Real mechanisms for mobile networks modeling and engineering

Real mechanisms for mobile networks modeling and engineering

Houda Khedher, Fabrice Valois*

Sami Tabbane

Ecole Supérieure des Communications de Tunis Route de Raoued – Cité El Ghazala 2083 Tunis, Tunisia, [email protected] *CITI, Bât. Léonard de Vinci, INSA Lyon, France [email protected]

Ecole Supérieure des Communications de Tunis Route de Raoued – Cité El Ghazala 2083 Tunis, Tunisia [email protected]

Abstract— As mobile networks are growing fast in size and complexity, radio network resource planning needs to be optimized to cope with both increased traffic demand and spectrum scarcity. Although, many studies have been carried on in mobile networks planning in terms of traffic distribution analysis and mobility modeling, relatively few studies have been published regarding the network planning based on real-world data and measurements. In this paper, a realistic model has been developed in order to evaluate the performance of a real GSM network providing a conversational service in an urban area. In this framework, our contribution addresses an important issue: how monitor the performance measurements collected and stored in the Operations and Maintenance Centre (OMC) to propose a complete study including a realistic mobility model and a qualitative and quantitative traffic characterization that constitute a generic planning model which can be customized by any opaerator for mobile networks deployment? Keywords- Mobile networks, mobility characterization, radio network planning.

I.

modeling,

traffic

INTRODUCTION AND MOTIVATION

With the continuous growth of subscribers in mobile communication services, the problems of system design optimization and radio network planning have become a major concern of mobile operators engineering teams. Moreover, mobile communication systems require increasing capacity, so imposing a revision of the overall approach of the process design in order to cope with the scarce radio resources. This implies careful consideration of user mobility and an accurate traffic analysis. To perform a proper capacity planning for mobile networks it is necessary to have a good knowledge of the behavior of mobile users, not only on an individual and on an aggregate level, but also on different time scales. Their design, planning and control must be supported by suitable traffic models that are able to deal with new sets of constraints in which the Quality of Service (QoS) management and mobility play an important role. Mobility and traffic models are needed in the design of strategies for location updating and paging, radio resource

management (e.g., dynamic channel allocation schemes) and technical network planning and design (network dimensionning). The choice of the mobility model and the traffic behavior have a significant effect on the obtained results. If the model is unrealistic, invalid conclusions may be drawn. In this paper, we present our approach to propose a generic model which can be customized and parameterized by any operator for planning and deployment of mobile networks. Our approach starts with the extraction of the relevant parameters from measurements realized on an operating GSM network providing a conversational service to a population of mobile users in an urban area covered by 197 cells. The outline of this paper is organized as follows: in section II we give a survey of traffic and mobility modeling approaches found in literature. Section III introduces our modeling approach based on the analysis of real measurements and outlines the statistical method used to characterize the arrival processes and the channel holding time. Section IV presents some simulation results. Section V concludes the paper. II.

VOICE TRAFFIC AND MOBILITY MODELS

In most analytical and simulation approaches, the performance evaluation of mobile networks is usually carried out by using some basic concepts of Queuing theory as well as assuming certain statistical distributions for offered traffic [5], channel holding time [3]. For past decades, the traffic modeling and dimensioning of voice networks has relied on the Erlang B call blocking model which was originally developed for the fixed network. The application of the Erlang theory assumes: that there are a large number of users, that the arrival process follows a Poisson process with inter-arrival times exponentially distributed, that there is full availability (i.e., an arriving call can use any free circuit) and that lost calls are cleared (i.e., they are rejected from the system and no repeat attempts). However, this is not always the case for real networks where mobiles within the coverage area are in general only able to seize circuits on the cell site serving that area and in such networks, blocked calls may be reattempted in neighboring cells. Also full availability may not be possible if a priority access scheme has been implemented.

