IET Computers & Digital Techniques

Research Article

Evaluation of simulator tools and power-aware scheduling model for wireless sensornetworks

ISSN 1751-8601Received on 4th January 2017Revised 15th March 2017Accepted on 13th April 2017E-First on 9th August 2017doi: 10.1049/iet-cdt.2017.0003www.ietdl.org

Rym Chéour1 , Mohamed Wassim Jmal1, Olfa Kanoun2, Mohamed Abid1

1National School of Engineers of Sfax, Computer and Embedded System Laboratory, University of Sfax, Sfax, Tunisia2Measurement and Sensor Technology, Chemnitz University of Technology, Chemnitz, Germany

E-mail: cheourr@gmail.com

Abstract: The sharp increase of the wireless sensor networks (WSNs) performance has increased their power requirements.However, with a limited battery lifetime it is more and more difficult to deploy many more sensors with today's solutions.Therefore, the authors need to implement autonomous WSNs without any human intervention or external power supply. To thisend, this study proposes an effective strategy to ensure an energy consumption gain that takes into account time constraintsthrough a power-aware model based on the dynamic voltage and frequency scaling and the dynamic power management thatare appropriate to the WSNs and on a global Earliest Deadline First scheduler. To select the most suitable simulator to integrateand simulate the developed models, >25 of the existing WSN simulators are outlined and evaluated. On the basis of thiscomparative study analysis, the authors chose the simulation tool for real-time multiprocessor scheduling (STORM) to validatetheir work for its multiple advantages.

1 IntroductionWireless sensor networks (WSNs) are the backbone of theconnected industry and have managed to make their way in theInternet of Things (IoT) [1]. WSNs integrating the IoT need to behighly flexible, scalable, interoperable and reliable. They invaded avery wide range of applications such as healthcare, military,environment supervision [2], precision agriculture [3], industry [4],aerospace [5] and so on. Nevertheless, a sensor node would drainthe battery more quickly in the absence of effective energyoptimisation techniques causing a break-down of the wholenetwork [6]. This fact has motivated researchers to developprotocols and mechanisms to reduce its use in all layers of theprotocol stack [7]. However, the nodes are increasingly providedwith advanced skills as well as additional services, such as the useof a global positioning system [8], multimedia or ultrasoundsensors. These features allow the emergence of complex andenergy-intensive applications such as cryptography which hasheavy constraints on the processing power. In this case, it becomesdifficult to apply security algorithms in some applications becausethe process of encryption and decryption consumes a lot of timeand energy [9]. Hence, time constraints become an influencingfactor and then strict deadlines should be ensured so that the timebetween the detection of an anomaly and the operator interventionwould prevent the incident. A particularly significant example isthe sudden increase of water pressure or oven temperature in aturbine that can quickly get critical [10]. Consequently, combiningthe energy optimisation techniques and scheduling is one way toincrease energy efficiency. Several experiments to reduce theenergy in WSNs are conducted. The authors in [11] investigatedthe low-power circuit design. Lamonaca et al. [12] proposed anovel WSNs synchronisation protocol with a trade-off betweenpreserving the lifetime of the whole network and highsynchronisation accuracy. Sankar Ramachandran et al. [13]presented a generic energy-aware reconfiguration model able tomeasure and model the energy cost of reconfiguration. Real-timescheduling was proposed by Harkut et al. [14] through theimplementation of an Earliest Deadline First (EDF) policydedicated to WSNs. The recent works of Kulau et al. [15] havedefined an energy model with undervolting capacities to reduce thetemperature profiles in a Smart Farming application. However, theabove techniques treat communication and routing protocols and

do not investigate the processor optimisation. Also, the harvestingopportunity for the node is less than its energy requirement.Besides, using solar panel is only operational outdoor which is notpractical and lacks of accuracy.

The two objectives of this paper are defining an energy-awarescheme and giving a comprehensive survey of the WSNsimulators. Indeed, unlike approaches based on a single strategythat may not adapt to all of the applications nor to the requiredmetrics, the power manager incorporated a hybrid approachcombining several multiobjective strategies: time-out2 dynamicpower management (DPM), intertask dynamic voltage andfrequency scaling (DVFS) and global EDF (GEDF). Thescheduling algorithm carries out the tasks so that the duration ofthe idle periods is optimised to activate the low-consumption statesthrough DPM. For more optimality, we invest in new condition tobypass the Dhall effect [16]. In corollary, during periods of activity,the voltage and frequency couple is modified while taking intoaccount the deadlines of the tasks. In our case, we will use the‘GEDF’ strategy. The theoretical results on multiprocessorscheduling show that a global scheduling has a better exploitationof the resources compared to a partitioning approach [14].

The second main objective of our work is to review theavailable WSN simulators with the aim of determining which is themost suitable for adapting to our purpose. We will define a high-level taxonomy of 25 simulators to have a better insight of theirfeatures. The classification method is developed according to manycriteria including their taking into account or not the criterion ofenergy, whether the simulator is generic, code, node or hybridoriented. The use of simulation will bring a great saving and willincrease efficiency at all levels of the WSNs.

The remainder of the paper is structured as follows. First,Section 2 surveys WSNs simulators and exposes a comparativestudy of different simulators treating energy. Section 3 presents arealistic power consumption model for WSNs. Next, Section 5reports some results with the STORM simulator. A discussion isgiven in Section 6. Section 7 includes a conclusion.

