Routing Heuristics in LTE-Advanced Networks

Bruno Alexandre Moreira Pinchobruno.pincho@ist.utl.pt

Instituto Superior Técnico, Lisboa, Portugal

June 2018

Abstract

Device-to-Device communications are recognized as a viable and resourceful approach to deal withcellular network capacity limitations. Described in most recent literature regarding mobile cellulararchitecture, these can provide an alternative traffic flow that mobile service providers can implementusing their own subscribers equipment in order to alleviate network congestion, increase availability andreduce infrastructural costs. In the presented work through the gathering and analysis of the availablestate of the art information, a theoretical model for the creation of a Device-to-Device routing protocolwith load balancing in shared spectrum for Long Term Evolution Advanced cellular network, will bepurposed. The proposed scheme was evaluated through simulation. The results show that betterperformance can be achieved for users of high data rate apps in congested networks, in comparison toa scheme in which only the shortest routes were used.Keywords: Device-To-Device, LTE-A, Routing, Genetic Algorithm.

1. Introduction1.1. Motivation

Wireless networks are predicted to handle InternetProtocol(IP) data in the order of 190 exabytes by2018, and may exceed 500 exabytes by 2020 [1].Besides the increase in connected devices, individ-ual traffic demand also increases, mainly due to thegrowth of bandwidth-hungry applications such asvideo streaming and file sharing [2].This evolutionin customer demand shows that existing cellularnetworks will soon fail to meet market expectations,either due to inflexible and expensive equipment ordue to a complex and non-agile control plane [3].Along the path from Long Term Evolution(LTE)networks towards the 5G, one of the most relevantadvances is presented in 3GPP release 12 in whichDevice to Device(D2D) communications are speci-fied for the first time [4]. The latter are expectedto improve spectrum usage and energy efficiency ofthe overall network system, increasing throughputand reducing end-to-end delay[5] .

Mobile service providers will have to be able toprovide the same or better service in order to main-tain revenue, this is not an easy task since, as userdemands increase, so does the user lack of toler-ance facing under performing services and applica-tions. The necessary upgrades to the network struc-ture will be expensive for providers, but the liter-ature states that a cost reduction can be achievedby implementing Cloudification and Programmabil-ity. The first consists in the virtualization of net-

work functions consequently eliminating the needfor specific hardware, while the second will providethe network with enough flexibility to incorporateservice differentiated innovations [3] . One couldargue that hybrid cellular/D2D networks wouldntbring any monetary advantage to service providers,but [2] show that a two-tier network, with parallelD2D and cellular traffic provide a better revenue forthe operator as the number of devices in the samenetwork increases, when in comparison to a con-ventional network. In currently deployed networks,mobile service providers already make use of trafficoffloading techniques such as using WiFi hotspotsto extend service availability for subscribed cos-tumers.

1.2. Objectives

The purpose of this work is to develop and evalu-ate a multiple variable routing protocol to assist thecreation of best-effort communication links betweencommunicating nodes in Long Term EvolutionAdvanced(LTE-A) cellular networks. Three deci-sion variables will be taken into account, namelyroute length, proximity to interfering User Equip-ment(UE) and energy consumption.

The proposed routing system will be simulatedusing the SimuLTE framework, an open source sys-tem level simulator for LTE and LTE-A, which isbased on OMNeT++ a discrete event simulator.

1

1.3. Related Works

This section will summarize some of the most rel-evant scientific work encountered when researchingthe addressed subject. In [6] Muhammad Usmanet al. present a software defined network controllerto couple the formation of D2D clusters with cen-tralized control signaling, resource allocation androuting. The clustering is promising in terms ofscalability by reducing the search space for pathcomputation by a factor proportional to the maxi-mum cluster size.

In [7] Hiroki Nishiyama et al. state the impor-tance regarding D2D communications in disasterscenarios, developing an application that enablesD2D multihop communication, through the estab-lishment of a MANET using WiFi direct technol-ogy.

[8] Giovanni Nardini et al., present a resourceallocation framework based on a best-fit heuristicthat helps the network decide which transmissionmode should a communication flow be assigned to.Always aiming to maximize throughput and min-imize losses due to mode switching. The statedwork focuses in one-to-one D2D communications,and only assumes pairs of devices which are in aminimum range from each other as eligible D2Dlinks. Therefore every direct communication is onlymade between pairs of devices, no kind of rout-ing computation is done by the framework, onlyresource allocation.

In [9], is presented a solution to enable mobility inD2D links from one communication cell to another,which yielded promising results, by succeeding inreducing network signaling overhead and latency.The proposed model provides improvements in ser-vice reliability specially when considering vehicle tovehicle communications, which have a higher han-dover ratio.

Jaesung Park in [10] proposes a fast and energyefficient D2D multihop routing framework, imple-mented separated from the infrastructure network.The framework uses a presence server to account fordevice location, in order to perform a location basedrouting by calculating the Euclidean distances be-tween each device and choosing the closest and leasttraffic congested nodes. Performance was evaluatedin a home-made event simulator, in which infras-tructure influence was not taken in account.