Several works [4, 6, 7] have modeled traffic in cellular networks on the basis of homogeneous cell size and regular frequency reuse . It’s important to note that homogeneous cell size and regular frequency reuse implicitly assume that traffic density is homogeneous. This does not reflect at all the real networks where the number of mobile users in a coverage area is itself random and time varying. Due to users mobility, traffic density may vary by several orders of magnitude, not only on national scale but even within a BSC area. Though this can be observed in practically all networks, few studies have been made to treat it in the presence of real conditions [8]. As mentioned above, homogeneous mobile network have become a standard model for cellular radio networks, the cells in the network are studied one by one with stochastic models, e.g. Hong and Rappaport [4]. For such a system, every cell in the network can be considered to be statistically identical and independent to each other. Thus, by analyzing the performance of one single cell, the performance of the whole network can be characterized. In order to study a single cell independently in an isolated manner, it is essential to take into account the interaction between the cell under investigation and its neighboring cells. Thus, the offered traffic to a single cell can be modeled as consisting of two call arrival processes. The two arrival processes are on one hand, new call origination (i.e., new call attempt) process in the cell under investigation (calls which are originated inside the limits of the cell). On the other hand, the handover call arrival from surrounding cells. The analysis of traffic and system capacity for mobile systems needs to take into account both new calls and handover traffic. Generally, arrival processes are often modeled as Poisson process for analytic simplicity and because such processes have attractive theoretical properties: the inter-arrival times are both exponentially distributed and independent. Another parameter that appears an important element to be considered for mobile networks planning, is the channel holding time. It can be defined as the time duration between the instant that a channel is occupied by a call and the instant it is released by either completion of the call or by handing over to another cell. From classical teletraffic theories, it has been widely used the negative exponential distribution to model the channel holding time in large and single-cell systems for the sake of tractability [4]. Guérin [3] has extended this by attempting to describe the channel holding time by the negative exponential distribution and he has shown that the channel holding time can be seen as directly dependent on two other exponential processes. Namely, the total call duration and the crossing cell boundaries. However the channel holding time has been showed to fit lognormal distribution better than the exponential [9]. Various other methods such as the sum of hyper exponential [11] and general distributions [12] were proposed but complexity of the analysis has increased considerably with these techniques.

III.

MEASUREMENTS ENVIRONMENT AND DESIGN APPROACH

After giving an overview over state of the art models for describing the traffic and mobility in mobile networks, we are interested in this section to present first the measurements environment and second our design approach. Our approach starts with the analysis of the BSS (Base Station Sub-System) measurements collected and recorded in different counters according to a schedule established by the OMC (Operations and Maintenance Centre) over the hours of a day for a period of 16 days. These counters are sorted according to four measurements domains: handover (number of incoming external TCH handovers successes), quality of service (number of immediate assignment-successes for mobile originating procedure), resource availability and usage (average number of available TCH radio timeslot), traffic load (time in seconds during which the TCH radio timeslot is busy). The urban area under study is covered by 197 cells gathered into 5 BSC. The covered area is representative (area of 1333 Km 2 ) since it includes all the characteristics of a typical urban area such as business centres, highways and shopping centres. Our framework is based on two levels. At the first level called “macroscopic study” we proceed to split the urban area up to several sub areas (in our context we have found 8 activity areas: residential district, highways, business centre, downtown, industrial zone, university campus, tourist zone and shopping centre) according to the population activity and traffic characteristics. In order to define each area by the nature of the population activities (residential district, downtown, highways,...), we have investigated time of day traffic intensity patterns derived from call traffic intensity curves which describe the mean values of the traffic over weekdays (Mondays-Saturdays) and weekends (Sundays). Each activity area gathers together a number of cells sharing the same characteristics (cell size, number of channels and traffic intensity). In this way all cells belonging to the same activity area present approximately the same qualitative and quantitative behavior of the traffic intensity (they achieve their busiest hour at the same times of the day and having the same temporal and spatial variations). This property is verified for each activity area (Fig.1). After defining each activity area, we have been interested in defining certain statistical distributions of the arrival processes [2] namely fresh traffic and handover traffic and the distribution of the channel holding time [1]. At this second level called “microscopic study” we focus on the traffic modeling in the same way of macroscopic study but with more refinement in order to deduce the impact of each area on other areas and all kinds of interactions that can be established between them. As mentioned earlier, our approach is entirely statistical to derive distributions to fit the measured data. This is done by means of the Kolmogorov Smirnov (K-S) goodness of fit test. The K-S test is very simple to apply and has been widely used to fit telecommunication traffic. The general approach was as follows. The basic parameters (mean value and standard

deviation) of the hypothetical distribution function were estimated from the sample data by making use of the Maximum likelihood Estimation (MLE) [10]. Then we test the suitability of the fitting distribution by using the K-S test which is based on the differences between observed and theoretical cumulative distribution probabilities. In this article, we focus on the presentation of the results obtained only for one activity area (business centre). Interested readers can refer to [1, 2] for details of the others results related to different activity areas. Fig.1 shows the traffic intensity derived from our performance data provided by BSS counters in the business centre which is constituted by 2 cells. We have examined averages over all days (Mondays- Saturdays) ans weekends (Sundays).