2 Comparative study of WSNs simulatorsThe simulation technologies have been widely used mainly, whenthe experimental approaches are difficult to achieve due to thecomplexity of the qualitative and quantitative changes applied

IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

173

frequently to control the behaviours of the physical platform [17].The simulation techniques aim to describe the behaviour of thenode [18]. So, this comparative study provides an analysis ofdifferent simulator tools and classifies them according to theenergy criterion. Fig. 1 shows a classification of the simulatorsaccording to these criteria.

Avrora [19] is a cycle-accurate instruction-level, large-scalesensor network simulator, providing the highest behavioural andthe most detailed timing description for the WSNs. Contiki OSJava (COOJA) [20] is a cross-level simulator including networkingand instruction simulation and is based on the ContikiOS. EmStar[21] is suitable for sensor nodes running on Linux and is a real-time simulator. It develops software for iPAQ-based microserverand MICA2 Motes. GLoMoSim [22] is an open-source, large-scale, scalable, extensible, parallel and packet-level simulator. NS2[23] is used to simulate routing protocols. It is general-purpose,extensible and network-level simulator. The OMNet++ [24] is a C++ discrete-event, general-purpose network simulation systemsupporting heterogeneity, media access control (MAC) and somelocalised protocols with graphical runtime environment. Ptolemy[25] is a simulator that models the distributed hybrid systems andconsiders the software components and its interaction with thehardware components of the node (time model, messages exchangeetc.). Optimized network engineering tool (OPNET) [26] is adiscrete-event, general-purpose, network behaviour simulator ableto define and test protocols, nodes and process models. Prowler[27] is an environment and wireless medium that is probabilistic,event-driven, communication-level and running in Matlabsimulator with easy and fast application prototyping. QualNet [22]is a derivative of GloMoSim modelling a large-scale network(thousands of nodes) thanks to the parallel processing. It provides agraphical user interface and has high scalability. Shawn [26] is fast,highly customisable and an algorithm-level simulator thatsimulates the impact of a phenomenon. Tiny OS simulator(TOSSIM) [28] is dedicated to TinyOS applications running onlyon MICA motes. It simulates the transmission conditions of largenumber of nodes but not with a real-environment condition.UWSim [29] is used for underwater sensor enabling underwatercommunication. It has high frequency, specific routing protocoland low bandwidth. A comparison of the non-power-aware WSNssimulators is made in Table 1.

2.1 Overview of the power-aware simulators

Various surveys were developed to compare the performances ofthe simulators but are still confined to the same tools of simulationsand neglect time-dependent issues, like energy consumption andradio channel utilisation. We will focus on power profiling inWSNs simulators and we classify them according to their ability tomodel the energy.

2.1.1 Accurate prediction of power consumption(AEON): AEON [35] is built on top of the Avrora simulator [19].It is based on real code which is as close as possible to an actualexecution. It is used to estimate the energy on variouscryptographic and security algorithms to transmit long and

protected messages. Being based on a real performance, it avoidserroneous lifetime estimation and coarse approximations forWSNs.

2.1.2 Atmel emulator (ATEMU): ATEMU [36] is the first and themost accurate instruction-level emulator. It is designed for MICAplatform and TinyOS. It allows parallel simulation of large amountof nodes on heterogeneous network and with distinct applications.It provides a graphical user interface (GUI) and realistic results. Itsupports the processor instructions, the radio description, timersand so on. However, it suffers from low scalability and longsimulation time which is 30 times slower than TOSSIM.

2.1.3 EnergySim: EnergySim [37] is a versatile event-driven andextensible WSNs MAC protocol simulator. It is based on C# visualprogramming language. It is a custom build simulator dedicated forapplications requiring different kinds of nodes and deployed indifferent environments. It generates Excel worksheets and plots ofthe energy states. It supports multiple topologies and offers an easyuser interface but does not follow a physical model.

2.1.4 Hierarchical design platform for sensor networksexploration (IDEA1): IDEA1 [38] is a SystemC-based, system-level, design and simulation framework for WSNs. This simulatorhas the ability to model both the networks, and the software andhardware components. Also, it allows a concise estimation of thepower consumption, thanks to its ability to support all the powermodes and the transitions between them. However, it has low datarate and necessitates a strong microcontroller and its correspondentcommunication protocol.

2.1.5 J-Sim: J-Sim [39] is a general-purpose, cross-level, large-scale and discrete-event simulator handling the communication, themobility and the energy levels (battery, CPU, radio, protocol stacketc.). Each node entity is seen as an independent component. Thesimulator provides a trace file and the interconnection among thenodes. However, its execution time is much longer than othersimulators and is of high complexity.

2.1.6 Power aware wireless sensors (PAWiS): PAWiS [40] is adiscrete-event WSN simulator based on OMNeT++ that carries outa module with the characteristics of the CPU. It takes intoconsideration the mobility of the nodes, the environmentaldynamic, the routing protocol and the hardware design issues. Itsimulates the internal structure, the surroundings and the powerconsumption of every single node. The output is a timing andpower consumption profiles of the model.

2.1.7 PowerTOSSIM: PowerTOSSIM [28] is a power modellingextension of TOSSIM. It keeps a trace of the power state, thetransitions messages, the processor energy usage and estimates thenumber of cycles in AVR. Although, it benefits from the scalabilityof TOSSIM, it loses some accuracy on capturing the interrupts.Being limited to MICa2, the node model does not execute the samecompiled code as real sensor nodes.

Fig. 1  Taxonomy of simulators for WSNs

174 IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

2.1.8 Sensor network simulator and emulator(SENSE): SENSE [41] accounts the dynamic battery effects andconsiders the MAC protocols. The ‘SimplePower’ componenttakes one of five modes which are transmit, receive, idle, sleep andoff. It is characterised by extensibility, reusability and a modelling

methodology. Nonetheless, the simulator engenders hard and nonerelevant outputs and depends on obsolete libraries due to absenceof active development.

