Paper Preictive Moeling in a VoIP System Ana-Maria Simionovici a, Alexanru-Arian Tantar a, Pascal Bouvry a, an Loic Dielot b a Computer Science an Communications University of Luxembourg, Luxembourg b MIXvoip S.a, Luxembourg Abstract An important problem one nees to eal with in a Voice over IP system is server overloa. One way for preventing such problems is to rely on preiction techniques for the incoming traffic, namely as to proactively scale the available resources. Anticipating the computational loa inuce on processors by incoming requests can be use to optimize loa istribution an resource allocation. In this stuy, the authors look at how the user profiles, peak hours or call patterns are shape for a real system an, in a secon step, at constructing a moel that is capable of preicting trens. Keywors particle algorithms, preiction, user-profiles, VoIP. 1. Introuction In the past few years, many researchers focuse on eploying an improving the Voice over IP (VoIP) technology. VoIP platforms are subject to extremely fast context changes ue to the ynamic pricing an automatic negotiation, availability an competition for resources in share environments. One thus nees to provie powerful preiction moels that are able to automatically evolve an aapt in orer to consistently eal with the varying nature of a VoIP system. Furthermore, as such systems generally o not allow having a centralize management, mainly ue to scale concerns, one woul also like to have a moular approach for, e.g., resource allocation or loa balancing [1]. This stuy is built on a collaboration between the University of Luxembourg an MixVoIP, a company that hosts an elivers commercial VoIP services, with an important market share in the Luxembourg area an, at this time, a significant number of subscribe clients [2]. The clients of MixVoIP are small businesses, libraries, small companies. VoIP is any type of technology that can transmit, in real-time, converte voice signals into igital ata packets over an IP network. A VoIP service, e.g., as implemente an provie by a public or private company, allows placing calls over the Internet via an orinary phone or computer. A VoIP phone only nees to connect to a home computer network using a special aapter. One of the main avantages of using VoIP in enterprises for example is the reuce cost of sharing a certain number of external phone lines, furthermore avoiing the allocation of a line per user. The most well known telephone system capable of switching calls an of proviing a powerful control over call activity (through the use of channel event logging) is Asterisk [3]. It is a framework for builing multiprotocol, real-time communications solutions. Asterisk is in charge of establishing an managing the connection between two en evices by sening the voice portion of the call an everything that is not voice, also known as overhea. The protocols in charge of controlling multimeia communication sessions, respectively of elivering information an transferring ata are Session Initiation Protocol (SIP), an Real Time Protocol (RTP). SIP is a signaling communications protocol, wiely use for controlling multimeia communication sessions such as voice an vieo calls over Internet Protocol (IP) networks. SIP is one of the most known protocols in charge of signaling, establishing presence, locating users, setting, moifying or tearing own sessions between en-evices. After the connection is establishe, the meia transportation is one via RTP. Coecs are use for converting the voice portion of a call in auio packets an the conversation is transmitte over RTP streams. The calls are store by the VoIP proviers in a Call Detail Recor (CDR) atabase for billing. An example of a telephone system solution for VoIP is given in Fig. 1. The users connect via Internet to a server that runs either on a physical machine or in a ata center or in the clou. All users must be registere to a SIP Registrar Server in orer to inicate the current IP aress an the URLs for which they woul like to receive calls. The security layer can inclue SSH, FTP an access control or sessions for passwor protecte phone login. The voice noes control call features such as voice mail, call transfer, conference functions. The solutions are mainly eploye on Linux istribution. After authorizing the user to place calls, two links are mae. First, the user calls Asterisk that, in turn, connects to the atabase server in orer to check the creentials of the user an if the call can be mae, e.g., creit on a pre-pay account. Secon, Asterisk tries to reach the other en-evice. After the call is finishe, when the user hangs-up, call etailes are store in the CDR (e.g. client i, estination, prefix, call uration, billing). The situation presente above stans for outboun calls place by the users of the VoIP service provier. Depening on the coecs of each evice on the call path, the ecoing an re-coing operations increases the CPU usage. The VoIP provier is in charge of forwaring the call to ifferent telecom operators. For incoming calls, the car- 32