Fig.1 Traffic intensity versus the hours of the day in the business centre

According to Fig.1, we notice that the traffic intensity during the week is almost the double of that of the weekend except for the first hours of the day between midnight and eight O’clock in the morning where the traffic intensity during the weekend exceeds that of the work days. It is shown that there are two peak hours along the work day; the first busy hour is around twelve O’clock and the second one is around six O’clock in the afternoon, this reflects the activity hours of the users when they usually leave their works. However, for the weekend, the first peak hour is at one O’clock in the afternoon and the second is a peak period between nine and ten O’clock in the afternoon. It’s important to note that before drawing the curves of this business centre, we verified that two cells included in this area have the same behavior (peak hours, traffic intensity ans same variations). Analyzing this graph, we found that determining the peak hours and principally the time evolution of the traffic along the day contribute not only to improve the utilization of resources but also to propose some tariff schemes. In order to characterize the arrival processes distributions, we have applied the Kolmogorov Smirnov test to verify the hypothesis that the new calls and handover calls data follow a particular probability distribution.

Fig.2 Poisson K-S goodness of fit test between 5 AM and 6 AM

Fig.3 Normal K-S goodness of fit test between 5 PM and 6 PM

Fig.2 is a plot of the empirical distribution function (new calls) with a Poisson cumulative distribution. The observations between 5 AM and 6 AM come from the poisson distribution at the 0.05 level of significance. Nevertheless, for the busy hour between 5 PM and 6 PM, we have observed (as shown in Fig.3) that empirical data follows perfectly the normal distribution. It’s important to note that our choice of these hours of the day came after applying the same test with the “fitted” poisson and normal distributions over each hour of the day. We deduced that empirical data satisfy the normal distribution over the work hours with parameters (mean value and standard deviation) different from one hour to another. However, for the little activity period from 1 AM to 7 AM, we have observed the suitability of the poisson distribution for each hour. This result appears quite logic since this period corresponds to the minimum blocking rate of the day. In the same manner of treating the new calls, we have applied the K-S test to handover calls. Fig.4 and Fig.5 display respectively the suitability of the poisson distribution and the normal distribution (handover calls) for two hours of the day in the same business centre, the first hour is between 5 AM and 6 AM and the second one is between 5 PM and 6 PM.

Fig.4 Poisson K-S goodness of fit test between 5 AM and 6 AM

Fig.6 Exponential K-S test between 7 AM and 8 AM

IV.

Fig.5 Normal K-S goodness of fit test from 5 PM to 6 PM

The fact that the handover arrivals follow the poisson process could be surprising. Indeed, during this period the blockage rate reaches its minimum. Thus this system can be approximated by the M / G / ∞ model where the outgoing flow is considered to obey the poisson distribution. To characterize the channel holding time, we have applied the K-S test to evaluate the agreement between the empirical data obtained from BSS measurements and the exponential distribution (Fig.6) which has usually been considered as a suitable distribution to model the channel holding time in most of the analytical and simulation studies. We found that the channel holding time follows an exponential distribution in each hour of the day with a mean channel holding time different from one hour to another at the 0.05 level of significance. It’s important to note that our motivation for channel holding time characterization research comes from the fact that it plays a very important role for our framework where we neglect all the intra-cell handovers and so in this case the channel holding time can be considered as the cell residence time. The latter has in turn huge influence on the call blocking and dropping probability.