Table 1 Comparison of the non-power-aware WSNs simulatorsSimulator Language Network HW SW Licence FeaturesAvrora [19] Java x xa open +simulation of thousands of nodes

+cycle accurate and fast−mobility

−device specificCastalia [30] C++ NED x academic +detailed radio state transition

+realistic wireless medium and communication interface−non-portable on node platforms

COOJA [20] Java x x x BSD +hardware and software simulation+different types of propagation models

−low efficiency−limited simultaneously simulated nodes

EmStar [21] C x Libre +same source code as on actual sensors+simple environmental model and network medium

−restricted number of simulated sensors−restricted to some platforms

GLoMoSim [22] PARSECb x open +extendable, large scalability+protocols designed to WSNs

−lack of accuracy−poor documentation

QualNet [22] PARSEC x commercial +high fidelity and modular network design+GUI and user-friendly tools−expensive licence annual

NS2 [23] C++ x GNUGPL +wide range of protocols+extendable

−low scalability, complexity and time consuming−no GUI

−lack of packet formats, energy models and MACOMNet++ [31] C++ x academic +component-based, hierarchical, modular and extendable architecture

−poor analysis of performance measures−limited available protocols

OPNET [32] C, C++ x commercial +handling large records+fast processing time−poor documentation

−requirement of sensor-specific hardwareProwler [27] Matlab x academic +accurate radio model

+fast and easy application prototyping−default TinyOs MAC protocol

−no support for sensors or physical phenomenaPtolemy [25] Java x open +support wireless communication and acoustic channel

+extensible and customisable−limited protocols (only sound)

TOSSIM [33] C x x BSDc +ideal transmission conditions+portability

−inappropriate for heterogeneous environments−restricted to TinyOS

−same code for each nodeShawn [34] C++ x BSD +large-scale WSNs

+high abstraction level−no GUI

UWSim [29] C++ x commercial +simulation of acoustic networks+novel routing protocol−limited functionalities

−only for underwater sensor networksaATMEL only.bParallel simulation environment for complex system.cBerkeley Software Distribution License.

IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

175

2.1.9 Sensim: Sensim [24] is an extension of OMNeT++ thatanalyses the networking and the scalability within the network. Itencapsulates an energy model that supervises the remaining energyin the battery and gives the different levels of power consumptionin different operation's modes. Besides, it integrates a simple radio,a simple CPU implementation and a network module combiningthe sensor and the network channel. However, it necessitates anabrupt learning curve which is not very popular with non-established simulators.

2.1.10 SensorSim: SensorSim [42] is an application and a packetlevel simulator. This extension of NS2 includes a sensor channel, aradio propagation model and a power model supporting all thehardware components that rely on the battery. Also, it allows adynamic and hybrid management combining both real andsimulated nodes. It simulates applications under actual conditions.The simulator is of high complexity due to hard coded applicationsand poor documentation support as it still is at pre-release stageand is now withdrawn.

2.1.11 STORM: STORM [43] is based on real-time schedulingsimulator that supports multiprocessor systems and is easy to use.Moreover, the ability to extend the features, to consider differentcomputation time task model, the characteristics and executionconditions of numerous hardware components, helps to improvecontinuously the results. Depending on the scheduling policy andthe resources described in an XML file, it runs every task over aspecified time interval. The results of the simulation are a set ofdiagrams analysing the behaviour of the system (tasks, processors,timing, performances etc.) and mainly the energy consumption.The behaviours of hardware components can be accuratelycaptured, which is the basis of energy consumption estimation.Each state is associated with a current consumption.

2.2 Comparison of WSNs power-aware simulators

The validation of sensors networks is difficult to achieve. Indeed,there is no precise modelling of energy consumption within thenode. Besides, rare are the models who consider the hardware.Some platforms only work with their own hardware architecturewhich limits their portability. Similarly, these simulators are notglobal and adopt either the hardware or the software or the networkgiven the difficulty of combining several aspects [24]. In addition,the simulation time is very long. We noted that essential functionsare missing in each simulator. Therefore, we tried through thissurvey to identify the best simulator who supports our needs. So,the simulator should adopt the relationship between functional,timing and energy behaviours. The combination of these threefactors ensures relevant abstractions of both functional behaviour

of the sensor nodes and the power consumption within them.Table 2 describes the key features of the power-aware simulatorsincluding their merits and limitations. We have noticed that themost used simulators such as NS2 or ATEMU focus on describingthe behaviour of the whole networks and fail to describe accuratelythe behaviour of a single node. However, PowerTOSSIM or AEONtries to estimate the energy at the level of instructions which is hardto achieve and not suitable to most of the nodes available in themarket. Few documentation are available for both EnergySim andPAWiS so it is difficult to use them. ATEMU, AEON andPowerTOSSIM provide a high behaviour and timing accuracy ofthe WSNs programs but do not cover new hardware blocks. Thus,they remain limited to some platforms which restrict their scope ofaction. ATEMU and EnergySim simulate heterogeneous networks.Unlike the PAWiS and SensorSim simulators, SENSE does notgive a detailed structure to capture energy consumption of differenthardware and software components. PAWiS, IDEA1 andSensorSim include the transition time between the different powermodes.

To sum-up an efficient simulator should be at the same timeextensible, accurate reusable, scalable, complete and have highfidelity characteristics. As a result of this comparative study, weselected the STORM simulator. Its multiple benefits meet our needto model the temporal execution of the tasks where each CPUemulates the operation of a real node. Also, it offers a realistic andeffective simulation and is capable of simulating several tens ofthousands of nodes. Besides, it allows an adequate level ofmodelling and a promising time of simulation with a simplicity ofmodelling and a relevance of the results.

