Approximate Analytical Models for Dual-Band GSM Networks Design and Planning

Michela Meo and Marco Ajmone Marsan Dipartimento di Elettronica Politecnico di Torino 10129 Torino, Italy michela,ajmone
Abstract— In this paper we consider dual-band GSM networks, where voice and data services are offered to users moving over an area covered with overlapping macrocells and microcells. For this wireless network context we develop simple approximate analytical models of the system dynamics, and we exploit such analytical models for the design and planning of the critical system parameters, with particular attention to the number of traffic channels to be activated within macrocells. Keywords— GSM Networks, Resource Planning, Performance analysis, Approximations.

I. I NTRODUCTION The new version of the GSM standard is based on the use of two separate frequency bands, around 900 MHz and 1.8 GHz, respectively. Cells served by frequencies in the 900 MHz band are rather large (up to 35 km around the base station), whereas cells served by frequencies in the 1.8 GHz band are much smaller (typically less than 1 km), due to the much worse propagation characteristics of microwaves in the latter frequency range through the atmosphere. For this reason, cells served by frequencies in the 900 MHz band are normally called “macrocells”, whereas cells served by frequencies in the 1.8 GHz band are often called “microcells”. The main advantage relating to the use of microcells in a GSM network lies in an improved spatial reuse of frequencies, hence in substantial capacity increases, with the consequent possibility of offering, in addition to telephony, data services at mediumhigh rates (up to hundreds of kb/s), and even multimedia services through the integration of voice and data traffic flows. The main disadvantage relating to the use of microcells in a GSM network lies in the fact that, given the user mobility pattern, the number of handovers during a connection increases for decreasing cell size. This can be a critical factor, since the design and planning of cellular communication networks aim at meeting specified constraints on handover failure probabilities. More precisely, the design and planning of GSM networks are based on a number of performance parameters, among which are: i) the average number of active calls within a cell, which is an indirect metric of the revenues generated by the installed equipment; ii) the handover failure probability, which must be kept very low in order to avoid user dissatisfaction; iii) the new call blocking probability (the probability that a new call cannot be established due to the lack of free channels), which must be kept small, specially if multiple operators offer mobile telephony services in the same area.
This work was supported in part by the Italian National Research Council.

The design and planning of GSM networks require accurate models for the computation of the number of frequencies to be activated in cells (each frequency can support up to eight traffic channels), so as to obtain acceptable performance. Markovian models have been traditionally used for the design and planning of mobile cellular telephony networks, considering one cell at a time (see for example [1], [2], [3], [4], [5], [6], [7]). While this approach proved adequate for networks comprising only macrocells, it cannot be directly transferred to the dualband environment, where the minimum network element that has to be considered consists of one macrocell and all the microcells comprised within the macrocell. This network element will be called a “cell cluster”. This subsystem is rather complex for the direct development of Markovian models. For this reason, approximate models were proposed in the literature. In [8], [9], [10] the flow of users from microcell to macrocell is based on a Markov Modulated Poisson Process representation, while in [11] the same flow of users is derived from the computation of the average residual service time. The approach we adopt in this paper to support design and planning of dual-band GSM networks, instead, is similar to the one in [12]. We focus on networks where users exhibit different mobility patterns; we also extend our approach to cope with networks where different classes of service are provided to end users in terms of different bandwidth requirements. The paper is organized as follows. In Section II we describe the characteristics of the cellular mobile communication network, together with the probabilistic assumptions that are needed to describe the system dynamics with Markov processes. In this first description we consider different user mobility patterns, but just one service class. We then present the approximate model and its solution: Sections III and IV respectively describe the models of the microcells and of the macrocell. In Section V we derive expressions for the performance indices of interest. Resource planning strategies are illustrated in Section VI, where numerical results are also shown and discussed. An extension of the basic model allowing data connections to be also considered, is presented in Section VII. Section VIII concludes the paper. II. S YSTEM


A dual-band GSM network is considered, where each cell (microcell or macrocell) is served by a different base station (called Base Transceiver Station – BTS – in GSM jargon). We focus on a particular area, covered by one macrocell and Ñ