With increasing complexity, Dr.Ketan Kotechaand Sonal Popat purpose in [11] a Multi-objective Genetic Algorithm (MOGA) based adap-tive routing algorithm in Mobile Ad hoc Net-works(MANET). Experimental results obtainedfrom the Network Simulator NS-2.28, show that theMOGA based protocol approach is better in termsof delay reduction, when compared to Ad hoc on de-mand Distance Vector(AODV) and Quality of Ser-

vice(QoS) based AODV routing.

2. BackgroundThis section will summarized the most importantinformation, regarding the developed work.

2.1. Long-Term Evolution NetworkLTE is one of the most recent wireless telecommuni-cations technology, following the Universal MobileTelecommunications System(UMTS) base architec-ture and improving it further. The designed pur-pose was to provide a high-speed, universally ac-cessible wireless service, enabling networking at alllocations for tablets, smartphones, computers andin general LTE enabled User Equipments. In com-parison with older architectures LTE is an All-IPpacket switched network, with average data ratesof 200 Mbps. Since LTE has no circuit-switching,voice services are provided over IP, defining a newstandard, voice over LTE.

LTE uses subcarriers 15 KHz apart, organizingthe radio interfaces Downlink(DL) and Uplink(UL),into radio frames of 10ms duration. Physical chan-nels can either be implemented using Time divi-sion Duplexing(TDD) or Frequency Division Du-plexing(FDD). In the first, UL and DL share thesame frequency band alternating only in the timedomain,while in FDD each channel has its own fre-quency bands.

Physical Resources are organized in a time-frequency grid, in which each column is 6/7 OFDMsymbols per slot and each row corresponds to a15KHz subcarrier. A Resource Block(RB) consistsin 12 subcarriers spread across 6/7 OFDM symbols.Considered the smallest capacity unit that can beassigned to a user, the Evolved Node B(eNB) as-signs the RBs with channel-dependent scheduling.

2.1.1 Power characteristics of LTE net-works

In order to be able to measure the impact of for-warded network traffic in terms of energy comsump-tion in each envolved UE in a LTE network, it is firstnecessary to choose the most correct possible modelin order to be able to simulate with acceptable pre-cision the energy consumption related to forwardingtransmission in the UL. In [12] the authors purposea empiric power model for data transfer in LTEnetworks,which accounts for separate UL and DLcontributions. For simultaneous UL and DL datatransfer, the power level (mW) is given by the fol-lowing formula [12]:

P = αutu + αdtd + β (1)

In which the terms αu = 438.39 and αd = 51.97(mW/Mbps) are empirically determined constants.The constant β = 1288.04 (mW) is the base power

2

when throughput is 0. The UL and DL throughputare respectively given by tu and td (Mbps).

2.2. LTE-AdvancedLTE-A architecture was specified in 3GPP release10, being further improved in following releases 11and 12. The key improvements over LTE, can besynthesized as follows:

• Carrier aggregation.

• Multiple Input Multiple Output(MIMO) en-hancements to support higher dimensionalMIMO.

• Relay Nodes.

• Heterogeneous networks involving small cellssuch as femtocells, picocells and relay.

• Traffic offload techniques to redirect trafficonto non-LTE networks.

• Enhancements to machine-to-machine comuni-cations.

2.2.1 Downlink Vs Uplink

Commercial handheld devices like cellphones andtablets do not have access to a dedicated spectrum.These must use the mobile service providers estab-lished radio resources. In case of D2D UEs, thesedevices will have to share the available spectrumwith normal cellular users in paired FDD duplex-ing or unpaired TDD duplexing networks.

D2D transmissions may use resources from thetwo available air interfaces,either UL or DL:

• UL interface: since the UL is modulated us-ing Single Carrier Frequency Division MultipleAccess (SC-FDMA) the UE devices will needto be equipped with a SC-FDMA receiver inorder to use this interface to communicate. Us-ing UL resources for D2D transmission meansthat the UE will experience strong co-channelinterference from nearby cellular devices thatare generating uplink traffic.

• DL interface: as the modulation used for theDL is Orthogonal Frequency Division MultipleAccess(OFDMA), D2D devices would have tobe equipped with OFDMA transmitters. Simi-larly to UL constraints, using DL resources forD2D communications means that the D2D UEsmay cause a high level of interference to nearbyco-channel cellular devices that are receivingDL traffic.

Choosing the right interface is imperative in or-der not to hurt the overall network performance.The literature advises to share uplink resources

with D2D. Since UL resources have less utiliza-tion and the SC-FDMA modulation scheme has alower peak to average power ratio than OFDMA,this approach improves spectral efficiency and al-lows a lower energy consumption which is advanta-geous in power constraint mobile devices. Concern-ing the DL, the presence of heavy control signal-ing and higher traffic congestion associated with amore complex hardware implementation, reinforcethe use of UL resources[5].