3 Power consumption model for WSNWe propose a realistic power consumption model for WSN byincorporating the characteristics of a typical low-power processorsuch as the frequency, the voltage and time constraints. Fig. 2describes how the model will be developed. STORM allows from aspecification of the characteristics of a software architecture (thetasks to schedule), a hardware architecture (resources of execution)and the choice of a scheduling policy, to simulate the execution ofthese tasks on these resources according to the rules of this policyover a given period. The inputs are provided before the simulationthrough an XML file. Whereas the results are visualised in the GUIof the simulator (Gantt charts) or as pdf reports to debug theimplementations of the model and to show the behaviour of thenetwork. In this paper, we expose the potential of the energy modelwhich reduces the power consumption locally, at the level of thenode, and globally targeting the entire network. In fact, sometechniques, at the local level, require decisions at the global leveland vice versa. Fig. 3 depicts each phase of the development. Atthe global level, the ‘Task Allocation’ phase assigns the tasks to the

Table 2a Comparison of the power-aware WSNs simulatorsSimulator Language Advantages LimitationsAEON [35] Java power and lifetime estimation SW restriction: limited to TinyOS

execution of real code for accurate prediction of consumption HW restriction: no cover for new hardwareblocks

ATEMU [36] C accurate slow: long simulation timestep-by-step few functions to simulate routing

breakpoints feature to debug TinyOS application codes clustering problemEnergySim [37] C# custom build simulator does not follow a physical model

supports multiple topologies few documentationfast and easy user interface

IDEA1 [38] SystemC model network, SW and HW componentsPower modes and transition time support low data rate

heterogeneity supportJ-Sim [44] Java easy installation on different platforms low efficiency of simulation

specific packages with both battery and power model only IEEE 802.11 as MAC protocolPAWiS [40] C++/NED/Lua supports cross-layer design to capture the whole system in one

simulationlack of accuracy

extracts power consumption figures from SW and HW modules

176 IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

appropriate motes. The inputs of the model are the data related totime parameters, number of tasks, number of the nodes and so on.Then, the tasks are scheduled, in the ‘Task Scheduling’ phase,under the GEDF algorithm. At the local level, the ‘EnergyManagement’ phase aims to adapt the choice of the appropriateenergy strategy to the needs of the application. Finally, we evaluatethe energy consumption of each task in the ‘PerformanceEvaluation’ phase.

3.1 Impact of the power consumption model at the globallevel

A global energy manager can perform a variety of managementcontrol tasks based on the collected data such as controlling theresponse time, measuring the energy consumption and monitoringthe quality of service (QoS) by tuning the use of resources. Itallows to adjust the WSNs characteristics to increase the lifetime.Also, the GEDF strategy will schedule the n tasks at the globallevel.

3.2 Impact of the power consumption model at the local level

The adjustable parameters of the processor at a local level includethe voltage and frequency of the processor, which can be specifiedto meet the power budget. DPM strategy is to identify idle time tounder-utilise hardware components and adapt their powerrequirements accordingly. If there are no tasks in the queue, thisstrategy forces its subsystems to operate at the most economicalpower mode or to put them into a sleeping mode. Local DVFS canset the clock frequency and the voltage separately. Each task isstretched to its deadline while the voltage and the frequency arereduced.

Table 2b Simulator Language Advantages LimitationsPowerTOSSIM [28] nesC CPU cycle counting SW limitation: designed especially for tinyOS

track HW power state transition HW restriction: support only mica nodeseasy to use inaccuracies in the simulation

Sense [41] C++ balanced modelling methodology and simulationefficiency

lacks of comprehensive set of models andconfiguration templates

memory efficient simulation output hard to interpretfast and extensible

Sensim [24] Java investigate networking and scalability requires a steep learning curvefree space and two-ray ground reflection propagation

modelSensorSim [42] C++ flexible topologies synchronisation between the simulation scheduler

and protocol execution in the real nodesrealistic applications and middleware architectures

power models supportSTORM [43] Java energy characteristics of CPU (consumption, sleep mode,

time penalties)high level of abstraction

independence from the platform

Fig. 2  Overview of the execution of the power consumption model

Fig. 3  Overview of the WSN energy model

IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

177

4 Implementation of the power consumptionmodel with STORMThis energy model is a periodic real-time system and combines thefrequency settings and the power modes of the ATmega128(Table 3). It is implemented on the Mica2 platform and can beeasily ported to other sensor nodes. The tasks are preemptive,independent and support migration from one node to another. Thedeadline of the task is equal to its period. The values of the voltage/frequency couples supported by ATmega128 are derived from thedatasheet as shown in Table 5. Each system power mode representsa combination of power domains states and corresponds to powerrequirements of a specific software scenario.

A GEDF scheduling of n periodic, synchronous and implicitdeadlines tasks, on m identical processors exists if and only if thecondition

Utotal = ∑0

n CiTi

(1)

Umax = max CiTi

(2)

is respected (in our case m = 5; 2.58 < 5). Unfortunately, manystudies have shown that the available tests (1) for this policy do notallow full use of the processors while ensuring schedulability(Dhall effect [16]). This will lead us to apply other conditions toprove the schedulability of our GEDF; in particular, the limit onUmax which are those of Goossens, Funk and Baruah (GFB) [45]and Srinivasan and Baruah [46] clarified in what follows:

• The GFB condition states that the limit of the total utilisationratio is at most m ∗ (1 − umax) + umax such as umax is the maximum

utilisation ratio of any task. We found 2.58 which is largelylower than the limit of GFB which is 3.08.

