reference architecture for a novel application,oriented to support complex SWAT operations in highrisk scenarios involving hostages and criminals or ter-rorists. The system, inspired to a firefighting support

system presented for the first time in [4], assists a num-ber of SWAT teams on the field by providing sensors,augmented reality tools, tactical information and lifeprotection support to enhance the effectiveness of theteam while raising the level of protection for agents,by a real time analysis of the conditions of the fieldand the possibility of computationally complex coordi-nation and data management applications to lower theimpact of possible menaces (e.g., terrorists hiding be-tween hostages or their organized reaction to the SWATaction).

In order to provide a flexible, yet cost effective solu-tion, that allows to combine highly available and timelytasks with heavy and variable workloads, to supportboth field oriented, tactical and strategic operations,the application is based on EC. This poses the problemof having an optimal dimensioning of the various com-ponents of the architecture according to the needs ofeach mission, or to the evolution of the scenario. Thisrequires the model to be optimized to find the right bal-ance between the different tasks on the system, and theevaluation and optimization have to be feasible with alow computational effort and in a short time, in orderto keep the system well tuned along all the duration ofthe mission.

The EC architecture has been chosen because of itshigh flexibility [12]. In the EC paradigm CC facilitiesare complemented with additional computing facilitieslocated where the application has to be used. These ad-ditional facilities are ”at the edge of the cloud” (hencethe name), so that they are close to the user and donot suffer from the problems that affect the commu-nications between the user and the CC facility. Theedge facilities may be common computers, mobile de-vices, low power/low cost devices (e.g. Raspberry Pi),or other devices.

As performances are a critical issue since the firstphases of the design cycle and along the life cycle of thesystem, and the system is subject to be deployed in dif-ferent alternative configurations in relation to scenar-ios, we complemented a performance evaluation tech-nique, based on QN, with a parameter optimizationtechnique, based on GA. The choice of QN is due to

Proceedings 32nd European Conference on Modelling and Simulation ©ECMS Lars Nolle, Alexandra Burger, Christoph Tholen, Jens Werner, Jens Wellhausen (Editors) ISBN: 978-0-9932440-6-3/ ISBN: 978-0-9932440-7-0 (CD)

the fact that QN are a consistent, yet simple way toconsider all relevant aspect of the system, easily un-derstandable for non specialists and, in general, widelysupported by available tools and easily included intomore complex modeling frameworks [2][1].

The choice of GA is due to their generality and thefast and reliable evaluation of the possible parameterconfigurations. They enable to find the near optimalsolution of the NP-hard optimization problems fasterthen traditional ’hard’ computing methods, [16], [17].

In this paper, we present a performance evaluationapproach for a complex, scalable, critical EC basedSWAT operation support and management system, de-signed to assist missions involving one or more SWATteams in high-risk scenarios, that leverages on wear-able and deployable sensors, augmented reality devicesand CC support for strategical mission supervision andadditional support, to allow an effective coordinationof SWAT teams, providing life parameters monitoring,extended environmental modeling and real time infor-mation enhancement.

The paper is organized as follows: after this intro-duction, Section II introduces the system, III describesthe modeling approach, IV presents performance eval-uation. Conclusions follow.

II. A CRITICAL EDGE COMPUTING BASEDSYSTEM

The proposed system is inspired to the one presentedfor the first time in [4] and described in details in [?].In these papers a different QN performance model hasbeen presented, together with some performance re-lated considerations, but no optimization support.

The architecture of the system leverages on the com-mercial availability of several key technologies, namelyWireless Sensors Networks (WSN) (including opticaland thermal cameras) and related Internet of Things(IoT) and Mobile Sensors Networks (MSN) [9][15],Augmented Reality (AR), Wearable Computing (WC),and the already introduced EC and CC. These tech-nologies have been considered, for their maturity leveland the existence of cost effective devices on the COTSmarket, to satisfy the 4 main functionalities that allowto accomplish the mission of the system: field control,agent support, tactical support and strategical support.For a matter of clarity, let us consider an example sce-nario: a terrorist attack into a skyscraper.