2.3. LTE Simulation engineThis section will address the chosen simulationframework and its underlying engine.

2.3.1 OMNeT++

OMNeT is an open-source discrete-event simulatorthat features extensible, modular, component-basedC++ simulation library and framework,which areused primarily for network simulations [13].

The basic building blocks for OMNeT++ arethe modules. These can be either simple or com-pound. Module communication is made throughmessages, which are sent and received through aconnection linking each module’s gates. The gate isan input/output communication interface presentin each simple module. A network is considereda compound module without any gates, it has thehighest modular hierarchy. In compound modules,connections must be declared in order to allow sub-modules to communicate with others outside thecompound. Connections are characterized by abit rate, delay and loss rate. When sub-modulesinside a compound module need to communicatewith a module outside of the network, their con-nections must go through the compound modulegate. OMNeT++ simulation modules have them-selves a modular architecture, divided in implemen-tation, description and value parameterization. Thebehavior or implementation is written in C++ inthe form of classes, the description is expressed inNetwork Description files, and the model values areinitialized in a specific .INI file.

2.3.2 SimuLTE

SimuLTE is a system-level simulator based on theOMNeT++ and INET frameworks, which uses thesame modular concept. It is designed as an ex-tension of a wireless network interface card moduleand allows the simulation of LTE/LTE-A in FDD.The framework provides models for UEs and eNBs,in which both contain the LTE Network InterfaceCard(NIC), along with other modules that imple-ment higher layer protocols taken from the INETframework. The LTE NIC implements the wholeprotocol stack, one sub-module per layer. In the

3

UE, the NIC module connects to the IP module,taken from the INET framework, allowing the useof TCP and UDP applications. In the eNB mod-ular structure it allows the connection of multi-ple eNBs through IP, via a Point-to-Point Proto-col(PPP). The binder can be designated as whatthe authors refer as a oracle module in [14]. It hasfull visibility of the nodes in the network, assistingin interference computation by keeping track of re-source allocations. The EPC is constituted by thepacket data gateway, which interfaces the eNB withexternal IP based networks. A representation canbe seen in Figure 1.

Figure 1: SimuLTE LTE-A module representationfrom [14].

2.4. Genetic AlgorithmsBy taking inspiration from Charles Darwin’s Natu-ral Selection theory,John Henry Holland pioneereda new area in computational science by establish-ing the field of evolutionary computation. Thebase conception was to mathematically abstract thecomplex and robust procedures that allowed bio-logical organisms to evolve and adapt to their en-vironment, and thus reach success. Holland pro-posed the idea of mapping information regardingcandidate solutions of a given problem in strings ofbinary symbols, referring to them as chromosomes,also proposing computational analogs of the cellularrandom mutation, crossover and natural selection,in order to guide the evolution of the given chro-mosomes towards a solution of the proposed math-ematical problem [15].

A Genetic Algorithm(GA) can be categorized inmultiple types as a consequence of their versatility.The classification is based on the number of objec-tives the algorithm has to deal with, being the mostcommon types the single objective, multi-objectiveand parallel GA.

2.4.1 Encoding

The data mapping or encoding into chromosomesis one of the first steps taken when implementing aGA. Heavily reliant on the scope of the problem, the

type of encoding that is chosen will greatly impactthe performance of the algorithm and its readabil-ity. The most common encoding schemes can beseen in Table 1.

Table 1: Encoding Schemes.Encoding Type ChromosomeBinary 1000011101Value ABEFGTEPermutation 1 2 5 4 3 1 7Hexadecimal 1 4 9 A 4 COctal 0 1 2 7 3 5

2.4.2 Fitness Function

The Fitness Function is a mathematical expressionthat will emulate the effect that the environmentwould have on an adapting and evolving organism.In essence it is just an objective function whose pro-pose, regarding problem solving, is to maximize orminimize a given numerical value, indicating howmuch each variable in the problem is going to con-tribute to the final and optimal solution.

More specific to the implementation in GAs anobjective function that translates the fitness of eachorganism in a population pool must be devised ina way that the chromosome itself acts as a vari-able input , where fitness is assigned according tothe quality of the encoded proposed solution andits respect for the established constraints. Thisensures that the algorithm only spends processingtime in promising solutions, by guiding the geneticoperands such as crossover and selection. In orderto guarantee deterministic results , fitness functionsare required to have a precise model which is feasi-ble to compute [15].

Fitness assignment can be separated in thetwo main categories, single objective and multi-objective. In single objective, the algorithm onlyneeds to satisfy one condition (in a network it couldbe maximize throughput,minimize latency,etc...).In Multi-Objective the fitness function will need toweight and assign fitness values taking in regardmultiple constraints in order to reach a solution thatis optimal according to the various objectives statedby the problem.