• Srinivasan and Baruah showed that the utilisation rate of eachtask is at most (m/(2 ∗ m − 1)) can be scheduled successfully onm processors if the total use is at most (m2/(2 ∗ m − 1)). In ourcase we have Ui < 0.55 and Utotal = 2.5 < 3.125.

Indeed, we note that the parameters of the considered tasks(Table 4) show that it is feasible to implement a scheduling viaGEDF policy.

4.1 Implementation of the DPM in the WSN with STORM

Our DPM policy differs from the classical DPM policies byconsidering transition penalties. Besides, we will use the ‘time-out2’ strategies. The ‘time-out2’ selects the ‘DeepSleep’ statewhen the CPU is idle for a considering time duration. This allowsus to prevent the wake-up time from reaching a too high level. Thetime-out2-DPM is external to the scheduler and is identified in theXML input file. The power modes we consider are running, sleep,deep sleep, stand-by, extended stand-by and idle. First, the GEDFdesignates the task to run and the appropriate resource. Then, wecompute the time between two successive activations of the sametask and compare them with the sum of the transition durationbetween the two states, e.g. ttransition = 160 ms if this value exceedsttransition than the task is put in the ‘DeepSleep’ state. The transitiondelay tdelay of this transition has to be inferior to the latency toensure the proper functioning of the system. A description of theimplemented DPM algorithm used for WSNs and with STORM isgiven in Fig. 4.

4.2 Implementation of the DVFS in the WSN with STORM

Each nodes have several levels of frequency/voltage which aredenoted as (16;5.5), (14;5.05), (12;4.05) (10;3.15), (1;2.7). Weassume that the supply voltages from Vi to Vm such thatV1 ≤ V2 ≤ ⋯ ≤ Vm and the operating frequencies f i to f m such thatf 1 ≤ f 2 ≤ ⋯ ≤ f m are adjusted discretely. For each level, there is apower consumption associated: 48.9, 20.16, 13.3, 7.2, 0.8 mW.

The decision for a particular (Vi, Fi) is based on preserving thecorrect timing of the operating system. We will use the ‘inter-task’method where the processor speed and voltage are chosen beforereactivating the task. We have analysed the performance and

Table 3 Experimental parametersParameter Valueduration of the simulation 1000 msprecision 10−9snumber of tasks (n) 2, 5, 10, 20number of CPU (m) 2, 3, 5, 10CPU type ATMega 128runtime model WCET and AET

Table 4 Parameters of the software architecturePeriod Activation WCET BCET Deadline

1 80 0 50 30 802 100 0 55 35 1003 120 0 60 40 1204 150 0 60 30 1505 200 0 85 45 2006 250 0 120 60 250

Fig. 4  Algorithm 1: Implementation of DPM with STORM

178 IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

energy trade-offs using different DVFS factor α = ( f / f max) asshown in Table 5.

If the processor has been scaled with the factor α, a load of 1time units will take 1/α time units. When we decrease thefrequency, the effective execution time will be increased such thatall tasks still respect their response time bounds. If a task ispreempted or blocked, α = 1. The GEDF calculates each time theactual execution time (AET) and the worst execution cases time(WCET) of all tasks at the state ‘ready’ and ‘running’ to calculateα. Choosing a frequency lower than f min would necessarily lead tooverruns of deadlines. A detailed description of the DVFS that isimplemented is provided in the algorithm (see Fig. 5).

5 Evaluation of the performances of the powerconsumption modelWhen combining the power-aware model with a schedulingstrategy, the consumed energy is minimised without a degradationof the QoS. The simulation results are evaluated to analyse thepower consumption, the time parameters and to detect a potentialfor further optimisation. We observed that STORM provides resultsvery close to the values obtained on platform and applied in thesame experimental conditions. To measure the QoS, we calculatethe ratio of the number of tasks that meet their deadline and thetotal number of tasks. The average value is 97%.

The energy model allows task migration from one node toanother, which results in energy savings as well as workloadbalancing as illustrated in Fig. 6.

Fig. 7 highlights the simulation results for five nodes byconsidering the parameters of Table 4 (Utotal = 2.5).

We evaluated the impact of some parameters like the frequency,the transition time, the number of tasks etc. on final consumption.As a result, we noted that the more the number of processorsincreases the more the processing time and the makespan (i.e. thedate of completion of the last task scheduled) are reduced.However, it creates an additional energy costs by rising the slacktime. The simulation demonstrates also that reducing further thefrequencies leads to more energy saving at the global level.

The results of Fig. 8 reflect a gain between 26 and 52%. Thisfigure shows that there is a non-linear reduction of the energydissipation when the frequency is reduced. This is because changesin frequencies are accompanied by variation in voltage. Sincepower is the product of current and voltage, the increase in powerdissipation for different frequencies must be non-linear.

Fig. 9 captures the CPU load throughout the execution of tasksand the simulation time. The column time describes the CPU timein seconds. We see through this figure that the computational CPUload, defined as the ratio between the number of slots when theCPU is running and the total number of slots, drops significantly.When t = 0, the ratio is equal to 100%. The decrease that followedis due to the allocation of tasks to the different availableprocessors.

Table 6 compares the output of the proposed energy model thatcombines using global and local DVFS/DPM techniques. It isdifficult to model the idle interval when applying the DPM at theglobal level, and it causes additional energy and latency penaltymore than the local level. Besides, the DVFS strategy even if it iseasier to achieve than in the local level, it offers limited energyefficiency. Thus, this table supports our choice towards the use ofan interplay of both DPM and DVFS at the local level offeringmore energy gain and efficiency than in the global level.