while slow users are preferentially assigned channels in their microcell. iii) handover requests out of the macrocell.. The handover procedure for slow users is quite similar. while each microcell is equipped with Æ ´Ñµ channels. due to the reduced spatial reuse of macrocell frequencies. III. 1. The process of call request arrivals into the microcell is the superposition of two Poisson processes corresponding to new call requests (with parameter ´Ñµ ) and to incoming handover ´Ñµ requests (with parameter ). A pictorial representation of a cell cluster containing seven microcells is given in Fig. then we shall also extend our techniques to data services requiring several channels per call. One cell cluster comprising seven microcells microcells (one cell cluster). Fast users are assumed to be served only by macrocells. showing that some of those can be released with no additional complexity in the model. their call is transferred to the macrocell. with the additional complexity that when no channels are available in the new microcell.Fig. T HE M ICROCELL M ODEL Each microcell is modeled as an M/M/K/0 queue with Ã Æ ´Ñµ servers which represent traffic channels available in the microcell. The call arrival process into each cell cluster is governed by four different dynamics. it needs not handover until the macrocell border is reached. If no channel is available in the new macrocell entered by the fast user. iv) handover requests out of the microcells. The impact of these exponential assumptions will be discussed later on. Calls taking place in different cells (micro or macro) are assumed not to interfere with each other. Fast users can roam from macrocell to macrocell during calls: an active fast user (i. The assumption of exponentially distributed call holding times is known to be inadequate for data services. ii) completions of microcell calls. that require one channel per call. whereas fast user requests are accepted until some free channel exists in the macrocell. iv) incoming slow user handover requests. as explained below). Once a slow user is allocated a macrocell channel. Different system policies can however be easily accounted for. a fast user that has established a call) that roams from a macrocell to another. the service time of means are respectively  ½ and the M/M/K/0 queue representing the microcell is thus exponen´Ñµ tial with rate · . must execute a handover procedure transferring his call from the channel in the old macrocell to a channel in the new macrocell without interrupting the communication. but if none of those are available. The process of call departures from the cell cluster is also driven by four types of events: i) completions of macrocell calls. fast users can be favored by reserving Æ channels in the macrocell for their exclusive use. The rationale for this assumption is that the allocation of microcell channels to fast users may entail an excessive number of handovers. such as In- . and assume that the macrocell is equipped with Æ ´Å µ channels. In the development of the analytical model to support the design and planning process we introduce the following assumptions: ¯ The aggregate process of new call requests from (fast or slow) users within a (macro or micro) cell is Poisson. ii) new slow user call requests. corresponding to: i) new fast user call requests. the handover fails and the call must be terminated (or dropped). We initially only consider voice services. with parameter ´Å µ (fast users in macrocells) or ´Ñµ (slow users in microcells) ( ´Å µ and ´Ñµ are estimated from the user population and the system geometry). which are established provided that channels are available in the cell (fast users only contend for channels in macrocells. 1. Fast users represent mobile terminals (MTs) used inside moving vehicles to establish voice connections.e. Two classes of users request services from the network. the call can access unreserved channels in the macrocell comprising the newly entered microcell. iii) incoming fast user handover requests. and are obtained from × the observation of the user mobility). with modifications of the model presented in this paper. otherwise the new call request is rejected. ¯ The time between two successive handover requests of a call (the call dwell time) is assumed to be an exponentially dis´Å µ (fast users in tributed random variable with parameter ´Å µ ´Ñµ (slow macrocells) or × (slow users in macrocells) or ´Å µ ´Å µ ´Ñµ users in microcells) ( . ¯ The flow of incoming handover requests from other (macro or ´Å µ (fast users in macrocells) micro) cells is Poisson with rate ´Ñµ ´Å µ ´Ñµ or (slow users in microcells) ( and are derived by balancing the incoming and outgoing handover flows. In order to avoid the risk that the quality of service perceived by fast users degrades due to slow users accessing the macrocell. Users can generate requests for new calls. they can also access the unreserved channels in the macrocell). while the allocation of macrocell channels to slow users may lead to a waste of resources. but if no channels are available within their microcell. This means that slow user requests are accepted as long as more than Æ channels are free in the macrocell. while slow users are preferentially served by microcells. The call duration and the call dwell time are modeled by random variables with negative exponential distribution whose  ½ ´Ñµ . ¯ The call duration is an exponentially distributed random variable with parameter ( is obtained from the observation of the user behavior). while slow users account for MTs used by pedestrians.