2.4.3 Crossover

The genetic operator crossover, is responsible forthe creation of new chromosomes that will populatea new generation. The recombination of the dataencoded in two or more chromosomes will allow theGA to test new candidate solutions that may pro-vide an optimal one or help the next generation toconverge towards it.

4

2.4.4 Mutation

The mutation operator is used for exploring differ-ent options in the problem’s solution space. Muta-tion provides a significant help in the diversificationof the population, having a great impact in avoidingthe local optima problem. Several types of muta-tion techniques exist and some are limited by thegene encoding [15].

2.4.5 Selection

The Selection Operator allows the GA to choose ateach iteration, or generation the individuals whosechromosomes have the highest quality from theavailable genetic pool. By selecting some individ-uals in favor of others, the selection operator isexplicitly implementing Elitism. Elitism in GAsmaintains efficiency for it prevents the prolifera-tion of unwanted or weak solutions, guaranteeingthe provided computational resources are used onlyfor the most promising individuals.

Selection relies heavily on fitness as a measure tocompare two or more different individuals and selectthe fittest. The Natural consequence of selection ispersistence: the considered fittest individuals willbe subjected to the crossover and mutation opera-tors, having a fraction of their chromosome stringtransmitted to the offspring of a given generation,as described in the crossover and mutation subsec-tion.

2.4.6 Multi-Objective optimization

In cases where the described problem holds multi-ple variables and it is neither clear the weight ofeach variable in the objective function or they sim-ply cannot be assigned relative weights beforehand,then multi-objective GAs are a popular approachaccording to Jones et al. [16], with particular us-age in theoretical and electrical engineering fields.Mainly due to the fact this class of algorithms donot require the user to prioritize, scale or weightobjectives.

if we assume that the multiple-objective problemoptimization can be represented as the minimiza-tion of a fitness vector then:

fff(xxx) = [f1(xxx), f2(xxx), ..., fN (xxx)] (2)

In which N represents the number of objectivefunctions fi : Rn 7→ R and each solution vector xxx =[x1, x2, ..., xn] is composed of n decision variables.

2.4.7 NSGA-II

MOGAs that used the non-dominated sorting tech-nique had been criticized for their computationalcomplexity of O(MN3) (in which M stands for the

number of decision variables and N the populationsize), for a non-elitist approach (that would com-promise convergence rates) and the need to specifya sharing parameter. All according to KalayanmoyDeb et al. [17], who proposed an improvement overthe previous NSGA in an attempt to solve the abovecited problems. Despite the novel algorithm notproving itself the dominant MOGA in all scenarios,it performed very well when solving problems hav-ing strong parameter interactions and convergingbetter than its competitors, as shown in the exper-imental results in [17].

The NSGA-II faster convergence rates are of spe-cial importance, since in a real life scenario, a quickresponse from the routing algorithm allows for asmaller delay in communications. In the statedproblem, it is not obvious which of the optimizationvariables will weight more heavily in the quality ofthe communication flows, as such the rank basedapproach of the NSGA-II will allow to dynamicallydetermine the importance of each variable.

3. ImplementationDuring this dissertation, an NSGA-II based rout-ing algorithm was developed, for use in LTE-A net-works. Additionally three applications and two newmodules where developed for the SimuLTE frame-work, refactoring version. The first two applica-tions implement a simple client-server UDP basedsystem, in order to simulate simple interactions be-tween the client’s UE and the routing server. Thedeveloped MOGA and the shortest path algorithmare executed in the routing server. The third de-veloped application was a Video on Demand(VoD)application. Regarding the module that representsthe UE, an extension to the base model was createdin which were added two new sub-modules, one be-ing a simplified routing table and the other a energyconsumption meter. The following figure shows howthe different sub-modules of the extended UE inter-act between each other.

Figure 2: Sub-module interaction within the ex-tended UE module in SimuLTE architecture.

3.1. Routing ApplicationThe routing system was implemented as a client-server architecture, using UDP as the underlyingtransport protocol. Divided in two applications, the

5

client side runs in the UEs, while its counterpartis executed in a server located in an external IPnetwork. The external network interfaces with theeNB through the packet data gateway.

Communication between the two components ofthe system follows a simple request-reply protocol,in which the client must request the server in or-der to have a route assembled. The server may ei-ther reply with an error message in case of failurein building the requested route, or with an updatemessage, which will be sent to the requesting UEand to all the UEs included in the computed route.

The client routing app has to perform two maintasks. The first is to update the simplified routingtable module, allowing other applications runningin the same UE to use the established routes. Thesecond task is to relay UDP traffic, in the imple-mented module the client application can only re-lay traffic from the Voice over IP(VoIP) and VoDapplications.

The server routing app requests the network forinformation regarding the deployed UEs and regis-ters geographical information associated with each.When a request message arrives, depending on theinitialization value the server will perform a spe-cific route building routine. The implemented pathmaking modules are the developed MORGA andDijkstra’s shortest path algorithm.