6 DiscussionTable 7 illustrates a comparative study between our work andexisting techniques. The tasks [47] are based on the constraints ofprecedence where the end of a task corresponds to the beginning ofthe following one. This case generally imposes a cumulation ofdelay causing deadline violations. That's why we privileged thetasks with implicit deadline. Although Dargie [51] considered threepower modes, our approach extends to five power states byexploiting deep sleep with a Ttransition less or equal to the length ofthe inactivity interval. Liu et al. [48] proposed an ‘Energy Aware

DVFS’ algorithm (EA-DVFS) that slows the tasks if the systemhas not enough available energy, otherwise, the tasks are executedat full speed.

In [49], using exclusively two couples of (V, F) leads to designerrors difficult to determine. Indeed, they cause an inconsistencybetween the decisions taken by the energy manager and the actualfunctional states produced during the execution. This problem wassolved in our work by the consideration of discrete frequencies andreciprocally the voltage. The interaction between the DPM in theidle state and the DVFS in the active state offers more gain incomparison with other techniques. The advantage of our PM is itsability to switch from one methodology to another by switchingbetween DPM, DVFS and GEDF resulting in better energysavings. It has been deployed as an intermediate layer at theapplication and OS level. On the energy-saving level we obtainedan estimated gain between 50 and 90%.

7 ConclusionEnergy conservation has played a large part in the technologicalevolution of sensor nodes as the available energy resources arelimited. In order to achieve high efficiency and establish optimisedenergy consumption in a node, we proposed a power efficientsystem for WSNs. This power-aware model was designed toimprove the energy efficiency thanks to a double solution: a globaland dynamic approach using the analysis of the behaviour of thenetwork by applying a global EDF scheduler and a local approachbased on the application requests and the available quantity ofenergy through the time-out2-DPM and an inter-task DVFS. Toensure a better schedulability we combined the conditions of GFBand Srinivasan and Baruah. This model is generic and can workwith different scheduling algorithms. As a deep optimisation ofpower consumption requires a simulation tool to profile the energycost of the internal work of each node, we developed in this paper acomprehensive evaluation of the important simulationsenvironments in order to select the most suitable to our needs. Wehave classified them according to the energy criterion. Finally, thesimulation results were given to illustrate the effectiveness of theproposed approach. We have also extended STORM features byimproving its performances and functionalities.

8 AcknowledgmentsThe author thanks Prof. Aimé Lay-Ekuakille, Dipartimentod'Ingegneria dell'innovazione, University of Salento and Prof.Faouzi Derbel, HTWK Leipzig, Fakultät Elektrotechnik undInformationstechnik, for their valuable suggestions and constantsupport throughout this research work.

9 References[1] Singh, S., Chand, S., Kumar, R., et al.: ‘Neecp: Novel energy-efficient

clustering protocol for prolonging lifetime of wsns’, IET Wireless SensorSystems, 2016, 6, (5), pp. 151–157

[2] Benhaddou, D., Al-Fuqaha, A.: ‘Wireless sensor and mobile ad-hoc networks:vehicular and space applications’ (Springer, 2015)

[3] Khriji, S., El Houssaini, D., Jmal, M.W., et al.: ‘Precision irrigation based onwireless sensor network’, IET Sci. Meas. Technol., 2014, 8, (3), pp. 98–106

[4] Ojuroye, O., Torah, R., Beeby, S., et al.: ‘Smart textiles for smart homecontrol and enriching future wireless sensor network data’ (Springer, 2017),22, pp. 159–183

[5] Varghese, B., John, N.E., Sreelal, S., et al.: ‘Design and development of anRF energy harvesting wireless sensor node (EH-WSN) for aerospaceapplications’, Procedia Comput. Sci., 2016, 93, pp. 230–237

[6] Chéour, R., Jmal, M., Abid, M.: ‘Hybrid energy-efficient power managementfor wireless sensors networks’. Int. Conf. on Smart, Monitored and ControlledCities (IEEE, Tunisia, 2017)

[7] Guru, R.: ‘Energy efficiency mechanisms in wireless sensor networks: asurvey’, Int. J. Comput. Appl., 2016, 139, (14), pp. 27–33

[8] Shaikh, F.K., Zeadally, S.: ‘Energy harvesting in wireless sensor networks: acomprehensive review’, Renew. Sustain. Energy Rev., 2016, 55, pp. 1041–1054

[9] Bartariya, S., Rastogi, A.: ‘Security in wireless sensor networks: Attacks andsolutions’, Environment, 2016, 5, (3), pp. 214–220

[10] Obeid, A.M., Karray, F., Jmal, M.W., et al.: ‘Towards realisation of wirelesssensor network-based water pipeline monitoring systems: a comprehensivereview of techniques and platforms’, IET Sci. Meas. Technol., 2016, 10, (5),pp. 420–426

IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

179

[11] Rawat, P., Singh, K., Chaouchi, H., et al.: ‘Wireless sensor networks: a surveyon recent developments and potential synergies’, J. Supercomput., 2014, 68,(1), pp. 1–48

[12] Lamonaca, F., Gasparri, A., Garone, E., et al.: ‘Clock synchronization inwireless sensor network with selective convergence rate for event drivenmeasurement applications’, IEEE Trans. Instrum. Meas., 2014, 63, (9), pp.2279–2287

[13] Sankar Ramachandran, G., Daniels, W., Matthys, N., et al.: ‘Measuring andmodeling the energy cost of reconfiguration in sensor networks’, IEEE Sens.J., 2015, 15, (6), pp. 3381–3389