A. System functionalities

Field control consists in the capability of providingreal time information about what is happening on thefield. In our example scenario, the SWAT teams maydeploy wireless and/or mobile sensors and cameras totake control and tactical superiority in the building,and may wear sensors to monitor the environment; ad-ditionally, a (partial) supplemental view on the envi-ronment may be obtained by means of drones, eithermoving inside or inside the building, so to obtain a par-tition of the overall scenario into an observable and anon observable part, in which the observable part is di-

rectly monitored by the SWAT team leader. Addition-ally, the building may have an existing sensor networkor cameras, eventually partially damaged by the attack-ers, that may be somehow remotely accessible by themission leader, outside the action scope of the SWATteam leader.

Agent support is provided by the equipment that isgiven to each agent. For the purposes of this paper, thisequipment consists of wearable sensors to monitor theenvironment and the health conditions of each agent,and AR support to provide real time additional datasuperposed to the physical perception of the environ-ment, eventually integrated with a WC device to pre-process data. Data are exchanged in real time with theSWAT team leader, that reports to the mission leader,by means of the tactical support.

Tactical support consists in allowing each SWATteam leader to coordinate his team, control the observ-able part of the scenario and to take decisions harvest-ing the information about the current tactical situationon the field. The SWAT team leader should be able tocommunicate with his team and to control mobile sen-sors, including drones and cameras or other complexdevices. Tactical support is provided by a real timeinformation system running on a dedicated server, lo-cally hosted on the field (e.g. into a van) and capable ofperforming real time analysis on gathered sensor dataand providing updated AR information to the agentsbelonging to a SWAT team. With reference to the ex-ample scenario, one or more SWAT teams are on thefield, each one supported by a van in which the SWATteam leader operates, located in close proximity withrespect to the skyscraper so to avoid communicationproblems with the SWAT team.

Strategical support consists in providing the missionleader, which operates remotely, with updated infor-mation about the SWAT teams on the field and thefield itself, by merging sensed data and external datasources, and additional advanced data analysis tools(e.g. face recognition applications) to obtain strate-gical superiority and to evaluate the possible alterna-tive scenarios resulting from possible course of eventsand the consequences of the various options that theSWAT team leader may decide to examine before tak-ing his decisions. In the example, the commander mayuse a mission supervision environment leveraging in-formation from the various SWAT teams and from ex-ternal sources such as the internal sensing system ofthe skyscraper, accessed by the World Wide Web, andHomeland Security department (or equivalent govern-mental or non-governmental institutions) databases tosimulate the course of events or wider scenarios, and theeffects of different actions of the SWAT teams. Process-ing is not real time, but may require elastic computingresources and interfacing with other local or nationalauthorities, such as police, air force, secret services andthe aforementioned Homeland Security authorities.

B. System architecture

In order to implement the 4 functionalities of the sys-tem, the architecture is organized into 4 main subsys-tems, namely Cloud Backend (CB), Edge Frontend(s)(EF), Personal Support (PS), Sensing Network (SN):they are meant to provide the implementation plat-form respectively for the strategical support, the tac-tical support, the cyberfireman support and the fieldcontrol. A representation of a possible configuration isin Fig. 1.

Fig. 1. System architecture layout

CB provides the needed elasticity for the executionof the needed strategical superiority and mission super-vision applications, that may vary from a simple de-cision support system to complementary applicationssuch as signal processing, OLAP, image recognition,simulation, integrating field sensed data and externalsources accessed on line, obtained from other authori-ties in real time and/or other databases already avail-able in the system or special purpose applications. Inour example, the CB may run advanced signal process-ing applications to find signals into the data complexprovided by the wearable sensors that agents cannotperceive, such as the voices of nearby hostages or ofterrorists, not directly audible by the agents, or canperform data fusion between data from SN and PS andfrom the building sensor network to provide referencesabout the actual viability of the skyscraper or the po-sition of people by thermal sensing, and integrate ARdata or lower the complexity of real time generation ofAR data for the EF. Consequently, workload may sig-nificantly vary according to the phases of the missionand different needs, that may in principle be not definedbefore they actually manifest. An important parame-ter is the volume of service requests coming from theEFs and the amount of supervisory requests generatedon the CB by the mission leader (and, possibly, otherauthorities).