SIMULATION RESULTS

In this framework, we have performed a simulation process in which we consider realistic hypotheses of a finite number of users moving in a geographical region represented by 35 cells arranged in 8 activity areas, taking into account users mobility and the arrival processes distributions obtained from our statistical analysis described in section III. In the simulation model, call duration, channel holding time and inter-arrival times are generated randomly by suitable exponential distributions. The cell transition process follows the transition probability matrix not reported in this paper. The simulator allow us to evaluate the main system performance parameters in terms of blocking probability of new calls and handover blocking probability. The motivation for studying the new call and handover blocking probabilities is that the Quality of Service (QoS) in mobile networks is mainly determined by these two metrics. To assert the validity of our model, the performances of a reference system will be computed, comparing BSS measurements with those produced by a suitable simulation program.

Fig.7 Comparison between simulated and measured probability of new calls blocking

V. CONCLUSION A realistic model for mobile communication networks has been developed using the management data collected in the BSS counters. It was found that the arrival processes in a real urban area could be well approximated by a Poisson process for little activity periods of users and by a Normal distribution for the other periods of the day. The channel holding time can still be taken exponentially distributed in the presence of real conditions. Also simulation results have been plotted; they confirm the BSS measurements in the considered system. The model developed in this paper can be employed to evaluate the performance of realistic networks and to describe accurately for a wide range of simulation scenarios the impact on planning process. Hours

Fig8. Comparison between simulated and measured probability of handover blocking

Fig.7 and Fig.8 show respectively the comparison between simulated and measured probability of new calls blocking and simulated and measured probability of handover blocking for the business centre from 7 AM to 8 PM. Also these two figures show that the two metrics are less than 2% along the day which is a threshold fixed in our context.

Fig.9 Flow traffic over the hours of the day in the business centre In Fig.9 we present the evolution of some parameters which have an important effect on the network performances such as the number of incoming handover, outgoing handover, initiated calls and terminated calls.

REFERENCES [1]

Houda Khedher, Fabrice Valois and Sami Tabbane, “Channel Holding Time Characterization in real GSM network”, PIMRC’03, September 2003, Beijing, Chine. [2] Houda Khedher, Fabrice Valois and Sami Tabbane, “Traffic Characterization for Mobile Networks”, VTC Fall 2002, pp. 1485-1489, September 2002, Vancouver, Canada. [3] Roch A. Guérin, “Channel Occupancy Time Distribution in a Cellular Radio System”, IEEE Transactions on Vehicular Technology, vol. Vt35, no. 3, August 1987. [4] D. Hong and S. S. Rappaport, “Traffic Model and Performance Analysis for Cellular Mobile Radio Telephone systems with Prioritized and NonPrioritized Handoff Procedures”, IEEE Transactions on Vehicular Technology, pp. 77-92, Aug. 1986. [5] G W.Tunnicliffe, A.R Murch, A.Sathyendran, PJ. Smith, , “Analysis of traffic distribution in cellular networks”, Proceedings of the IEEE Vehicular Technology Conference, vol.48, no.3, pp. 1984-1988, Ottawa 1998. [6] Kin K.Leung, William A.Massey, Ward Whitt, “ Traffic models for wireless communication networks”, IEEE Journal on Selected Areas in Communications, vol. 12, no.8, October 1994. [7] Sanjiv Nanda, “Teletraffic Models for Urban and Suburban Microcells: Cell Sizes and Handoff Rates”, IEEE Transactions on Vehicular Technology, vol. 42, no. 4, November 1993. [8] Uwe Gotzner, Andreas Gamst, Roland Rathgeber, “Spatial traffic distribution in cellular networks”, 48th IEEE Vehicular Technology Conference, pp. 1994-1998, Ottawa 1997. [9] C. Jedrzycki and V. Leung, “Probability Distribution of Channel Holding Time in Cellular Telephony Systems”, Proc. IEEE Transactions on Vehicular Technology. Conf., pp. 247-251, May 1996. [10] Raj Jain, “The art of computer systems performance analysis”, John Wiley & Sons, Inc, 1991. [11] P. Orlik and S. Rappaport, “A Model for Teletraffic Performance and Channel Holding Time Characterization in Wireless Cellular Communication with General Session and Dwell Time Distributions”, IEEE J. Sele. Areas of Commun., vol. 16, pp. 788-803, June 1998. [12] Y. Fang and I. Chlamtac, “Teletraffic Analysis and Mobility Modeling of PCS Networks”, IEEE Trans. Commun., vol. 47, pp. 1062-1072, July 1999.