[14] Harkut, D., Ali, M., Lohiya, P.: ‘Scheduling task of wireless sensor networkusing earliest deadline first algorithm’, Int. J. Sci. Res. Comput. Sci. Eng.,2014, 2, (2), pp. 1–6

[15] Kulau, U., Rottmann, S., Schildt, S., et al.: ‘Undervolting in real world WSNapplications: a long-term study’. Int. Conf. on Distributed Computing inSensor Systems (DCOSS), 2016, 2016, pp. 9–16

[16] Dhall, S.K., Liu, C.L.: ‘On a real-time scheduling problem’, Oper. Res., 1978,26, (1), pp. 127–140

[17] Helkey, J., Holder, L., Shirazi, B.: ‘Comparison of simulators for assessingthe ability to sustain wireless sensor networks using dynamic networkreconfiguration’, Sust. Comput. Inf. Syst., 2016, 9, pp. 1–7

[18] Lattanzi, E., Freschi, V., Dromedari, M., et al.: ‘A fast and accurate energysource emulator for wireless sensor networks’, EURASIP J. Embed. Syst.,2017, 2016, (1), p. 18

[19] Titzer, B.L., Lee, D.K., Palsberg, J.: ‘Avrora: scalable sensor networksimulation with precise timing’. Fourth Int. Symp. on Information Processingin Sensor Networks, 2005, pp. 477–482

[20] Osterlind, F., Dunkels, A., Eriksson, J., et al.: ‘Cross-level sensor networksimulation with COOJA’. 31st IEEE Conf. on Local Computer Networks,2006, pp. 641–648

Table 5 Voltage/frequency couples supported byATmega128level 1 2 3 4 5frequency, MHz 16 14 12 10 1voltage, V 5.5 4.05 3.6 3.15 2.7α 1 0.875 0.75 0.625 0.0625energy, mJ 48.9 20.6 13.3 7.2 0.8

Fig. 5  Algorithm 2: Implementation of the DVFS in the WSN with STORM

Fig. 6  EDF-DPM experiment results for a node A

Fig. 7  Simulation results for n = 5 motes

Fig. 8  EDF-DVFS experiment results

180 IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

[21] Girod, L., Elson, J., Cerpa, A., et al.: ‘Emstar: a software environment fordeveloping and deploying wireless sensor networks’. Proc. of the GeneralTrack: USENIX Annual Technical Conf., 2004, pp. 283–296

[22] Chhimwal, M.P., Rai, D.S., Rawat, D.: ‘Comparison between differentwireless sensor simulation tools’, IOSR J. Electron. Commun. Eng., 2013, 5,(2), pp. 54–60

[23] Issariyakul, T., Hossain, E.: ‘An introduction to network simulator NS2’(Springer, 2012)

[24] Chéour, R., Jmal, M., Lay-Ekuakille, A., et al.: ‘Choice of efficient simulatortool for wireless sensor networks’. Int. Workshop on Measurements andNetworking Proc. (M&N), Naples, Italy, 2013

[25] Liu, J., Liu, X., Lee, E.A.: ‘Modeling distributed hybrid systems in PtolemyII’. Proc. of the American Control Conf., 2001, vol. 6, pp. 4984–4985

[26] Nayyar, A., Singh, R.: ‘A comprehensive review of simulation tools forwireless sensor networks (WSNs)’, J. Wirel. Netw. Commun., 2015, 5, (1), pp.19–47

[27] Korkalainen, M., Sallinen, M., Karkkainen, N., et al.: ‘Survey of wirelesssensor networks simulation tools for demanding applications’. Fifth Int. Conf.on Networking and Services, 2009, pp. 102–106

[28] Shnayder, V., Hempstead, M., Chen, B., et al.: ‘Simulating the powerconsumption of large-scale sensor network applications’. Proc. 2nd Int. Conf.on Embedded Networked Sensor Systems, 2004, pp. 188–200

[29] Dhurandher, S.K., Misra, S., Obaidat, M.S., et al.: ‘Uwsim: a simulator forunderwater sensor networks’, Simulation, 2008, 84, (7), pp. 327–338

[30] Boulis, A.: ‘Castalia, a simulator for wireless sensor networks and body areanetworks, version 2.2’, User's manual (NICTA, 2009)

[31] Varga, A., et al.: ‘The OMNET++ discrete event simulation system’. Proc.European Simulation Multiconf. (ESM2001), 2001

[32] Hammoodi, I., Stewart, B., Kocian, A., et al.: ‘A comprehensive performancestudy of Opnet modeler for Zigbee wireless sensor networks’. Third Int. Conf.Next Generation Mobile Applications, Services and Technologies, 2009, pp.357–362

[33] Levis, P., Lee, N., Welsh, M., et al.: ‘Tossim: accurate and scalable simulationof entire Tinyos applications’. Proc. First Int. Conf. on Embedded NetworkedSensor Systems, 2003, pp. 126–137

[34] Kroeller, A., Pfisterer, D., Buschmann, C., et al.: ‘Shawn: a new approach tosimulating wireless sensor networks’, arXiv preprint cs/0502003, 2005

[35] Landsiedel, O., Wehrle, K., Götz, S.: ‘Accurate prediction of powerconsumption in sensor networks’. Proc. Second Workshop on EmbeddedNetworked Sensors, 2005

[36] Polley, J., Blazakis, D., McGee, J., et al.: ‘Atemu: a fine-grained sensornetwork simulator’. First Annual IEEE Communications Society Conf. onSensor and Ad Hoc Communications and Networks, 2004, pp. 145–152