EF provides computing resources for the real timeexecution of the SWAT team coordination tasks. It isbasically provided by a high performance server (e.g.as in [5]), that is connected to the CB by a high speedmobile connection (e.g. 4G or 5G) and is installed on avan. In the following, we will not consider the connec-tion problems with PS and SN that may arise becauseof the difficult conditions, as it is out of the scope ofthis paper: we hypothesize, without loss of generality,

that the quality of the connection is kept with the useof proper repeaters (e.g. by setting up an ad-hoc net-work). The EF workload is supposed to be regular anddependent on the number of agents in a SWAT teamand on the number and type of sensors in the controlledSN.

PS provides an agent (similarly to what presentedin [8]) what is needed to implement wearable sensingand AR by means of a WC system. A minor com-puting effort is performed locally, basically devoted toservice tasks and devices management, and communi-cations with the related EF. Data from sensors (e.g.position, health monitoring, environmental sensors andadditional deployable ones), with partial preprocessing,are sent to EF, and commands and AR information arereceived from the EF. Workload is mainly related to theconfiguration of the equipment, and sensors contributeto the workload of the EF.

SN consists in a traditional sensor network, in whichsensors may be fixed, mobile, disposable, actuable, mul-timedia or complex sensors. All sensors are controlledby the EF, and affect its workload. The kind and num-ber of sensors in the network may vary, even duringa same mission, because of damages, of new needs, ofenergy exhaustion, of replacement (e.g. see [9][15]).

According to the scenario, the actual configurationmay vary, and proper criteria for a correct parametersetting is crucial in order to ensure that performancesare sufficient to match the requirements. A configu-ration will be composed of one CB, one EF per eachSWAT team operating in the field, one PS per eachagent and one or more SN for each EF.

III. PERFORMANCE MODEL

We model system with the multi-class open queue-ing network shown in Figure 2. Queuing stations areused to model EF nodes (Edge Frontend), networkcommunication (Edge Network) and CC nodes (CloudBackend). Arrivals to the Edge Frontend nodes repre-sent data acquired from the field. Since different data,of different complexity, can be collected at differentspeeds, we use costumer classes to model traffic het-erogeneity. Without loss of generality, in this work wefocus on two different classes of data: the one gener-ated by SN devices, and the one originating from PSnodes. We call λSN and λPS respectively the data rateof the two sources. Since we consider scenarios wheremore EF are present, we model each one with a dif-ferent queuing station. We call NEF the number offrontend nodes, and we suppose that each of them ischaracterized by the number of devices it receives from/ sends to: for this reason we might have different ar-rivals for each EF station. For sake of simplicity, weconsider EF nodes to be homogeneous, all character-ized by the same arrival rate from same classes, andeach node providing the same performance and thusserving its requests in the same amount of time. How-ever, the service time for different traffic classes mightdiffer: in particular we call SEF,SN and SEF,PS the av-erage service time required by the two considered data

acquisition classes. To simplify the evaluation of thesystem, we suppose that EF nodes are equipped by anoperating system that, when more requests needs to beserved, processes them concurrently. In this way EFnodes can be modeled adopting the processor sharingservice discipline.