3.2. VoD and VoIP ApplicationsDue to technical limitations regarding both applica-tions,their respective message packet structure hadto be altered by adding an extra address field re-garding the concrete destination of the transmittingdata, which is decided previously to the simulation.This modification does not alter either the payload,or allows the routing application to read the contentin the relayed traffic.

Since the VoD application was not originally im-plemented within the SimuLTE framework, a newimplementation was made regarding the multi-hopscope.

3.3. User Equipment Model ExtensionSome of the features of the previously existingmodel that represented the UE, were either insuffi-cient or not fully adequate. For example, the exist-ing routing table was largely dependent on the met-rics used in IP based fixed networks, being overlycomplicated for the desired purpose. Also the lackof a battery like module, made it impossible to an-alyze the impact a given route would have on theequipment’s energy consumption.

In order to add new features, an extension ofthe existing UE module was created, in which thenew sub-modules were added. The first sub-moduleconsisted in a simplified version of the pre-existingrouting table, that could be accessed by any app

in the UE. The second sub-module is essentially anenergy consumption meter, which implements themodel discussed in Section 2.1.1.

3.4. Implemented Genetic AlgorithmThe main focus of this dissertation, was the devel-opment of a GA that could be used in route com-putation within the SimuLTE framework. Throughthe analysis of available state of the art academicwork, the NSGA-II algorithm was chosen as thebase of the algorithm, due to its proven efficiencyand superior performance in comparison with otherexisting MOGAs. The new algorithm, will hence-forth be designated as Multiple Objective GeneticAlgorithm(MORGA). The only specifications re-garding the genetic operators, was the utilizationof the binary tournament selection scheme. TheMORGA’s computations had to take in account thefollowing objectives, total route length, the numberof interfering UEs within a certain range and energyconsumption.

The algorithm was implemented in C++ ,so thatit could be integrated in the OMNeT++ environ-ment. Initially the MORGA was intended to lever-age between D2D routes and simple two hop eNBroutes, but due to technical limitations in the sim-ulation engine, which only allowed for one type ofcommunication flow to be active at once during asimulation run. A. V. Bastos also took notice ofthis technical limitation in [18].

3.4.1 Data Encoding

The chosen encoding scheme was value encoding,which allowed increased readability during develop-ment, and avoided data size fluctuations since everynode was represented by a single value which had al-ways the same data size. Each value is implementedas a 8 bit character, obtained from the OMNeT++simulation source id string.

3.4.2 Value translation

The name translation routine compresses the sim-ulation name string, into an 8 bit character valuerepresentation which will represent a specific UEduring computations. The compression routine issimple, given an N character long id string,the rou-tine will check each character for an associated UEduring a given execution.The first available charac-ter will then be associated with the UE whose nameis under translation procedure. If the name stringis exhausted, the routine will search for the firstavailable character in the 8 bit ASCII table.

3.4.3 Objective Function Model

The optimal solution the algorithm tries to reachwill contains the minimum values of each fitness

6

regarding their correspondent decision variables.Therefore the problem can be defined as the mini-mization of a vector objective function given by thefollowing equation:

fff(ccc) = [f1(ccc), f2(ccc), f3(ccc)] (3)

Where ccc represents a possible solution. Eachsolution is composed of three vectors contain-ing the decision variables for each objective,ccc = [c1c1c1, c2c2c2, c3c3c3]. The first component isc1c1c1 = [c11, c12, c13, c14, ..., c1n], where n is equalto the number of nodes encoded in a givenchromosome. With the second beying c2c2c2 =[c21, c22, c23, c24, ..., c2n], containing the densityvalue for each node encoded in the chromosome.Finally c3c3c3 = [c31, c32, c33, c34, ..., c3n] which containsthe spent capacity of every node’s battery.

Let f : R3 → R3, where fi : R3 → R for each0 < i ≤ n. The first objective function f1, trans-lates the total number of nodes in the encoded net-work, as the sum of the lengths of all genes en-coded in a given chromosome f1(ccc) =

∑ni=0 c1i = n.

The second objective function, defined as f2(ccc) =Max{c2i} ∗ Penalty, will determine the maximumvalue and apply the corresponding penalty. Thethird objective function, f3(ccc) = Max{c3i}, will de-termine the maximum spent capacity value.

3.4.4 Termination Criteria

A solution is considered a global optimum if it main-tains itself as the best during a given number ofgenerations. The number of generations the bestsolution must remain unchanged in order for thealgorithm to terminate, is given by the followingequation:

a ∗ Ni

(4)

In which N corresponds to the population size ofthe NSGA-II implementation, i to the generation ofthe current population in a given execution, and athe termination coefficient which is a value empiri-cally obtained.

3.4.5 Genetic Operators

This section describes the genetic operators imple-mented in the MORGA. The implemented selec-tion scheme is based on binary tournament selec-tion. Pairs of chromosomes are randomly chosenfor tournament rounds until N have been selected.