[37] Nasir, A., Soong, B.-H.: ‘Energysim a novel, fast, extensible wireless sensornetwork mac protocol simulator for evaluating energy efficiency’. Eighth Int.Conf. Information, Communications and Signal Processing (ICICS), 2011, pp.1–5

[38] Du, W., Mieyeville, F., Navarro, D., et al.: ‘Idea1: a validated SystemC-basedsystem-level design and simulation environment for wireless sensornetworks’, EURASIP J. Wirel. Commun. Netw., 2011, (1), pp. 1–20

[39] Fahmy, H.M.A.: ‘Simulators and emulators for WSNs’, ‘Wireless sensornetworks’ (Springer, 2016), pp. 381–491

[40] Weber, D., Glaser, J., Mahlknecht, S.: ‘Discrete event simulation frameworkfor power aware wireless sensor networks’. Fifth Int. Conf. on IndustrialInformatics, 2007, vol. 1, pp. 335–340

[41] Chen, G., Branch, J., Pflug, M., et al.: ‘Sense: a wireless sensor networksimulator’, ‘Advances in pervasive computing and networking’ (Springer,2005), pp. 249–267

[42] Park, S., Savvides, A., Srivastava, M.B.: ‘Sensorsim: a simulation frameworkfor sensor networks’. Proc. 3rd ACM Int. Workshop on Modeling, Analysisand Simulation of Wireless and Mobile Systems, 2000, pp. 104–111

[43] IRCCYN: ‘Présentation de storm’, 2015, http://storm.rts-software.org/doku.php?id=start=

[44] Sobeih, A., Hou, J.C., Kung, L.-C., et al.: ‘J-sim: a simulation and emulationenvironment for wireless sensor networks’, Wirel. Commun., 2006, 13, (4),pp. 104–119

[45] Goossens, J.: ‘Scheduling of offset free systems’, Real-Time Syst., 2003, 24,(2), pp. 239–258, Available at: http://dx.doi.org/10.1023/A:1021782503695

[46] Srinivasan, A., Baruah, S.: ‘Deadline-based scheduling of periodic tasksystems on multiprocessors’, Inf. Process. Lett., 2002, 84, (2), pp. 93–98

[47] Chen, G., Huang, K., Knoll, A.: ‘Energy optimization for real-timemultiprocessor system-on-chip with optimal DVFS and DPM combination’,ACM Trans. Embed. Comput. Syst. (TECS), 2014, 13, (3s), p. 111

[48] Liu, S., Qiu, Q., Wu, Q.: ‘Energy aware dynamic voltage and frequencyselection for real-time systems with energy harvesting’, ‘Design, automationand test in Europe’ (IEEE, 2008), pp. 236–241

[49] Hoang, V.T., Julien, N., Berruet, P.: ‘Increasing the autonomy of wirelesssensor node by effective use of both DPM and DVFS methods’, ‘Faibletension faible consommation (FTFC)’ (IEEE, 2013), pp. 1–4

[50] Popovici, E., Magno, M., Marinkovic, S.: ‘Power management techniques forwireless sensor networks: a review’. Fifth IEEE Int. Workshop on Advancesin Sensors and Interfaces (IWASI), 2013, pp. 194–198

[51] Dargie, W.: ‘Dynamic power management in wireless sensor networks: state-of-the-art’, IEEE Sens. J., 2012, 12, (5), pp. 1518–1528

[52] Brihi, A.: ‘Investigation into the dependency between resource utilization,power consumption and performance in multimedia servers’. PhD thesis, MScdissertation, Department of Computer Science, Technische UniversitatDresden, 2012

[53] Kansal, A., Hsu, J., Zahedi, S., et al.: ‘Power management in energyharvesting sensor networks’, ACM Trans. Embed. Comput. Syst. (TECS),2007, 6, (4), p. 32

[54] Zhu, N.: ‘Simulation and optimization of energy consumption in wirelesssensor networks’. PhD thesis, Ecole centrale de Lyon, Ecully, 2013

Fig. 9  Impact of CPU load

Table 6 Comparison of global and local DVFS/DPM outputGlobal Local

DPM define a wide sleep state provide the necessary energyfor each task

limitations: mode switching in CPU causes additional energyand latency penalty

advantages: easier to achieve advantages: scalabilitylimitations: difficult to model the

idle intervallimitations: transition delay

DVFS a single global (F, V) to thewhole network

assign to each component'sits (F, V) separately

limitations: lack of means to reduce static power consumptionadvantages: frequent, easy,

cheapadvantages: flexible, more

efficientlimitations: limited energy

efficiencylimitations: complexity

Table 7 Comparative study of existing techniques with the energy manager EM developedAlgorithms Tasks Online offline Harvesting Scheduler Transition

delaysAffinity Overhead

Chen et al. [47] DPM-DVFS periodic, dependent x — time triggered x xLiu et al. [48] EA-DVFS periodic, preemptive x x EDF no negligibleHoang et al. [49] DPM-DVFS periodic, preemptive x no noPopovici et al. [50] DVFS — x x — — noDargie [51] DPM-DVFS periodic x FIFO — — noBrihi [52] DVFS periodic noKansal et al. [53] KAN-PM periodic x x — noZhu [54] iMASQO sporadic x x x no noLe [55] BQS-PMa periodic x x — x no yes

EM DPM-DVFS-GEDF periodic, preemptive,independent

x GEDF x yes no

aQoS-based Power Management.

IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017

181

[55] Le, T.N.: ‘Global power management system for self-powered autonomouswireless sensor node’. PhD thesis, Université Rennes 1, 2014

182 IET Comput. Digit. Tech., 2017, Vol. 11 Iss. 5, pp. 173-182© The Institution of Engineering and Technology 2017