All nodes send data to the cloud using the same net-work connection, hence communication is modeled witha single queuing station to which all the EF nodes areconnected. Service times might be different for the con-sidered sensor classes, and in particular we set them toSN,SN and SN,PS . The main feature of edge computingis that data collected by the sensors is first analyzed bythe EF nodes, and only a fraction of them is requiredto be sent through the network. We model this by al-lowing jobs of the two classes to immediately leave theEF stations respectively with probabilities 1−pSN and1 − pPS . In this way only a fraction pSN and pPS ofdata produced by the sensors reaches the cloud.

The cloud part of the application is considered to dis-tributed between NC equivalent virtual machines thatshare the arriving load in an uniform way. Besidesdata received from the EF nodes, the cloud backendis queried also by other users that monitor the eventfrom different institutions. This is modeled by addingan extra class, characterized by a global arrival rate λC .Traffic generated by this particular class immediatelyleaves the system after being served. Data incomingfrom the sensors (SN and PS classes) go instead backto the EF nodes to allow feedback from the centralinfrastructure to the server deployed on the field. Wecall respectively SC,SN , SC,PS , SC,C the average servicetime required by the three considered classes. Table IIIsummarizes model parameters and shows the base-linevalue considered in the following part of the paper. Val-ues have been validated as realistic from experts of thefield.

Edge

Frontend

Cloud

Backend

Edge

Network

Edge

Frontend

Edge

Frontend

...

Cloud

Backend

...

Fig. 2. Queuing network model for the system

IV. SCENARIOS AND ANALYSIS

The model presented in Figure 2 can be used to studyseveral deployment scenarios. We will focus on a coupleof them: maximizing the system computational power,while minimizing the cost for acquiring resources; andconsidering the impact of the coordination activity onthe cloud on the system response time.

Param. Description ValueNEF Num. of EF nodes 2λSN Arrival rates of 1.25 data/s.λPS SN and PS classes 1.25 data/s.

SEF,SN Service times 0.2 s.SEF,PS at the EF nodes 0.35 s.pSN Probability of 0.18pPS going to the cloud 0.12SN,SN Network 10 ms.SN,PS transfer times 15 ms.NC Number of VMs 2λC Coordination traffic 0.1 req./s.

SC,SN Service times 1.5 s.SC,PS at the cloud 1 s.SC,C backend 2 s.

TABLE I: The model parameters

A. MAXIMIZING SYSTEM POWER

As defined by Kleinrock in [19], system power cor-responds to Π = X

R : the ratio of the throughput andthe response time of a system, or of one of its com-ponents. The main rationale about this performanceindex is that it increases when either the throughputincreases, or when the response time decreases. In otherwords, the configuration in which a component worksat its maximum system power corresponds to the pointwhere users receives the best tradeoff between speedand latency. In this first study we want to determinethe point that maximizes system power, considering afixed total data generation rate. In particular, we imag-ine that data is collected at a total rate of Λ readingsper second, which is a structural parameter that de-pends on the specific sensing technology and commu-nication protocol between the sensors and the EF thatis employed. The choice is then to properly balancereading from the two class of sensors. In particular, wecall 0 ≤ β ≤ 1 the traffic mix parameter, and we defineλSN = βΛ and λPS = (1 − β)Λ. Another parameterwe want to take into consideration is the number of EFnodes that is being used. Indeed, the higher the num-ber of EF, the higher will be the system power of theapplication. However, considering a larger number ofEF nodes implies a higher cost: in particular we con-sider a discount function gµ(NEF ), that describes theimpact of having a higher cost due to a larger numberof EF nodes, on the average system power we aim tomaximize. In this example we set gµ(NEF ) = µ