The implemented crossover operator is based onthe 1-point crossover scheme. The cut point is equalto the index of the first equal value, that is neitherthe head or tail, of the value string in both genes.

The implemented mutation operator guaranteesthat every new gene has a chance to mutate, by

allowing at least one mutation each. During amutation routine, two indexes are randomly se-lected,these may range from 0 to the length of themutating gene minus one. The index selection pro-cess follows a uniform distribution. These two in-dexes,let us refer to them as I0 and I1 such asI0 < I1, will translate into two values correspon-dent to two existing nodes. After having chosentwo random node in the encoded route, the muta-tion operator will test the following conditions:

1. If the two nodes are in communication range.

2. If |I0 − I1| = 2.

if the first condition is verified,the value stringwill be truncated between the I0 and I1 indexes.Otherwise if the second condition is verified, theoperator will search for a third node, that is withincommunication range of both nodes encoded in I0and I1. If the third node actually exists, and is notthe same that is in position I2 such as I0 < I2 < I1,then the old value in I2 will be swapped for thenewly found one.

4. ResultsThis section will start by describing the simulationscenarios. Afterwards the simulation results will bepresented, alongside with the respective analysis.The performed simulation’s purpose is to verify theeffectiveness of the developed genetic algorithm incomparison to Dijkstra’s shortest path algorithm,in route computation for LTE-A networks.

4.1. Simulation ScenariosThree main network scenarios were used in the Si-muLTE simulations, the mesh, the dense mesh andthe congested dense mesh scenario. The mesh sce-nario consists of ten UEs , with two transmitters,two receivers and six relays. Each device is geo-graphically placed in such a manner that it is pos-sible to create a transmission bottleneck. In thisnetwork, the first transmitter to query the serveris transmitter1 followed by transmitter2. A sugges-tion of this scenario inside the OMNeT++ environ-ment can be seen in Figure 3(a).

The dense mesh scenario is a setup containing 32UEs, with three transmitters and three receivers,contained in a 70x70m2 area. This network hasa higher user density in comparison with the first.UEs are randomly placed within the limited area,with the intent of testing both algorithms in a ran-dom network topology. A suggestion of this setupcan be seen in Figure 3(b).

The congested dense mesh scenario is essentiallya spacial rearrangement of the second, in which itis divided in three different sub-spaces, of 40x20m2

, 10x10m2 and 20x20m2 areas. The first subs-spaceis connected to the third, through the second, and

7

(a) Mesh network scenario. (b) Dense mesh network scenario.

Figure 3: Mesh and dense mesh scenario.

each contain approximately 10 UEs, with all thetransmitters being located in the first sub-space andthe receptors in the third sub-space. Having the sec-ond sub-section as an intentional choke-point, thescenario will have the tendency to create traffic bot-tlenecks, allowing to evaluate the MORGA’s effi-ciency in a scenario similar to the mesh scenario,but denser.

4.1.1 OMNeT++ Simulations

Simulations are divided in two phases, the setupand transmission phase.

In the setup phase, each UE will initialize theroute application module, and either wait for com-munication or request the server to compute a rout-ing path. After receiving a query, the server ana-lyzes if it is possible to establish a route. If not,it will return a error message. Once the requestingUE receives the error message, it will enter a retrycycle until it receives an update to its routing tablemodule. An update message is sent by the server toall the intervening UEs in a successfully computedroute.

After a small delay, in order to allow all updatemessages to be received, the simulation will enter itssecond phase. During this phase, certain UEs thatwere chosen as transmitters will begin to generatetraffic, from either a (Voice Over IP)VoIP or (VideoOn Demand)VoD program. The generated trafficwill be forwarded through the established commu-nication routes.

4.2. Simulation Results

This section contains the results of OMNeT++ sim-ulations with the SimuLTE framework. The sim-ulation parameters are described in Table 2. Pa-rameters regarding the VoIP and VoD applicationsare described in Tables 3 and 4. Simulations wheremade in a Ubuntu 16.04 virtual machine with 4 pro-cessors and 4828 MB of base memory, using OM-NeT++ version 5.1.1 and SimuLTE refactoring ver-sion.

Table 2: Simulation parameters.Parameter Value

Carrier frequency 2 GHz

Bandwidth 5.4 MHz

Mobility model Stationary (OMNeT++ model)

Path loss model ITU Urban Macro

Fading Model Rayleigh fading(Jakes)

eNB Tx power 40 dBm

eNB Antenna gain 18 dB

Noise figure 5 dB

Cable Loss 2 dB

Simulation time 10 seconds

Repetitions 20

Scheduling Policy Proportional Fairness

AMC mode D2D

CQI 12

N 100

Termination coefficient 4

Table 3: VoIP application parameters.

Parameter Value

Packet size 40 bytes

Sampling time 0.02 seconds

Table 4: VoD application parameters.