Preictive Moeling in a VoIP System Fig. 1. Basic example of VoIP telephone system. rier connects to the VoIP provier which, in turn, looks up for the number an returns the number an location where the user can be reache. Together with MixVoIP the authors work on a next generation clou-base prouct range, a moel that aapts to an copes with the highly ynamic evolution of requests, loa or other stochastic factors. At the same time, in orer to eal with a moel base on an extremely large number of parameters, all in a time-epenent framework we look at ynamic factors an stochastic processes while taking into account repeating activity patterns, failures, service level agreements. Thus, one nees to know how to efine time epenency in orer to brige, in a coherent manner, online (real life) processes, the concepts employe in moeling those aspects, an the (classically) static representations use in optimization. As an example, a specific aspect one can exploit is that activity patterns insie such systems ten, over specific perios of time, to express regular or localize characteristics, e.g., the evolution of energy price when relying on renewable sources, cyclic loa over a 24 hours span. The final goal of the project evelope by the University of Luxembourg an MixVoIP company is to implement such an approach within MixVoIP s environment, i.e., capable of proviing a preiction system for clou-base platform. The project can be split into two main research areas: preiction an loa balancing. Currently, we focus on builing the preiction moel (in charge of anticipating the number of the calls in the system) an be use as an input for the ynamic loa balancing. The transition to a irect implementation for the clou environment, to evelop intelligent loa balancing mechanisms that optimally sprea the traffic insie the clou, is therefore neee. To this en, with the support of MixVoIP, the paraigms esigne insie the project will also be put into practice, hence resulting into a fully functional preictive optimization system for real clou base VoIP platforms. The mechanism shoul be robust, flexible an scalable. Moreover, the loa balancing mechanisms shoul aress several requirements like increase scalability, high performance, high availability or failure recovery. The nee for resource capabilities arises as the cost-saving benefit of ynamic scaling is brought by the clou phenomenon an given that clou computing uses virtual resources with a non eligible setup time. Preiction is therefore necessary. Statistical moels that escribe resource requirements in clou computing alreay make the object of pay-as-you-go, e.g., proviers of environments that scale transparently in orer to maximize performance while minimizing the cost of resources being use [4]. By preicting the loa, allocating tasks to processors an ynamically turning off reserve computational resources, the high power require by clou computing system can be reuce rastically [5]. Clou computing can be seen as an overlay where seamless virtualization is implemente while ealing with privacy constraints. It relies on sharing computing resources rather than on having local servers or personal evices to hanle applications. Clou computing is use to increase capacity or a capabilities on the fly without investing in new infrastructure, training new personnel, or licensing new software. As a specific aspect, one may consier a computational eman an offer scenarios, where iniviuals or enterprizes negotiate an pay for access to resources through virtualization solutions (aministere by a ifferent entity that acts as provier) [6]. Clou computing encompasses any subscription-base or pay-per-use service that, in real time over the Internet, extens IT s existing capabilities [7]. Demaning parties may however have iverging requirements or preferences, specifie by contractual terms, e.g., stipulating ata security, privacy or the service quality level [8]. Moreover, ynamic an risk-aware pricing policies may apply where preictive moels are use either in place or through intermeiary brokers to assess the financial an computational impact of ecisions taken at ifferent time moments. Legal enforcements may also restrict access to resources or ata flow, e.g., ata crossing borers or transfers to ifferent resource proviers. As common examples, one can refer to Amazon Web Services [9] or Google Apps Clou Services [10]. The impact of task scheuling an resource allocation in ynamic heterogenous gri environments, given inepen- 33