NEF,

with µ > 0. The rationale is that a smaller numberof nodes has an higher impact on the system power weaim to maximize. We thus create an objective functionf1(β,NEF ) defined as:

f1(β,NEF ) = αλSN

RSN (β,NEF )+

(1 − α)λPS

RPS(β,NEF )+

µ

NEF(1)

where RSN (β,NEF ) and RPS(β,NEF ) represent theresponse time of the two sensor data classes. Constantα is used instead to balance the interest between thesensors (α = 1) and the personal devices (λ = 0).Figure 3 shows the system power of the two data classesand the value of the objective function 1 for differenttraffic mixes β and for a different number of nodesNEF .As it can be seen, the system power ΠSN of the sen-sor network has a monotonic behaviour, while the oneof the personal support class ΠPS shows a maximumwhen the number of EF nodesNEF = 1. The figure alsoshows the values of the objective function f1(β,NEF )for α = 0.5, and µ = 0.5 or µ = 1, to consider twodifferent cost scenario where data classes are equallyimportant. As it can be seen, in some cases behaviouris monotonic, while in some other it experiences someinternal convexity. This motivates the use of optimiza-tion algorithms to study the values of β and NEF wherethis maximum is reached.

Figure 4 shows the different values of the parameter

ΠSN

ΠPS

f1, µ=0.5

f1 µ=1.5

1

2

3

4

NEF

0

0.2

0.4

0.6

0.8

1

β

0

0.5

1

1.5

2

2.5

f1

Fig. 3. Ibjective function f1(β,NEF ) for different values ofresource cost parameter µ and α = 0.5.

that maximizes the objective function for different costparameters µ, and the balance parameters α. When thecost is very high, it is better to keep as few EF nodes aspossible: however, depending on the balance parameterα there could be optimal non-trivial traffic mixes thatoptimizes the system. When the cost is very low, it isnot always true that it is better to use a larger numberof EF nodes: the balance parameter α can make thesystem work at its optimum configuration even withNEF = 1, and NEF > 2 are never needed.

B. THE EFFECT OF THE COORDINATORS

Different players in the management of the eventmonitor the evolution of the scenario accessing thecloud back-end of the application. Depending on theirupdate rate λC , they might have a clearer and more up-to-date picture of the scenario, increasing their abilityto help and their effectiveness in solving the situation.However, a larger upload rate might overload the sys-tem, leading to an unstable behaviour. In this study,we consider the effect on the response time of the trafficgenerated by the coordinators, varying 0.1 ≤ λC ≤ 0.5

β

NEF

0 1

2 3

4 5

µ 0 0.2

0.4 0.6

0.8 1

α

0

0.5

1

1.5

2

Opt

Fig. 4. Optimum values of β and NEF as function of the costparameter µ, and the balance parameter α.

req./s., and we combine the analysis with the traf-fic mix introduced in the previous section. We mea-sure the non-effectiveness of the coordination actionswith a function lθ(λC) > 0: larger values denotes aless effective coordination of forces, while values thattends to 0 denotes optima interactions. In this workwe use lθ(λC) = θ

λC, where θ > 0 is a scaling pa-

rameter to make the measure compatible with the ex-pected response times. Again we want to account thecost on the infrastructure in our performance assess-ment. In this case, however, since response time mustbe minimized, it must increase with the number of re-sources. Moreover, since we are studying the coordina-tors, which mainly affect the cloud, we must accountnot only for the number of EF nodes NEF , but alsofor the number of provisioned VMs NC . In particu-lar, we use a cost function hν(NEF , NC), and we set itto hν(NEF , NC) = ν · (NEF + NC) (with ν > 0 as ametric conversion parameter), and we set the objectivefunction f2(β,NEF , NC , λC) to:

f2(β,NEF , NC , λC) =λSNΛ∗ RSN (β,NEF , NC) +

λPSΛ∗ RPS(β,NEF , NC) +

λCΛ∗RC(β,NEF , NC) +

θ

λC+ ν · (NEF +NC) (2)

where RC(β,NEF , NC) is the response time of thecoordinator class, and Λ∗ = Λ + λC is the total systemthroughput.Figure 5 shows the value of the objective functionf2(β,NEF , NC , λC) for different traffic mix β and co-ordinator traffic λc in different resource configuration.As it can be seen, there are many cases in which thesystem is not stable (represented in the figure withf2(β,NEF , NC , λC) = 0), and the objective functionis concave in some cases. Figure 6 shows for which val-ues of β and λC the minimum of the objective functionis reached. It is interesting to see that, when the cost