Parameter Value

Packet size 51.2 kilobytes

Video size 524.43 megabytes

Send interval 10 miliseconds

The Table 5 makes a comparison in terms of net-work performance between the two algorithms re-garding the VoiP application execution. It is pos-sible verify that the MORGA allows the receptionnodes to achieve a 3.36% higher received through-put, with 34.6% less frame delay and 71.55% lessjitter while reducing the energy consumed per bitreceived by 9.88%. The MORGA used 21.68% moreUL resources. Though it may be explained by thehigher received throughput, which would have re-quired more resources.

In the same scenario Table 6 shows that for the

8

Table 5: Network performance VoIP in the meshscenarioAlgorithm Dijkstra’s shortest Path MORGA

Energy consumption(mJ/bit) 3.521 3.173

Received throughput(bps) 526.5172 544.810

Frame delay(s) 0.069328 0.045341

Jitter(s) 0.0811 0.023071

UL utilization(%) 0.65 0.83

DL utilization(%) 0.0001 0.0001

VoD application, the MORGA can achieve a 40.44%higher throughput rate, when comparing to theshortest path algorithm, with less 54.78% energyconsumption, despite showing an increase in end-to-end delay of 20.61%.

Table 6: Network Performance of VoD app in themesh scenarioAlgorithm Dijkstra’s shortest path MORGA

Energy consumption(µJ/bit) 2.373 1.073

Throughput 0.0757 0.1271

End-to-end delay(s) 2.0553 2.5900

UL utilization(%) 53.56 53.48

DL utilization(%) 0.0005 0.0004

In Table 7 it is possible to observe the networkperformance of the studied algorithms, regardingthe execution of a VoiP app in the dense mesh sce-nario. Performance results for the VoIP applicationin the dense mesh scenario are very similar for bothalgorithms, with Dijkstra’s shortest path algorithmachieving a slightly better performance. Regard-ing energy consumption, Dijkstra’s shortest pathalgorithm consumed 9.88% more energy per bit re-ceived, achieving a 5.40% higher received through-put. The delay was 19% higher for the shortestpath algorithm, but jitter was 17.95% lower, all incomparison to MORGA’s performance values.

Table 7: Network Performance for VoIP app in thedense mesh scenario.Algorithm Dijkstra’s shortest Path MORGA

Energy consumption(mJ/bit) 15.288 13.777

Received throughput(bps) 215.481 203.839

Frame delay(s) 0.06493 0.05259

Jitter(s) 0.0811 0.09885

UL utilization(%) 0.38 0.39

DL utilization(%) 0.00025 0.0003

The results of the simulations for the congesteddense mesh scenario, with the UEs running a VoIPapplication are shown in Table 8. For the congesteddense mesh scenario, the MORGA achieved a 7.60%higher received throughput, with 20% less energyconsumed. The frame delay was 27.65% lower, with65.59% less jitter, all in comparison with the Dijk-stra’s shortest path performance.

Table 8: Network performance for VoIP app in thecongested dense mesh scenario.Algorithm Dijkstra’s shortest path MORGA

Energy consumption(mJ/bit) 15.626 12.475

Received throughput(bps) 201.359 217.921

Frame delay(s) 0.0781 0.0565

Jitter(s) 0.1575 0.0542

UL utilization(%) 0.3814 0.3901

DL utilization(%) 0.0001 0.00015

5. ConclusionsIn this dissertation, the problematic of networkrouting for cellular communications in LTE-A, wasaddressed. Initially the purpose was to design arouting system which was fully integrated in thenetwork structure, in order to manage resource allo-cation, alongside route computation. Due to tech-nical limitations in the simulation framework, theidea of combining both features was dropped. De-spite these limitations, the system’s core design wasto decide when it was worth to establish a D2Dcommunication flow, in alternative to a more stableand better performing two hop communication flowusing the eNB.The system was established with aclient-server model, using UDP as the underlyingtransport protocol. UEs would run the client ver-sion, and the server version would be executing ona computer in an external IP network connected tothe eNB. Development efforts resulted in the cre-ation of the MORGA, which was written in C++in order to be integrated in the server applicationof the simulated LTE-A network. The implementedalgorithm had to leverage the impact of three dif-ferent decision variables, namely route length, in-terference and energy consumption.

Through multiple simulations it was possible toreach the conclusion that the MORGA was speciallyeffective in congested networks in which communi-cation flows tended to funnel, namely the mesh andthe congested dense mesh scenarios. Achieving anoverall higher throughput, less delay and less energyconsumption.

Future work should try to finish implementingthe battery module and shutdown function for theUE module in the SimuLTE framework. Testthe MORGA’s effectiveness in comparison to otherrouting algorithms. A different system model couldbe implemented, instead of a centralized client-server model, the system could be decentralized.Regarding non-congested scenarios, an heuristicroutine that could manage interference at a resourceallocation level in conjunction with a shortest pathalgorithm, could prove itself even more effectivethan the MORGA.