Ana-Maria Simionovici, Alexanru-Arian Tantar, Pascal Bouvry, an Loic Dielot ent jobs has been stuie in [11]. The authors evelope a hierarchic genetic scheule algorithm, capable of elivering high quality solutions in reasonable time; makespan an flowtime are the two objectives consiere for the optimization problem (minimization). The algorithm improves the task execution time across the omain bounaries an the results are compare with mono-population an hybri genetic-base scheulers. The allocation of tasks to processors in new istribute computing platform is challenging when interconnecting a large number of processing elements an when hanling ata uncertaintie. In [12] a scheuler in charge of stabilizing this process is presente. The authors iscuss an algorithm where the application graph is ecompose into convex sets of vertices an it reuces the effects of isturbances in input ata at runtime. With respect to the previous consierations, the paper presents an algorithm capable of preicting the number of calls place in a VoIP system uring a given time frame. The ata collecte from MixVoIP is analyze an user profiles are outline. It has been observe that there are patterns in the number of calls place uring the ifferent hours of a ay, peak hours or in abnormal situations. During public holiays the number of calls heas to less than 10 per hour an, uring a normal ay, there are peak hours where the servers are overloae or hours when the traffic is very low. While MixVoIP has chosen to have over capacity in orer to avoi ropping the calls when there is a big traffic in the system, there are reasons to improve this approach an to implement a new moel capable of aapting to the number of calls while optimally allocating resources. Statistics are use as an input for the preiction moel which, base on a chosen time frame, can estimate the number of calls place uring the next time frame. The rest of the paper is organize in the following manner. Section 2 iscusses relevant backgroun relate to preiction moels, an motivates the importance of preicting the loa. In Section 3 the preiction moel an etaile information is introuce. Section 4 presents experiments, results. Finally Section 5 conclues an gives remarks about future work. 2. Relate Work A series of ifferent stuies looke in the past few years on preiction moels an VoIP. However, most existing approaches are base on preicting the speech quality of VoIP. The impact of packet loss an elay jitter on speech quality in VoIP has been stuie for example by Lijing Ding in [13]. He proposes a formula use in Mean Opinion Score (MOS) preiction an network planning. A parametric network-planning moel for quality preiction in VoIP networks, while conucting a research on the quality egraation characteristic of VoIP, was presente by Alexaner Raake in [14]. The work aresses to the ifferent technical characteristics of the VoIP networks linke with the features perceive by user. He gives a etaile escription of VoIP quality an iscusses how wieban speech transmission capability can improve telephone speech quality. In [15] the authors present a solution for non-intrusively live-traffic monitoring an quality measuring. Their solution allows aapting to new network conitions by extening the E-Moel propose by the International Telecommunication Union-Telecommunication Stanarisation Sector (ITU-T) to a less time-consuming an expensive moel. Another moel for objective, non-intrusive, preiction of voice quality for IP networks an to illustrate their application to voice quality monitoring an playout buffer control in VoIP networks is presente in [16]. They evelop perceptually accurate moels for nonintrusive preiction of voice quality capable of avoiing time consuming subjective tests an conversational preiction voice non-intrusive moels quality for ifferent coecs. The problem of traffic anomaly etection in IP networks has been stuie in [17]. The author presents an easy close form for preiction of the mean bit-rate of one conversation generate by SID-capable speech coecs as a function of the coec an the number of frames per packet use. Reference [18] explores the cumulative traffic over relatively long intervals to etect anomalies in voice over IP traffic, to ientify abnormal behavior when ifferent threshols are exceee. While extensive stuies exist in each of the mentione areas, only few sources consier such a holistic approach an analysis. Also, no conclusive agreement in the optimization omain exists on how to eal with highly ynamic timeepenent systems, causality an impact in a preictive framework, information coherence or escriptive power of the moels when facing a fast changing environment where ifferent scenarios are possible. Last, to the best of our knowlege, the use of such highly integrative techniques in a real worl setup has not been aresse to this extent ever before an woul therefore represent a premiere for the clou base VoIP commercial omain. 