NEF=1, NC=1

NEF=1, NC=2

NEF=2, NC=1

NEF=2, NC=2

NEF=3, NC=3

NEF=4, NC=4

0 0.2

0.4 0.6

0.8 1

β 0.1

0.2

0.3

0.4

0.5

λc

0

1

2

3

4

5

f2

Fig. 5. Objective function f2(β,NEF , NC , λC) for different com-binations of EF nodes NEF and cloud VMs NC .

is very high; the minimum is obtained for β = 1 andλC = 0.18 req./s. When the cost is lower, the minimumis obtained in different combinations of β and λC . Theoptimization algorithm also determined the optimumnumber of nodes NEF and NC : even if not explicitlyshown in the figure to make it more readable, the mini-mum of the objective function is reached when the min-imum number of resource is used when β < 1. In theflat area with β = 1 instead, the optimum is reachedwhen the maximum number of resources is used.

β

λc

0 0.002

0.004 0.006

0.008 0.01

ν 0

0.0025

0.005

0.0075

0.01

θ

0

0.2

0.4

0.6

0.8

1

Opt

Fig. 6. Optimum values of β and λC as function of the costparameter ν, and coordination effectiveness θ.

V. RELATED WORKS

The integration of Cloud computing, mobile comput-ing and Internet of Things (IoT) allows to exploit agreater flexibility in performance engineering and tun-ing with respect to cloud architectures, and extendstheir great available power and elasticity by allowingan efficient collection and transmission of massive andubiquitous data sources, such as IoT devices and mobiledevices, to overcome the limitations of the weak pointof cloud architectures, the network connection towardsthe edge of the cloud. The readers may refer to [23] fora first approach to edge architectures problems and fea-tures, while the general context of application of these

solution is presented in [12] and [21], that report aboutsecurity related and law related topics. Of course, acorrect balance, and related tuning and adaptation, ofsystem parameters is crucial to achieve desired over-all performance figures, as discussed in [22], [24], [26],[20] and [6], that may help readers in understandingthe general framework in which responsiveness, scala-bility, privacy and fault tolerance in edge architecturesmay be achieved and modeled or measured. Due tothe interest about edge-based solutions, there is an in-tense research activity that spans over performancesand system/software management, communication pro-tocols and architectural solutions. In [10] the fog com-puting perspective and some related applications arepresented. In [18] the role of containers in edge-basedsoftware architectures is discussed and quantitativelydocumented, and in [14] the general software architec-tural aspects are examined. In [7] the focus is on therelation of edge architectures with IoT, with specialattention to communication problems. In [11] an au-tonomic approach to edge architecture management isdiscussed. In [25] the edge approach is extended to-wards osmotic computing. For the aspects focusingspecifically on cloud related problems and performancemodeling, readers may refer to other previous papersof the authors, stating their positions that influencedthis work ([16], [17], [3], [13]. In the first two papersauthors successfully applied a specifically designed GAfor solving NP-hard task scheduling problems for Com-putational Clouds. Their novel approach assumed gen-erating every consecutive population from ’freezing’ theworst parents instead of removing them from the pop-ulation. New population was of the equal size as theprevious population. This enabled to generate morebroad sub-optimal solutions space and hence find thesolution faster.

VI. CONCLUSIONS AND FUTURE WORK

Results show the effectiveness of the presented mod-eling approach for the performance evaluation of criti-cal edge based systems, that showed a non trivial be-havior. Future work include a better exploitation of GAin the technique and a more detailed characterizationof the system.