´Ñµ ´Ñµ ½ ¼ s. This independence assumption is justified by the results presented in [16]. we can take advantage of the fact that their steady-state distribution is insensitive to the service time distribution type. we need to characterize this overflow process. Again. In order to accurately model the macrocell.ternet access. For the computation of Ó we operate in the following way. where it was shown that more complex (multi-cell) models do not lead to significant improvements in the accuracy of performance predictions. and  ½ is the mean service time. assuming that both the mobility model and the call duration are exponential. and may lead to drastic approximations (see for example [13]). Furthermore. 2 we report the coefficient of variation values versus the load per channel. In order to validate the approximation. We rely on this result. we used a simulator of the call-level behavior of a cluster comprising Ñ microcells and one macrocell. However. ¾ and Æ ½ . Furthermore. and corresponds to an assumption of independence among cell behaviors. depending only on the service time average value. However. we still have to consider the superposition of Ñ such processes in order to characterize the overall overflow process entering the macrocell from all Ñ microcells. after having obtained this distribution. We derive the average number of calls which overflow from a microcell in the time unit: Ó ´Ñµ ´Ñµ · ´Ñµ ´ Æ ´Ñµ µ Æ ´Ñµ µ and we set: Ó ´ Ñ Óѵ Ñ ´Ñµ · ´Ñµ ´ Of course. in [15] the handover request arrival process is taken to be Poisson. some authors also questioned the assumption of exponential mobility models (see for example [14]). [4]. it tries to access a channel in the macrocell. phase-type or Weibull. the superposition of Ñ renewal processes is not a renewal process. we can assume that in steady-state the average handover flow entering a cell equals the average handover flow leaving ´ÓÙص . defined as Æ ´Ñµ . [3]. it tends to a Poisson process (see [18]). Let ´ µ be the steady-state probability that channels are busy in the microcell. ´Ñµ can be iteratively computed by imposing that . The results in [17] state that the overflow process from a GI/M/K/0 loss system is a renewal process with interval distri´ÓÙص ´Ñµ The computation of coefficients and parameters is rather difficult. and was validated by comparison with simulations in [2]. Thus. that corresponds to an M/M/K/0 queue with Ã Æ ´Ñµ . as Ñ tends to infinity. From classical queueing theory. We can easily specialize (1) to our case. instead of exactly deriving (2). but recently some researchers have even argued that the exponential assumption is not realistic in the context of telephony. and approximate the slow user call request process overflowing from all microcells to the macrocell with a Poisson process with parameter Ó . we approximate it. and in our notation ´Ñµ È ´Ñµ ´Æ µ. other distributions. In Fig. such as Pareto. The solid with ½ . since the steady-state distribution of a pureloss system is insensitive to the service time distribution. but. ´ÓÙص Æ ´Ñµ ½ ´Ñµ ´ µ Then. It can be shown that this distribution is hyper-exponential: ´ µ Ø Æ ´Ñµ ¼   Ø Ø ¼ (2) for ½ Æ ´Ñµ with Æ ´Ñµ ½ with ´Ñµ · ´Ñµ ´Ñµ and ´¼µ ¼ ´Ñµ Æ  ¼ ½ ½ ¼ · ´Ñµ The incoming handover flow is not known a-priori. paying attention to the evaluation of average values. and for a given value of Ñ we measured the coefficient of variation of the interarrival intervals for the overflow process out of a single microcell and for the superposition of the overflow processes from all microcells. is given by the Erlang-B formula. obtaining: ¼´ µ × ´Ñµ × · ´Ñµ · ´Ñµ · ´Ñµ From the inverse transform of (1) we can obtain the distribution of the interarrival times of the overflow process of call requests of slow users from a microcell towards the macrocell. This approach has been widely and successfully used in the literature when considering individual cells. as long as we can exploit M/G/K/0 queuing models. The probability that all channels in the microcell are busy. the accuracy of this approximation increases for increasing number of microcells in the cell cluster. In general. When a new call request finds all microcell channels busy. we have: ´ µ ´¼µ bution whose Laplace transform can be derived from the recursive formula: à ´×µ ½   à  ½ ´× · µ à  ½ ´×µ · à  ½ ´× · µ (1) where ¼ ´×µ is the characteristic function of the interarrival times at the queue. the latter can be computed as: the cell. and we can thus adopt an M/M/K/0 queuing model for convenience. which forms the arrival process of slow user call requests at the macrocell. È ´Ñµ . but the mobility model assumes an hyper-Erlang distribution for the dwell time. obtained from simulation experiments. we can adopt an exponential mobility model for convenience. are today much more fashionable.