3. Algorithm The final goal of the project evelope by the University of Luxembourg an MixVoIP is to implement an approach that, first, improves the VoIP service an that, secon, scales to a clou-base-solution. In this section an algorithm base on interactive particle algorithms [19] is presente, aapte, e.g., for VoIP. It allows preicting the traffic in servers, the number of calls, base on previous observations. The moel is namely base on interacting particle algorithms use for parameter estimation, given a Gaussian moel [20]. Pierre Del Moral an Arnau Doucet efine interactive particle methos as extension of Monte Carlo methos, allowing to sample from complex high imensional probability istributions. The algorithm can estimate normalizing constants, while approximating the target probability istributions by a large clou of ranom samples terme particles. Each particle can evolve ranomly in the space an, base on its potential, can sur- 34
Preictive Moeling in a VoIP System Algorithm 1 Estimation of parameters pseuo-coe Generation of Particles, P(µ, Σ) Fix some population size, N Draw X N (µ,σ), for i = 1 N o { For each particle } Generate µ from N (0,1) e.g. Box Müller Determine the (lower) triangular matrix A via a Cholesky ecomposition of Γ as AA T Calculate X µ + AN (0,I ), ii variable from N (0,1) e.g. Box Müller Calculate Σ n i=1 X ixi T Likelihoo L = (2π) N /2 Σ N/2 exp N (x µ) T Σ 1 (x µ) i=1 2 en for Perturbation of Particles for k = 1 steps o Perturb the encoe vector W {X i N (0,Σ)} Draw samples {Y i N (0,Σ)}, 1 i n Construct a new vector {X i a X i + 1 a Y i }, 1 i n Perturb µ, µ µ + val, val generate from N (0,1) Calculate new sigma, Σ n i=1 X ixi T Calculate Lnew, likelihoo with new Σ an µ if Lnew > L then Set the new values of Σ, µ in the particle en if en for Algorithm 2 Preiction Step Choose the preiction metho Select the particle with maximum likelihoo an extract µ an Σ that best escribe the observe ata Let the istribution be N (µ,σ) For Z N (µ,σ) we consier the following partition: [ ] µ α µ =, Σ = µ β [ Σ α 1 Σ α ] 2 Σ β 1 Σ β 2 Z α Z β = z N (µ c,σ c ) µ c = µ α + Σα 2 (Σβ 2 ) 1 (z µ β ) Σ c = Σ α 1 Σα 2 (Σβ 2 ) 1 Σ β 1 Preiction base on the likelihoo of each particle, weighte likelihoo an the training test Let Ltotal = N i=1 L for i = 1 size(samples) o Z β = en for N i=1 Z α L i Ltotal vive or not. If a particle oes not survive it will not be brought back. Many applications took benefit of this intuitive genetic mutation-selection type mechanism in areas as nonlinear filtering, Bayesian statistics, rare event simulations or genetic algorithms. The sampling algorithm applies mutation transitions an inclues an acceptance rejection selection type transition phase. To this en, the approach is closely relate to evolutionary-life algorithm, though while proviing theoretical error bouns an performance analysis results. For a full escription, please refer to the work of Pierre el Moral [19]. The algorithm starts with N particles, enote by ξ0 i, 1 i N, that evolve accoring to a transition ξ0 i ξ 1 i, given a fix set A. If ξ 1 i A, it will be ae to the new population of N iniviuals (( ˆξ 1 i) 1 i N), else being replace by an iniviual ranomly from A. The sequence of genetic type populations is efine by ξ n := (ξn) i 1 i N selection ˆξ n := ( ˆξ n) i 1 i N mutation ξ n+1. ˆξ i n 1 ξn i, can be seen as a parent. An overview of convergence results incluing variance an mean error estimates, fluctuations an concentration properties is also given in [19]. For the estimation of parameters, a population of particles is generate as in the following. Each particle encoes a mean vector µ of size, a matrix, W {X i N (0,Σ)} sample from a Wishart istribution W(Σ,,n). It is use to moel ranom covariance matrices an escribes the probability ensity function of ranom nonnegative-efinite matrices [21]. The parameter n refers to the egrees of freeom while Γ in Algorithm 1 an enotes a scale matrix. After generating the initial population of particles, a perturbation step is repeate for a given number of times with a pre- specifie constant value, a. The perturbation step consists in moifying the parameters of each particle subject to istribution invariance constraints. We generate a new value val from N (0,Σ) an µ µ + val is calculate, an raw new samples {Y i N (0,Σ)}, 1 i n. The encoe vector is recalculate base on vector Y an value a, as escribe in the algorithm. After builing the new Σ, the likelihoo of the perturbe particle is etermine. If the value of the likelihoo is improve, then the perturbe particle is accepte an the ol one is elete from the population. This step is repeate for a number of times, in orer to etermine the values of the parameters that maximize the likelihoo. The perturbation can be seen as a transition or mutation of each particle. After the last perturbation step we get the final population. There are two preiction methos presente in Algorithm 2. First, the particle with the maximum likelihoo is chosen, an then the parameters with the first component (Z α, input for testing) are obtaine, the secon component (Z β ) is extracte. Secon, we take in consieration the likelihoo of all the particles. The component Z β is calculate via a weighte sum of the likelihoos multiplie by the first component an ivie by the sum of likelihoos. The result consists in an estimation of the traffic uring the secon time frame, for every working ay of the chosen year. 35
Ana-Maria Simionovici, Alexanru-Arian Tantar, Pascal Bouvry, an Loic Dielot Fig. 2. Data collection, ata analysis, preiction, ecision. 4. Experiments an Results The preiction moel has been traine an teste on real ata from MixVoIP. In Fig. 2 is given a escription of the path leaing from ata collection to ata analysis, Fig. 3. The total number of calls place in the thir week of April 2012. Fig. 4. The total number of calls place in the secon week of November 2012. preiction an then moification of the system. The first step collects an analyzes the information from the etaile recors of place calls, e.g., the account coe an profile i. One profile i can have multiple account coes. We can categorize the users base on the number of calls place an treat the main customers with a higher priority. The estination of the call is also recore an, after analyzing the ata we are able to see that, base on the location of the specific services, the calls are more likely to be place in that area. The uration of the calls is an useful information along with the ate of the call. Figures 3 an 4 are two examples of the total calls place uring two ifferent weeks from 2012. Throughout a ay, the highest traffic is uring working hours (8 11 an 13 18). Due the profile of the clients at MixVoIP, we can see in Figs. 5 an 6 that, over a week, the traffic is high from Monay till Friay, uring working hours. Abnormal situations can be reveale if the loa of the server becomes high outsie the etecte normal intervals an ays as we can see in Fig. 7. This information is use as input to preict the loa of the servers uring working ays an peak hours. As training set, the calls place in 2012 from hour 10 to hour 11 is chosen. We compare the accuracy of the presente preiction methos with a fee-forwar neural network approach [22]. Neural networks are wiely use for preiction problems, classification or control, in areas as iverse as finance, meicine, engineering, geology or physics. An artificial neural network is a computational moel inspire from the structure an function of biological neural networks. It is consiere to be a strong nonlinear statistical ata moeling tool where the complex relationships between inputs an outputs are moele or patterns are foun. Pattern classification, function approximation, object recognition, ata ecomposition are only a few examples of problems where artificial neural networks were applie. An artificial neuron receives a number of inputs corresponing to synaptic stimulus strength in a manner that mimes a biological neuron, are fe via weighte connections, an has a threshol value. Each neuron can receive x 1,x 2,...,x m inputs 36
Preictive Moeling in a VoIP System Fig. 5. Distribution of calls per weekay from the first week of February 2012. that can be represente by a irect graph with noes (neurons) an eges (the links between them). The fee-forwar neural network receives as input values that are associate with the input noes while the output noes are associate to output variables. Hien layers may also appear in the structure of the network [24]. Various types of ata can be preicte using neural networks, e.g., future value or trens of a variable (value increase or ecrease). In our moel, each ay is efine by two parameters: number of calls at hour 10, respectively 11, D(X,Y). Depening on the number of segments chosen, e.g. nrseg = 300, we can calculate, for each time frame, the size of the intervals as being the maximum number of calls for hour 10, respective by 