VII. ACKNOWLEDGEMENTS

This article is partially based upon work from COSTAction IC1406 High-Performance Modelling and Simu-lation for Big Data Applications (cHiPSet), supportedby COST (European Cooperation in Science and Tech-nology)

References

[1] S. Balsamo and A. Marin. Queueing networks. InM. Bernardo and J. Hillston, editors, Formal Methods forPerformance Evaluation, 7th International School on For-mal Methods for the Design of Computer, Communication,and Software Systems, SFM 2007, Bertinoro, Italy, May28-June 2, 2007, Advanced Lectures, volume 4486 of LectureNotes in Computer Science, pages 34–82. Springer, 2007.

[2] E. Barbierato, M. Gribaudo, and M. Iacono. Exploiting mul-tiformalism models for testing and performance evaluationin SIMTHESys. 2011.

[3] E. Barbierato, M. Gribaudo, and M. Iacono. Modeling andevaluating the effects of Big Data storage resource allocationin global scale cloud architectures. International Journal ofData Warehousing and Mining, 12(2):1–20, 2016.

[4] E. Cavalieri d’Oro, S. Colombo, M. Gribaudo, M. Iacono,D. Manca, and P. Piazzolla. Modeling and evaluating per-formances of complex edge computing based systems: a fire-fighting support system case study. In Valuetools 2017,,2018.

[5] D. Cerotti, M. Gribaudo, M. Iacono, and P. Piazzolla. Mod-eling and analysis of performances for concurrent multi-thread applications on multicore and graphics processingunit systems. Concurrency and Computation: Practice andExperience, 28(2):438–452, 2016. cpe.3504.

[6] H. Chang, A. Hari, S. Mukherjee, and T. V. Lakshman.Bringing the cloud to the edge. In 2014 IEEE Conference onComputer Communications Workshops (INFOCOM WK-SHPS), pages 346–351, April 2014.

[7] M. Chiang and T. Zhang. Fog and iot: An overview ofresearch opportunities. IEEE Internet of Things Journal,3(6):854–864, Dec 2016.

[8] S. Colombo, M. Gribaudo, M. Iacono, D. Manca, and P. Pi-azzolla. A low-cost distributed IoT-based augmented real-ity interactive simulator for team training. In Proceedings ofthe 31th European Conference of Modelling and Simulation,May 23-26 2017, Budapest, page to appear, 2017.

[9] M. D’Arienzo, M. Iacono, S. Marrone, and R. Nardone. Es-timation of the energy consumption of mobile sensors in wsnenvironmental monitoring applications. In Advanced Infor-mation Networking and Applications Workshops (WAINA),2013 27th International Conference on, pages 1588–1593,March 2013.

[10] A. V. Dastjerdi and R. Buyya. Fog computing: Helping theinternet of things realize its potential. Computer, 49(8):112–116, Aug 2016.

[11] M. Desertot, C. Escoffier, and D. Donsez. Towards an au-tonomic approach for edge computing: Research articles.Concurr. Comput. : Pract. Exper., 19(14):1901–1916, Sept.2007.

[12] C. Esposito, A. Castiglione, F. Pop, and K. K. R. Choo.Challenges of connecting edge and cloud computing: A se-curity and forensic perspective. IEEE Cloud Computing,4(2):13–17, March 2017.

[13] M. Gribaudo, M. Iacono, and D. Manini. Improving relia-bility and performances in large scale distributed applica-tions with erasure codes and replication. Future GenerationComputer Systems, 56:773 – 782, 2016.

[14] Z. Hao, E. Novak, S. Yi, and Q. Li. Challenges and softwarearchitecture for fog computing. IEEE Internet Computing,21(2):44–53, Mar. 2017.

[15] M. Iacono, E. Romano, and S. Marrone. Adaptive monitor-ing of marine disasters with intelligent mobile sensor net-works. In Environmental Energy and Structural MonitoringSystems (EESMS), 2010 IEEE Workshop on, pages 38–45,2010.

[16] A. Jakbik, D. Grzonka, and F. Palmieri. Non-deterministicsecurity driven meta scheduler for distributed cloud organi-zations. Simulation Modelling Practice and Theory, 76:67 –81, 2017. High-Performance Modelling and Simulation forBig Data Applications.