0 0.5 1. the blocking probabilities for new call and handover requests coincide. V. which are most interesting for the system design. IV. which of course do not depend on Ñ. the arrival process at the macrocell is the superposition of two Poisson processes corresponding to new call requests and incoming handover requests (with rates ´Å µ and ´Å µ . T HE M ACROCELL M ODEL The behavior of the macrocell is more complex than that of the microcell. respectively). the fast user service time is a random vari´Å µ able with negative exponential distribution with rate · .5 1 1.5 2 2. and for the superposition of Ñ overflow processes (dashed lines) Ë · Æ ´Å µ Æ ´Å µ   Æ (4) line refers to overflow processes out of individual microcells.5 microcell m=2 m=4 m=8 m=16 due to the characteristics of call duration and dwell time for fast users. Indeed. where Æ is the number of channels reserved to fast users. the service time is a random variable with negative exponential ´Å µ distribution with rate · × . and the call duration and dwell time are random variables with negative exponential dis´Å µ tributions with rates respectively equal to and × . i. Dashed lines instead provide results for the superposition of Ñ overflow processes. 2. we have: ´Å µ · ´Å µ ´Å µ · × Ó · ´Å µ × The steady-state probability ´ µ that channels are busy serving fast user calls and channels are busy serving slow user calls is: × ´¼ ¼µ ´ µ (3) for all states × ´ ´ µ µ in the state space Ë : Fig. even for large systems. and even when Ñ is smaller than in typical dual-band GSM networks. ρ/N(m) 3 3.5 Load per channel. ´Å µ The value of is derived by balancing incoming and outgoing handover flows for fast users. low channel loads surely lead to acceptable performance. because of either a failure in the access to the system at call setup. We denote such probain the case of fast users and × in the case of slow bilities by users. with Ñ ¾ ½ . the coefficient of variation gets very close to 1 for any Ñ at higher values of the channel load. For slow users. are given by: tively denoted by È× È×´Å µ ´ µ ¾Ë× ´ ´ ´ µ and È ´Å µ ´ µ ¾Ë ´ µ (5) with Ë× Ë µ µ · · Æ ´Å µ Æ ´Å µ Æ ´Å µ   Æ For both classes. due to the presence of two classes of users (slow and fast). For fast users. the probabilities that calls are not successfully completed. The value of ´¼ ¼µ is obtained from the normalization equation: ½ ¼  ½ ´¼ ¼µ  ´ × µ ¾Ë The number of states is rather small. the coefficient of variation rapidly approaches 1. We can see that as Ñ increases. which is the value characterizing a Poisson process. Thus. Coefficient of variation of interarrival intervals of overflow processes. the coefficient of variation is close to 1 for most load values.5 2. P ERFORMANCE I NDICES The performance indices that we consider in this paper for GSM network design and planning are the call blocking probabilities. Denoting by and × the loads generated by fast and slow users. for individual microcells (solid line). the arrival process at the macrocell is assumed to be Poisson with rate Ó . we can model the macrocell as an M/M/K/0 queue with two classes of users.5 4. so that the normalization is not critical. . of the order of ´Æ ´Å µ µ¾ ¾.e.5 3.4. respec´Å µ ´Å µ and È . that may induce unacceptable performance.. Having assumed that all the macrocell dynamics have Markovian characteristics. The outgoing handover flow is computed as: ´ÓÙص ´ µ ´Å µ ¾Ë ´ µ The blocking probabilities for slow and fast users. or a failure during handover. and thus evaluate the performance of this system with generalizations of Erlang’s loss formula. Moreover. respectively. the situations where the system design instruments must be accurate refer to high loads. The convergence is rather fast.0 2.0 1.0 Coefficient of variation 3.

× .The probabilities and × can be computed as the complement to 1 of the probabilities of successful completion. i. given that it requested handovers. a slow user call that was accepted in the system completes with success (possibly after some handovers in other microcells) either in a microcell or in the macrocell. the remaining channels are shared between fast and slow user calls. the resource planning objective is taken to be guaranteeing a fair treatment to slow and fast users. such that a specified call blocking probability can be guaranteed to fast users. deriving the number of channels necessary to guarantee that the call blocking probabilities for both classes of users are smaller than a given threshold. and serve fast user calls as well as slow user calls that cannot find free channels in microcells. can finally be   ½ ½  È ´ µ ½ ´Å µ ·½ À µ ´½  À  È   ´½µ ´ µ ½ ¼ È ´¾µ · È ´¿µ ´ µ µ´½ µ ½  È Å  À      È Å À ½ ´½ ½ ¼ ½ ´ µ (6)  È ´Ñµ ´½   È× Å  È Ñ     È Ñ À× µ · ´½ ½ ´½ ´ µµ   À× µ (7) VI. and system parameter values are sought. The following three subsections separately consider the three cases. it fails if no channel is available either in the microcell or in the macrocell (the probabilities of ´Å µ such events are denoted by È ´Ñµ and È× . leaves the microcell because of handover: ´Ñµ À× ´Ñµ · Neglecting handovers of slow users at the macrocell borders (which are rather infrequent). Therefore. the failure of the ´ · ½µ st handover forced the call to access a channel in the macrocell. which in turn are obtained as the sum of all conditional probabilities that the call succeeds. using the parameter values shown in Table I for the generation of numerical results. since they permit a limited spatial reuse. The approach that is used to provide quality of service (QoS) guarantees to fast users consists in reserving some channels for their exclusive use. times the probability of requesting handovers. as regards resource planning in macrocells. where at call setup a channel was allocated. Finally.e. because no channel was available in the microcell. some channels are reserved to fast users. instead. we