11, ivie by the number of segments (size 1,size 2 ) (Algorithm 3). Each ay will be classifie as belonging to an interval for X, respectively Y. For, e.g., uring a public holiay X an Y of the respective ay will equal zero. We count how many ays belong to each pair (X,Y) an the classes will be use as an input for training the moel (Fig. 8). After training, the calls from hour 10 have been use to test the preiction moel, to anticipate the calls at hour 11. Algorithm 3 Extraction of samples pseuo-coe Fig. 6. Distribution of calls per weekay from the first week of August 2012. Store values from atabase in file, rea file, set values X an Y for each ay size 1 = max(x) nrseg, size 2 = max(y) nrseg Count the ays belonging to each interval where (i 1) size 1 X i size 1 an ( j 1) size 2 Y j size 2, i, j nrseg return samples We generate the particles accoring to Algorithm 1. The number of components we use (to train the algorithm) is = 2 which efines the imension of the space. The number of egrees of freeom chosen is n = 5. The N = 1000 particles will encoe the vector µ an the matrix Σ. Base on the parameters of the particles an the samples use Fig. 7. Abnormal situation, possible attack in the first week of September 2011. that have w 1,w 2,...,w m weights. Using the weighte sum of the inputs an the threshol, the activation of a neuron is compose. To prouce the output of the neuron, the activation signal is passe through, the neuron acting like the biological neuron. A fee-forwar [23] neural network is a collection of neurons that connect in a network Fig. 8. Before preiction sample of calls place in 2012 from 10:00-10:59 AM an 11:00-11:59 AM. 37
Ana-Maria Simionovici, Alexanru-Arian Tantar, Pascal Bouvry, an Loic Dielot Fig. 9. After preiction secon component calculate base on the weighte likelihoo of the particles. Fig. 10. After preiction secon component calculate base on the particle with maximum likelihoo. as an input, the likelihoo of each particle is calculate. This will be consiere as the initial population. After the generation, for each particle the values of the parameters µ an Σ will be perturbe accoring the algorithm. If the likelihoo is improve, the new values will replace the ol ones an move to the next step. The perturbation is applie for a number of steps, steps = 100. The final population of particles will be use for preiction that is one either by consiering the particle with maximum likelihoo (Fig. 9) or using the weighte likelihoo of the of particles (Fig. 10). For comparison, we traine an teste a neural network, using the same training set an testing set Fig. 11. Number of preictions below a given threshol. Mean absolute percentage eviation. (the first component) as previously iscusse. The neural network has nr = 10 neurons, one input, one output an two layers. We obtain the output of the neural network an we compare the preiction given by each metho. In Fig. 11, we can see a comparison of the mean absolute percentage error that was calculate using the expecte outcome an the output given by the three ifferent methos. For the threshol 0.3, the neural network metho has better results while for higher values, the best results are given by the average maximum likelihoo. This is what interests us: the preictor that works better for higher errors. The main reasons is that, for example, minus or plus 10 calls will not make a huge ifference. Miss-estimating 150 calls can have a negative impact on the system that can en up ropping calls because of lack of resources. For each metho, the Mean Absolute Percentage Error (MAPE) is calculate. MAPE is use in statistics to measure the accuracy of a metho for constructing fitte time series values. P(i) Z(i) P(i), MAPE is efine by the formula: M = 1 N N i=1 with N the size of the set, P(i) the actual value an Z(i) the forecast value. The smaller the value of MAPE is, the better is the moel. In Table 1, base on the calculate error we can see that the weighte likelihoo is the best solutions an we are consiering it. Table 1 Mean Absolute Percentage Error MAPE Value Maximum likelihoo 0.1359 Weighte likelihoo 0.0998 Neural network 0.1159 5. Conclusion an Future Work In this paper, a preiction moel we have built for a real Voice over IP system is presente. The authors have namely iscusse a metho capable to preict the loa of the traffic uring a time frame. It was also shown that there are patterns regaring the behavior of the users when placing calls. During public holiays, Sunays an Saturays, the number of calls is very low while uring the weekays, peaks are likely to appear. This information is helpful when taking ecision regaring resource allocations. The preiction moel propose in this work combines ifferent methoologies an will be use as an input for a future loa balancing moel. We also explore two ifferent methos for preicting the loa of the servers an compare the results. The first one takes in consieration the entity with the maximum likelihoo while the secon one is a weighte likelihoo base preiction. The propose methos were compare with the preiction given by a neural network that was traine an teste with the same ata. Since the VoIP implementations an the technology in itself eman a sustaine computational an banwith support, the ecrease of operation costs at infrastructure 38