AcknowledgementsFirstly, i would like to express my deep gratitude forprofessor Antonio Grilo, whose endless availability

9

and constructive critique made the writing of thisdissertation a valuable learning experience.

Would also like to express my gratitude to SergioSabino for all the help related with the developmentof the genetic algorithm.

Finally i would like to thank Giovanni Nardini,for the recommendations regarding the simulationengine.

References[1] A. Virdis, G. Stea, and G. Nardini, “Simulte

- a modular system-level simulator for lte/lte-a networks based on omnet++,” in 2014 4thInternational Conference On Simulation AndModeling Methodologies, Technologies And Ap-plications (SIMULTECH), Aug 2014, pp. 59–70.

[2] M. N. Tehrani, M. Uysal, andH. Yanikomeroglu, “Device-to-device commu-nication in 5g cellular networks: challenges,solutions, and future directions,” IEEE Com-munications Magazine, vol. 52, no. 5, pp.86–92, May 2014.

[3] V. Yazc, U. C. Kozat, and M. O. Sunay, “Anew control plane for 5g network architecturewith a case study on unified handoff, mobility,and routing management,” IEEE Communica-tions Magazine, vol. 52, no. 11, pp. 76–85, Nov2014.

[4] 3GPP, “Overview of 3gpp release 12,” 3GPP,Tech. Rep., June 2014.

[5] X. Lin, J. G. Andrews, A. Ghosh, and R. Rata-suk, “An overview of 3gpp device-to-deviceproximity services,” IEEE CommunicationsMagazine, vol. 52, no. 4, pp. 40–48, April 2014.

[6] M. Usman, A. A. Gebremariam, U. Raza,and F. Granelli, “A software-defined device-to-device communication architecture for publicsafety applications in 5g networks,” IEEE Ac-cess, vol. 3, pp. 1649–1654, 2015.

[7] H. Nishiyama, M. Ito, and N. Kato, “Relay-by-smartphone: realizing multihop device-to-device communications,” IEEE Communica-tions Magazine, vol. 52, no. 4, pp. 56–65, April2014.

[8] G. Nardini, G. Stea, A. Virdis, D. Sabella, andM. Caretti, “Resource allocation for network-controlled device-to-device communications inlte-advanced,” Wireless Networks, vol. 23,no. 3, pp. 787–804, Apr 2017. [Online].Available: https://doi.org/10.1007/s11276-016-1193-3

[9] O. N. Yilmaz, Z. Li, K. Valkealahti, M. A.Uusitalo, M. Moisio, P. Lundén, and C. Wi-jting, “Smart mobility management for d2dcommunications in 5g networks,” in WirelessCommunications and Networking ConferenceWorkshops (WCNCW), 2014 IEEE. IEEE,2014, pp. 219–223.

[10] J. Park, “Fast and energy efficient multihopd2d routing scheme,” International Journalof Distributed Sensor Networks, vol. 12,no. 5, p. 2148734, 2016. [Online]. Available:http://dx.doi.org/10.1155/2016/2148734

[11] K. Kotecha and S. Popat, “Multi objective ge-netic algorithm based adaptive qos routing inmanet,” in Evolutionary Computation, 2007.CEC 2007. IEEE Congress on. IEEE, 2007,pp. 1423–1428.

[12] J. Huang, F. Qian, A. Gerber, Z. M. Mao,S. Sen, and O. Spatscheck, “A close exami-nation of performance and power characteris-tics of 4g lte networks,” in Proceedings of the10th international conference on Mobile sys-tems, applications, and services. ACM, 2012,pp. 225–238.

[13] “”omnet++ discrete event simulator - home”,”[Accessed: 31- May- 2017]. [Online]. Available:https://omnetpp.org

[14] A. Virdis, G. Nardini, and G. Stea, “Mod-eling unicast device-to-device communicationswith simulte,” in 2016 1st International Work-shop on Link- and System Level Simulations(IWSLS), July 2016, pp. 1–6.

[15] U. Mehboob, J. Qadir, S. Ali, and A. Vasi-lakos, “Genetic algorithms in wireless net-working: Techniques, applications, and is-sues,” Soft Comput., vol. 20, no. 6, pp.2467–2501, Jun. 2016. [Online]. Available:http://dx.doi.org/10.1007/s00500-016-2070-9

[16] D. F. Jones, S. K. Mirrazavi, and M. Tamiz,“Multi-objective meta-heuristics: An overviewof the current state-of-the-art,” European jour-nal of operational research, vol. 137, no. 1, pp.1–9, 2002.

[17] K. Deb, A. Pratap, S. Agarwal, and T. Me-yarivan, “A fast and elitist multiobjective ge-netic algorithm: Nsga-ii,” IEEE transactionson evolutionary computation, vol. 6, no. 2, pp.182–197, 2002.

[18] A. V. Bastos, C. M. Silva, and D. C.da Silva Júnior, “Algoritmo de roteamento as-sistido para comunicação d2d em sistemas desegurança pública sobre redes lte avançadas.”

10