get: ½ The slow user call blocking probability. unless otherwise stated. the probability of this event is: ¯ È ´¿µ ´ µ ´½  È Ñ ´ µ µ ·½ À× ·½ È ´Ñµ ´½   È×´Å µ µ . the number of channels available in the considered macrocell is taken to be fixed. controlling the effect that channel reservation has on the performance of slow users. the probability of this event is: È ´½µ È ´Ñµ ´½   È×´Å µ µ ¯ the call completes in a microcell different from the one in which at call setup a channel was allocated. written as: × ½ × . for fast users: ½ TABLE I VALUES OF THE PARAMETERS FOR THE NUMERICAL RESULTS Parameter Æ ´Ñµ ½ Value 16 ½ ¼   ½ ´Å µ ´Å µ ´Ñµ × È success handovers È handovers ¼ × 0 ¼ ½ ¼ ¼ ℄ ¾ Denoting by À the probability that a fast user releases a channel because of handover. in the third case. the probability of this event is: È ´¾µ ´ µ ´½  È Ñ ´ µ µ ·½ À× ´½   À× µ the call completes in the macrocell. the most ‘valuable’ resources are macrocell channels. Resource planning in this case amounts to the identification of the minimum number of reserved channels necessary to guarantee the desired QoS to fast users. which are different for the two classes of users. and after successful handovers. we have: ´Å µ ´Å µ ´Ñµ ´Å µ s ½ À · ´Å µ Assuming independence for failures of the different handovers within the duration of a fast user call. with no channel reservations. after successful handovers. trying to guarantee the same QoS to both classes of users. Resource planning in this case must take into account that call blocking probabilities depend on mobility models. In this section we consider three different situations. we have to take into account that when a slow user attempts to access the system. careful resource planning in macrocells is a must. We recognize three different cases of successful completion of a slow user call: ¯ the call completes in the macrocell. In the first case. for example.. rather than call completion. Let À× be the probability that a slow user that succeeded in establishing a connection in a microcell. respectively). a hybrid approach is taken to resource planning. In the second case. For the computation of the call blocking probability for slow users. but at call setup a channel was allocated in a microcell. so that the total number of channels in the macrocell is minimized. R ESOURCE P LANNING In the context of dual-band GSM networks.

because no bound is guaranteed on their call blocking probability. Slow users. Consider for example the case Æ ´Å µ ¿¾ and ¼ ½. The plots in Fig.01 25 20 15 10 È ´Å µ ½      À ½ ½ (8) 5 0. Observe that the differences in Æ induced by varying the handover rates are not drastic. Æ .01. 4. Æ ´ µ ¿¾ Å .06 Fast user traffic. Suppose that the number of channels in the macrocell is fixed and equal to Æ ´Å µ .04 0. as ´Å µ increases. 4 show the values of blocking probability experienced by fast users when the number of reserved channels is as dictated by graph in Fig. In order to provide the desired QoS to fast users.It must be noted that all planning procedures that will be described in the following three subsections are based on iterations. the number of channels to be reserved to guarantee the desired QoS increases. Blocking probability for fast users when Æ is set to guarantee a specific QoS. In order to guarantee the desired QoS to fast users. and overall. 3. QoS guarantees to fast users In this situation. 7 report blocking probabilities for Fig. λ(M) 0. This fact must be controlled.1 Fig. 5 instead report the values of Æ required ¼ ½ for different values of the average numto guarantee ber of handovers per call. denoted by À ℄. In the plots we assume that the minimum with Æ ´Å µ admissible value for Æ is 7. Correspondingly. the average number of channels used by fast users increases too. and at every handover. N f 30 B=0. As already observed. the greater is the number of incoming handovers. but also the smaller is the time a user spends in the cell. We can plot curves that allow the identification of the minimum acceptable value for the number of channels to be reserved to fast users in the macrocell.1 In order to guarantee the desired QoS to fast users.10 B=0. A.02 0. 3 show the number of channels reserved to fast users at the macrocell necessary to guarantee equal to 0. The reason for this behavior is that the effective macrocell load does not change significantly with the handover rate: ´Å µ the higher is the handover rate . Æ increases also. fast users must be served according to specified QoS guarantees. Minimum values of Æ that guarantee a specified QoS. receive a best-effort kind of service. The plots in Fig.10 B=0. Æ channels are reserved for their exclusive use in the macrocell. by slow users. Given the value for . while the number of channels used by slow users decreases. 3.01 1e-04 0. 6 report the average numbers of channels used by fast users. Reserving channels to fast users to guarantee their QoS can deteriorate the QoS perceived by slow users. Bf 1e-01 1e-02 1e-03 B=0.08 0.02 0.04 0. The curves in Fig. and possibly avoided.08 0. which are expressed in terms of an upper bound on their call blocking probability. must be such that probability È is as specified in (8). λ(M) 0. from (6) we can obtain the maximum ´Å µ value that È can assume in order to guarantee : Number of reserved channels. Of course. the QoS corresponding to values of ¿¾. The plots in Fig.1 and 0. in order to reduce the dissatisfaction of slow users. the number ´Å µ of reserved channels. instead. The plots in Fig. the macrocell has to be designed in such a way that fast user call requests are refused with probability not larger than . and are thus made possible by the simplicity of the microcell and macrocell models that were previously described. as the macrocell traffic increases.06 Fast user traffic. both at call setup. Æ ´ ŵ ¿¾ 1e+00 Fast user blocking probability. so as to guarantee their QoS.