Preictive Moeling in a VoIP System level is expecte together with the improvement of the VoIP quality. As future work, we will compare our moel with Support Vector Machines (SVM) [25], [26], an approach frequently use in machine learning community for classification problems an regression analysis. SVMs are for example use for learning an recognizing patterns for a given input. Moreover, we are investigating the use of a higher number of time frames for the input an testing. We believe that aing this feature coul help to improve our results. The behavior of the propose algorithms in cases like uniformly istribute loa, requests focuse on one specific set of resources or fast changing profiles that switch between low an high activity perios, will also be analyze. The iea of builing a moel per user is also taken into account. Depening on the perio of the year, the clients are more or less active. For example, knowing that a school will be close for summer holiays gives us the iea that less resources will be neee for that specific client. Fining patterns for the normal behavior of the system an efining a confience interval (in which the traffic can fluctuate normally) is consiere. Two ifferent situations will be istinguishe: peaks in normal situations (aapt the system to scale an loa balance the resources) an peaks in abnormal situations (etection an reaction against attacks). For the current experiments, the samples have been filtere an the abnormal traffic was remove from the training set. The existing preiction moel will also be extene an, base on the outcome, a highly efficient loa-balancing algorithm will be evelope, allowing to eal with constraints an performance measures not aresse to such an extent an in such an integrative manner before in the literature. Acknowlegements This stuy is conucte uner the support from the CNRS, France, with the National Research Fun, Luxembourg project INTER/CNRS/11/03 Green@Clou an Luxembourg ministry of economy, project DynamicMixVoIP. The aims an context of this research project are built on a collaboration between the Computer Science an Communications (CSC) Research Unit, University of Luxembourg, an MixVoIP, a Luxembourg base company specialize in VoIP services. National Research Fun, Luxembourg, http://www.fnr.lu. References [1] T. C. Wilcox Jr., Dynamic loa balancing of virtual machines hoste on Xen, Master thesis, Dept. of Computer Science, Brigham Young University, USA, April 2009. [2] Mixvoip Home page [Online]. Available: http://www.mixvoip.com/ [3] L. Masen, R. Bryant, an J. V. Meggelen, Asterisk: The Definitive Guie, 3r eition. O Reilly Meia, 2011. [4] A. Ganapathi, C. Yanpei, A. Fox, R. Katz, D. Patterson, Statisticsriven workloa moeling for the Clou, in Proc. IEEE 26th Int. Conf. Data Engin. Worksh. ICDEW 2010, Long Beach, CA, USA, 2010, pp. 87 92. [5] S. Kim, J.-I. Koh, Y. Kim, an C. Kim, A science Clou resource provisioning moel using statistical analysis of job history, in Proc. 9th IEEE Int. Conf. Depen. Autonom. Sec. Comput. DASC 2011, Los Alamitos, CA, USA, 2011, pp. 792 793. [6] M. Armbrust et al., Above the clous: A Berkeley view of clou computing, Tech. Report no. UCB/EECS-2009-28, Electrical Engineering an Computer Sciences University of California, Berkeley, USA, 2009. [7] D. F. Parkhill, The challenge of the computer utility. Reaing: Aison-Wesley, 1966. [8] V. Stantchev an C. Schrpfer, Negotiating an enforcing qos an slas in gri an clou computing, in Avances in Gri an Pervasive Computing, N. Abennaher an D. Petcu, Es. LNCS, vol. 5529. Berlin-Heielberg: Springer, 2009, pp. 25 35. [9] Amazon Elastic Compute Clou (Amazon EC2) [Online]. Available: http://aws.amazon.com/ec2/ [10] Google Clou Platform [Online]. Available: https://clou.google.com/ [11] J. Kołoziej an S. U. Khan, Multi-level hierarchical genetic-base scheuling of inepenent jobs in ynamic heterogeneous gri environment, Information Sciences, vol. 214, pp. 1 19, 2012. [12] A. Mahjoub, J. E. Pecero Sánchez, an D. Trystram, Scheuling with uncertainties on new computing platforms, J. Comp. Opt. an