[17] A. Jakobik, D. Grzonka, J. Ko lodziej, and H. Gonzalez-Velez. Towards secure non-deterministic meta-schedulingfor clouds. In 30th European Conference on Modelling andSimulation, ECMS 2016, Regensburg, Germany, May 31 -June 3, 2016, Proceedings., pages 596–602, 2016.

[18] M. B. A. Karim, B. I. Ismail, W. M. Tat, E. M. Goor-tani, S. Setapa, J. Y. Luke, and H. Ong. Extending cloudresources to the edge: Possible scenarios, challenges, andexperiments. In 2016 International Conference on CloudComputing Research and Innovations (ICCCRI), pages 78–85, May 2016.

[19] L. Kleinrock. Power and deterministic rules of thumb forprobabilistic problems in computer communications. In ICC’79; International Conference on Communications, Volume3, volume 3, pages 43.1.1–43.1.10, 1979.

[20] H. D. Park, O.-G. Min, and Y.-J. Lee. Scalable architecturefor an automated surveillance system using edge computing.J. Supercomput., 73(3):926–939, Mar. 2017.

[21] R. Roman, J. Lopez, and M. Mambo. Mobile edge comput-ing, fog et al.: A survey and analysis of security threats andchallenges. Future Generation Computer Systems, 2016.

[22] M. Satyanarayanan. The emergence of edge computing.Computer, 50(1):30–39, Jan 2017.

[23] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu. Edge com-puting: Vision and challenges. IEEE Internet of ThingsJournal, 3(5):637–646, Oct 2016.

[24] W. Shi and S. Dustdar. The promise of edge computing.Computer, 49(5):78–81, May 2016.

[25] M. Villari, M. Fazio, S. Dustdar, O. Rana, and R. Ran-jan. Osmotic computing: A new paradigm for edge/cloudintegration. IEEE Cloud Computing, 3(6):76–83, Nov 2016.

[26] S. Yi, Z. Hao, Z. Qin, and Q. Li. Fog computing: Platformand applications. In 2015 Third IEEE Workshop on HotTopics in Web Systems and Technologies (HotWeb), pages73–78, Nov 2015.

AUTHOR BIOGRAPHIES

MARCO GRIBAUDO is an as-sociate professor at the Politecnico diMilano, Italy. He works in the perfor-mance evaluation group. His currentresearch interests are multi-formalismmodeling, queueing networks, mean-field analysis and spatial models. The

main applications to which the previous methodologiesare applied comes from cloud computing, multi-core ar-chitectures and wireless sensor networks.

MAURO IACONO is a tenured as-sistant Professor and Senior Researcherin Computing Systems at Dipartimentodi Matematica e Fisica, Universita degliStudi della Campania ”L. Vanvitelli”,Caserta, Italy. His research activityis mainly centered on the field of per-

formance modeling of complex computer-based sys-tems, with a special attention for multiformalism mod-eling techniques. More information is available athttp://www.mauroiacono.com.

AGNIESZKA JAKOBIK (KROK)received her M.Sc. in the field ofstochastic processes at the JagiellonianUniversity, Cracow, Poland and Ph.D.degree in the field of neural networksat Tadeusz Kosciuszko Cracow Univer-sity of Technology, Poland, in 2003 and

2007, respectively. From 2009 she is an Assistant Pro-fessor. Her e-mail address is: agneskrok@gmail.com

JOANNA KO LODZIEJ is an as-sociate professor in Research and Aca-demic Computer Network (NASK) In-stitute and Department of ComputerScience of Cracow University of Tech-nology. She is a Head of the De-partment for Sciences and Develop-

ment. She serves also as the President of thePolish Chapter of IEEE Computational IntelligenceSociety. She is also a Honorary Chair of theHiPMoS track of ECMS. Her e-mail address is:joanna.kolodziej68@gmail.com. The detailed informa-tion is available at www.joannakolodziej.org