B. also on the microcell pa´Å µ and rameters. for slow users. Minimum values of Æ that guarantee Æ ´Å µ ¿¾ À ℄ fast and slow users. that show the blocking probabilities for fast and slow users.1 The plots in Fig. 9. Fair treatment of slow and fast users The goal of resource planning in this case is to guarantee the same QoS level to slow and fast users. and × .02 0.06 Fast user traffic. for fast and slow users.02 5 0.1e+00 30 E[Hf]=1 E[Hf]=5 E[Hf]=10 1e-01 Blocking probability Number of reserved channels. while from (7) it can be derived different.02 0. and.04 0. λ(M) 0. for high traffic the QoS perceived by slow users becomes unacceptable. This is reflected by the plots in Fig. 6. while the QoS guaranteed to fast users is actually provided.08 0. the average number of busy channels. Number of busy channels in the macrocell: total. Fig.06 Fast user traffic. While the values of Æ ´Å µ are just enough to guarantee the QoS to fast users. Fig. and total. The macrocell design is thus based on the selection of the number of channels Æ ´Å µ (with Æ ¼) such that for a given QoS level : × 30 total fast slow 25 Number of busy channels (9) 20 15 10 The call blocking probabilities and × depend on the user mobility pattern. 7. Clearly.1 ½ ½¼.04 0. For this reason. the values of probabilities È ´Å µ È× which guarantee that the bound on and × is met are ´Å µ is given by (8). 8 report the values of Æ ´Å µ that are neces¼ ½. È ´Å µ that È× must be such that: È×´Å µ 5 ´½   À× · È ´Ñµ À× È ´Ñµ µ 0 0.06 Fast user traffic.1 ¼½ 0. the average number of channels used by fast users grows much more slowly. We can see that both Æ ´Å µ and the average number of channels used by slow users grow sharply with the input traffic. N f 25 1e-02 20 15 1e-03 Bf Bs 10 1e-04 0. Blocking probabilities with Æ ´ ŵ ¿¾ and Æ set so that Fig.04 0. 5. λ(M) ¼ ½ for 0.08 0. are also shown in the same graph. Instead. avoiding channel reservations to favor fast users. λ(M) 0. for fast and slow users ¼½ with Æ ´ µ ¿¾ and Æ set so that Å .08 0. versus increasing values sary to guarantee a QoS level of ´Å µ .

and for those values of Æ ´Å µ the QoS perceived by slow users is much better than what needs to be guaranteed.04 0. 9 clearly show that the access of fast users to macrocell channels is the critical element in resource planning. Blocking probabilities with no channel reservation. Indeed. For this reason. for fast and slow users. 8 and Fig.5 t=1.05 Fig. 9. has to be chosen rather large. Ø 0. the total number of channels in the macrocell can be smaller than with no channel reservation. λ(M) ¼ 0. in order to guarantee the QoS experienced by fast users. Average number of busy channels in the macrocell: total. Up to about Blocking probability . λ(M) 0. for a fixed QoS guarantee . with no channel reservation.0 80 Number of channels 60 60 40 40 20 20 0 0. the number of channels in the macrocell. and Æ ´Å µ is minimized. Fair treatment with channel reservation 1e+00 1e-01 1e-02 1e-03 1e-04 1e-05 Bf Bs 1e-06 0. 8. and Æ ´ µ set so that ¼ ½ is guaranteed to both fast and slow users Å the service experienced by slow users is much better than what needs to be guaranteed. with no channel reservation. λ(M) 0. 10.0 t=1. aiming at a condition in which fast users and slow users perceive about the same QoS. Æ ´Å µ . where we consider difØ ´Å µ .04 0. ¯ Increase Æ until becomes close to × . Similar results are reported in Fig.02 0. The macrocell planning problem in this case thus consists in the determination of a pair of values ´Æ ´Å µ Æ µ.03 Fast user traffic. ferent values for the slow user traffic. 10. setting ´Ñµ with Ø ¼ ½ ¼ ½ ¾ ¼. 11 show the values of Æ ´Å µ (solid line) and Æ (dashed line) that result from the application of the above procedure. and Æ ´Å µ set so that ´Ñµ Ø ´Å µ . and ¼ ½ is guaranteed to both fast and slow users that Æ ´Å µ set so The numerical results in Fig. The plots in Fig. for increasing values of ´Å µ . C.03 Fast user traffic. ¯ Decrease Æ ´Å µ to the smallest values which still guarantees (9).02 0.01 0. For the solution of this problem it is possible to operate in the following way: ¯ Set Æ ¼ and find Æ ´Å µ such that condition (9) holds.5 t=2.05 Fig.04 0.02 0. such that condition (9) holds.100 100 N(M) tot fast slow 80 Number of channels t=0.03 Fast user traffic.01 0. a reasonable approach to resource planning consists in the reservation of some channels to fast users. By so doing.05 ¼ ½ is guaranteed with no channel reservation ½¼ ½ ¾¼ Fig.01 0 0.