Appl., vol. 48, no. 2, pp. 369 398, 2011. [13] L. Ding, Speech quality preiction in VoIP using the extene E-moel, in Proc. IEEE Global Telecom. Conf. GLOBECOM 2003, San Francisko, USA, 2003, vol. 7, pp. 3974 3978. [14] A. Raake, Speech Quality of VoIP: Assessment an Preiction. Chichester: Wiley, 2006. [15] M. AL-Akhras, H. Zean, R. John, an I. ALMomani, Non-intrusive speech quality preiction in VoIP networks using a neural network approach, Neurocomput., vol. 72, iss. 10 12, pp. 2595 2608, 2009. [16] L. Sun an E. C. Ifeachor, Voice quality preiction moels an their application in VoIP networks, IEEE Trans. Multim., vol. 8, no. 4, pp. 809 820, 2006. [17] R. Estepa, Accurate preiction of VoIP traffic mean bit rate, Elec. Lett., vol. 41, pp. 985 987, 2005. [18] M. R. H. Manjes, I. Saniee, an A. Stolyar, Loa characterization, overloa preiction, an anomaly etection for voice over IP traffic, in Proc. ACM SIGMETRICS Int. Conf. Measur. Moel. Comp. Sys., Cambrige, MA, USA, 2001, pp. 326 327. [19] P. Del Moral an A. Doucet, Particle methos: An introuction with applications, LNCS/LNAI Tutorial book no. 6368. Springer, 2010 2011. [20] P. Del Moral, A.-A. Tantar, an E. Tantar, On the founations an the applications of evolutionary computing, in EVOLVE A Brige between Probability, Set Oriente Numerics an Evolutionary Computation, E. Tantar et al., Es. Stuies in Computational Intelligence, vol. 447. Springer, 2013, pp. 3 89. [21] S. W. Nyick, The Wishart an Inverse Wishart Distributions, 2012 [Online]. Available: http://www.math.wustl.eu/ sawyer/ hmhanouts/wishart.pf [22] I. Aleksaner an H. Morton, An Introuction to Neural Computing. Lonon: Chapman an Hall, 1990. [23] M. Buinich an E. Milotti, Properties of feeforwar neural networks, J. Phys. A: Mathem. Gen., vol. 25, no. 7, 1992. [24] D. Svozil, V. Kvasnicka, an J. Pospichal, Introuction to multilayer fee-forwar neural networks, Chemometrics Intell. Lab. Sys., no. 39, pp. 43 62, 1997. [25] N. Cristianini an J. Shawe-Taylor, An Introuction to Support Vector Machines. New York: Cambrige University Press, 2000. 39
Ana-Maria Simionovici, Alexanru-Arian Tantar, Pascal Bouvry, an Loic Dielot [26] T. Joachims, Text categorization with Support Vector Machines: Learning with many relevant features, in Proc. 10th European Conf. Machine Learn. ECML 98, Lecture Notes in Computer Science, vol. 1398. Springer, 1998, pp. 137 142. Ana-Maria Simionovici is a Ph.D. caniate at the University of Luxembourg. She is currently working on evolutionary computing, preiction, loa balancing. She hols Master s egree in Computational Optimization from University of Al. Ioan Cuza, Iasi, Romania (2012). E-mail: ana.simionovici@uni.lu Computer Science an Communications University of Luxembourg Campus Kirchberg, E 009 6 rue Couenhove-Kalergi L-1359 Luxembourg Pascal Bouvry earne his Ph.D. egree ( 94) in Computer Science with great istinction at the University of Grenoble (INPG), France. His research at the IMAG laboratory focuse on mapping an scheuling task graphs onto Distribute Memory Parallel Computers. He is now Professor at the Faculty of Sciences, Technology an Communication of the University of Luxembourg an heaing the Computer Science an Communication research unit. Professor Bouvry is currently holing a full professor position at the University of Luxembourg in computer science. His current interests encompass optimization, parallel/clou computing, a hoc networks an bioinformatics. E-mail: pascal.bouvry@uni.lu Computer Science an Communications University of Luxembourg Campus Kirchberg, E 009 6 rue Couenhove-Kalergi L-1359 Luxembourg Alexanru Tantar receive his Ph.D. iploma in Computer Science in 2009 from the University of Lille. He is now a Research Associate at the University of Luxembourg, conucting research on parallel evolutionary computation, the moeling an optimization of large scale, energy-efficient ynamic systems an Monte Carlo base algorithms. E-mail: alexanru.tantar@uni.lu Computer Science an Communications University of Luxembourg Campus Kirchberg, E 009 6 rue Couenhove-Kalergi L-1359 Luxembourg Loic Dielot is the current Founer at Pino S.a., CEO at Corpoinvest holing, Co-Owner at Forschung-Direkt Company. Since February 2008 he is also the Founer an Co-owner of MIXvoip S.A. He has competences in builing highly scalable an secure proucts base on linux an open source solutions while consiering the commercial alternatives. He has great know-how in web evelopment (PHP, CSS, Javascript, AJAX), Linux an security, Asterisk. He is intereste in atabase applications that nee to scale out. E-mail: lielot@mixvoip.com MIXvoip S.a Luxembourg Z.I. Rolach L-5280 Sanweiler, Luxembourg 40