In order to model systems where data connections overflow´Å µ ½). data connections are forced to use a smaller amount of bandwidth. In particular. ing to the macrocell are allocated just one channel (Æ we still use a multi-class model as before.Number of channels ¼ ¼¾ . and are provided with just one channel.05 ¼ ½ is guaranteed and Æ ´ Å µ is Ë ´ µ · Æ Æ ´Å µ The planning strategies presented before can be used also in this context. In order to simplify the system scenario. λ(M) set so that 0. or. and is the number of data connections. where Æ nels are provided to data connections at the macrocell. Comparing the curves of Fig. Æ ´ minimum Å µ and Æ 0. This is a remarkable improvement. For higher values of traffic. ´Å µ 100 N(M) Nf 80 60 40 VII. and voice calls with fast users.03 Fast user traffic. 8. can be used to cope with this system configuration.04 0. The plots in Fig. The microcell model remains the same as presented in Section III. Therefore. in order to meet their QoS requirements. Each data connection is assumed to require Æ ½ channels. 13. the data traffic load at the macrocell. The steady-state distribution of this model is as in (3). and the macrocell must serve also slow user calls. E XTENSION TO DATA C ONNECTIONS In this section we briefly describe how the proposed planning approach can be extended to deal with GSM networks where data connections requiring greater bandwidth than voice calls are provided to users with low mobility. but a further extension to the case in which slow and fast users can generate both voice and data connection requests is possible. the number of channels reserved to fast users in the macrocell. that shows the curves of Æ ´Å µ and Æ versus the macrocell traffic load. The macrocell state is again defined by the pair ´ µ where is the number of voice calls. and does not case Æ ´Å µ exhibit the fluctuating behavior of the case Æ . We can conclude that the allocation of just one channel to data connections that overflow to the macrocell seems quite interesting from the point of view of resource planning. and ei´Å µ ´Å µ ½ or Æ macrocell channels. × . instead of being is in this case Æ . is given by: × Ó Æ · ´Å µ Ë ´ µ · Æ Æ ´Å µ Æ Æ ´Å µ  Æ × where Æ is. The multi-class model of the macrocell behavior.01 Fig. 11. . we can observe that with this type of macrocell planning. 12 show that the QoS perceived by slow and fast users for high traffic values is about the same. the service provided to slow users by the resources in the microcells is so good that all channels in the macrocell can be reserved to fast users. If we assume that the average duration of data connections depends on the allocated bandwidth. presented in Section IV. as before. the data connection completion rate at the macrocell. 11 with those reported in Fig. Results refer to the planning strategy presented in section VI-C where fair treatment is provided to users adopting a channel reservation scheme. the service offered to slow users in the microcells deteriorates. The blocking probabilities for data connections (slow users) and voice calls (fast users) are: È×´Å µ ´ µ ¾Ë× ´ µ and È ´Å µ ´ µ ¾Ë ´ µ with Ë× ´ µ · Æ Æ ´Å µ   Æ Æ ´Å µ   Æ   Æ Æ Call blocking probabilities remain as in (5). The state space of the resulting model is: 20 0 0. We can observe that the ´Å µ ´Å µ ½ requires smaller values of Æ . ´Å µ Æ chanWe initially consider the first case. trying to save the valuable resources of the macrocell. provided that some changes are introduced. except for the fact that now the maximum number of ´Ñµ simultaneously active connections is Æ . are reported ther Æ in Fig. we associate data connections with slow users. The results generated by this planning procedure in the case of data connections using Æ microcell channels. Two alternatives are possible when data connections reach the macrocell: the data connection bandwidth requirements can either be satisfied with Æ channels. the Æ analysis of the microcell allows the derivation of the parameters of the overflow process of data calls towards the macrocell. the state space definition remains as in (4).02 0. the values of Æ ´Å µ are up to 30% smaller than in the case of no channel reservation.

R. V. pp. New York. Lo Cigno.” IEEE ICUPC’94.F. The simplicity of the approximate analytical models is instrumental for the accurate design and planning of the system parameters. [6] .B.” IEEE Trans. “Performance modeling and analysis of hierarchical wireless communications networks with overflow and take-back traffic. Massey. Beraldi. Meo. July 1998. Munich. Lin. October 1994. Chang. Noerpel.” IEEE Infocom’99. 3. N. C ONCLUSIONS We presented resource planning strategies for dual-band GSM networks. 1.Fang.S. Bolotin. 1987. Hong. Godlewski.03 Fast user traffic. N. I. Co. where service is offered to users moving over an area covered with overlapping macrocells and microcells. Ottawa. 1987.05 0.” ICC’95. so as to meet QoS constraints. λ(M) 0. “Traffic Model and Performance Analysis for Cellular Mobile Radio Telephone Systems with Prioritized and NonPrioritized Handoff Procedures. Lagrange.03 Fast user traffic. 1994. July 1995. M. X. Lin. North-Holland Publ. San Diego. “Modeling Hierachical Microcell/Macrocell PCS Architecture.F. Mastroianni. J. Architecture and Performance Issues”.” IEEE Communication Magazine. “Performance of a Reversible Hierarchical Cellular System. Marano. Vol. S.02 0. Zhou. pp. R. Y. Seattle. M. February 1997. W. W. 15. “A New Mobility Model and its Application in the Channel Holding Time Characterization in PCS Networks. Meo. Minimum value of Æ ´ µ . S. 102-107. Mastroianni. Vol. 8.” International Journal of Wireless Information Networks.” IEEE Journal on Selected Areas in Communications. n. Hu. P. NY. considering the provision of services with QoS guarantees expressed in terms of the probability of successful completion of connections. N. October 1999. S.02 0. Palm. S. pp. Rappaport. K.A. in this paper we have illustrated a few approaches for the optimization of the number of traffic channels to be activated within macrocells. A. Y.” IEEE MASCOTS’98. Resource planning is based on the iterative solution of simple approximate analytical models of the dynamics of microcells and macrocells. 89-99. F. Cox. Guerin.04 0. September 1994. Germany. Canada. Vol. A.” IEEE ICUPC’93. VT-35. Issue on Personal Communication – Services. Whitt. Intensity Variations in Telephone Traffic. M. Leonardi. As an example of the utilization of the approximate analytical models. Canada.” IEEE Transactions on Vehicular Technology. K. “Traffic Models for Wireless Com[16] [7] [8] [9] [10] [11] [12] [13] [14] [15] [2] [3] [4] [5] [17] [18] munication Networks. and corresponding value of Æ that guarantee ¼ ½ for voice calls and data connections requiring Æ ´ µ channels at the microcell and Æ ½ channels at the macrocell Å Å VIII. 2. 13. pp. September 1998. May 1997.Chlamtac. John Wiley & Sons. C. 12. Y. λ(M) 0. 4. 77-92. L. C. M. CA. “How Many Cells Should Be Considered to Accurately Predict the Performance of Cellular Networks?”. 35.R. 1406-1414. 1353-1364.” IEEE JSAC Spec. D. B. pp. Jabbari. June 1995. Leung. Marano. Ajmone Marsan. “Personal Communication Systems Using Multiple Hierarchical Cellular Overlays. 3. “Modeling of Customer Retrial Phenomenon in Cellular Mobile Networks. “Performance Analysis of a Hierarchical Cellular Mobile Communication System. Indoor and Mobile Radio Communications. WA. 1997. L. October 1997.P. Vol. N. R EFERENCES [1] D. Y. 12.100 1e+00 Nd(M)=1 Nd(M)=4 1e-01 Number of channels Bf Bs 80 Blocking probability 1e-02 60 1e-03 40 1e-04 20 1e-05 0 1e-06 0. A. Vol.05 Fig. R.04 0. R. 8. P. pp. “Channel Occupancy Time Distribution in a Cellular Radio System. Vol.” 47th IEEE Vehicular Technology Conference. Mastroianni. Tech. New York. De Carolis.01 0. S. N. Ajmone Marsan.. M. October 1993. “Modeling Techniques for Large-Scale PCS Networks.” 45th IEEE Vehicular Technology Conference. 433-438. Tran-Gia. “A Hierarchical Network Scheme for Multilayered Cellular Systems. G. March 1999. Veh. August 1986. 35. Marano. C.” IEEE JSAC. “Modeling Call Holding Time Distributions for CCS Network Design and Performance Analysis. Blocking probability and ¼ ½ is guaranteed and Æ ´ µ is minimum × Å are set so that Fig. Vol. Chang. Renewal Theory. European Wireless’99 and 4th ITG Mobile Communications. C. “Teletraffic analysis of a hierarchical cellular network. E. Rappaport. “Performance Analysis of Cellular Communication Networks Supporting Multimedia Services. 1962. Montreal. 12.” The Ninth IEEE International Symposium on Personal. N.01 when Æ Å µ and Æ ´ 0. Mandjes. Chin. 3.