10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

How user throughput depends on the traffic demand

in large cellular networks
Bartłomiej Błaszczyszyn∗ , Miodrag Jovanovic†∗ and Mohamed Kadhem Karray†

in function of the mean traffic demand locally (for each cell)

Abstract—We assume a space-time Poisson process of call and globally in the network (which involves appropriate spatial
arrivals on the infinite plane, independently marked by data averaging in conjunction with the temporal one, already present
volumes and served by a cellular network modeled by an infinite
ergodic point process of base stations. Each point of this point in the classical definition of the throughput) is a key engineering
process represents the location of a base station that applies a task in cellular communications. In this paper we propose an
processor sharing policy to serve users arriving in its vicinity, analytic approach to the evaluation of the mean user throughput
modeled by the Voronoi cell, possibly perturbed by some random in large irregular cellular networks, validated by real-network
signal propagation effects. User service rates depend on their measurements performed in operational networks.
signal-to-interference-and-noise ratios with respect to the serving
station. Little’s law allows to calculate the mean user throughput as
Little’s law allows to express the mean user throughput in any the ratio of the mean traffic demand (number of bits requested
region of this network model as the ratio of the mean traffic per unit of time) to the mean number of users in the steady state
demand to the steady-state mean number of users in this region. of the network. This argument can be used to express mean
Using ergodic arguments and the Palm theoretic formalism, we user throughput locally in any region of the network. Statistics
define a global mean user throughput in the cellular network and
prove that it is equal to the ratio of mean traffic demand to the usually collected in operational networks allow to estimate the
mean number of users in the steady state of the “typical cell” of mean traffic demand and the mean number of users for each cell
the network. Here, both means account for double averaging: over and hour of the day. Even if they carry important information
time and network geometry, and can be related to the per-surface about the local network performance, they exhibit important
traffic demand, base-station density and the spatial distribution of variability over time (24 hours) and network cells; cf. Figure 1.
the signal-to-interference-and-noise ratio. This latter accounts for
network irregularities, shadowing and cell dependence via some This can be explained by the fact that mean user throughput
cell-load equations. in a particular cell does not depend only on the traffic in
Inspired by the analysis of the typical cell, we propose also this cell, but also on the neighbouring cells. Moreover, the
a simpler, approximate, but fully analytic approach, called the geometry of different cells in a real network may significantly
mean cell approach. The key quantity explicitly calculated in
differ. For these two reasons, the family of local (established
this approach is the cell load. In analogy to the load factor of
the (classical) M/G/1 processor sharing queue, it characterizes for each cell) throughput-versus-traffic laws usually exhibits a
the stability condition, mean number of users and the mean lot of variability both in real data and in network simulations,
user throughput. We validate our approach comparing analytical and hence does not explain well the macroscopic (network-
and simulation results for Poisson network model to real-network level) relation between the mean traffic demand and mean user
measurements.
throughput. Finding such a macroscopic relation is an important
Index Terms—user-througput, traffic demand, cell-load, Little’s
law, cellular network, typical cell, point process, ergodicity, Palm task for network dimensioning. It is clear that an appropriate
theory, measurements spatial averaging is necessary to discover such a macroscopic
law.
I. I NTRODUCTION Spatial averages of point patterns, modeling in our case the
Mean user throughput is a key quality-of-service metric in geographic locations of base stations, can be studied using the
cellular data networks. It describes the average “speed” of data formalism of Palm distributions naturally related to the ergodic
transfer during a typical data connection. It is usually defined results for point processes. Within this setting one considers a
as the ratio of the average number of bits sent (or received) typical base station with its typical cell (zone of service) whose
per data request to the average duration of the data transfer. probabilistic characteristics correspond to the aforementioned
Since coexisting connections in a given network cell share some spatial averages of the characteristics for all base stations in
given cell transmission capacity, mean user throughput depends the network. Adopting this formalism, we define the mean
inherently on the requested data traffic. It also depends on user throughput in the infinite ergodic network as the limit
the network architecture (positioning of the base stations) and of the ratio of the mean volume of the data request to the
in fact may significantly vary across different network cells. mean service duration in a large, increasing to the whole plane,
Moreover, extra-cell interference makes performance of differ- network window. As the main result, we prove that such defined
ent cells interdependent. Predicting the mean user throughput (macroscopic) throughput characteristic is equal to the ratio of
∗ INRIA-ENS, 23 Avenue d’Italie, 75214 Paris, France Email: the mean traffic demand to the mean number of users in the
[email protected] typical cell of the network. Both these means account for double
† Orange Labs; 38/40 rue Général Leclerc, 92794 Issy-les-Moulineaux,
averaging: over time and network geometry.
France Email: {miodrag.jovanovic, mohamed.karray}@orange.com
This paper reports the results of a research undertaken under the contract A key element of the analysis of the cellular network is the
CRE No D09262 between Inria and France Télécom. spatial distribution of the signal-to-interference-and-noise ratio

978-3-901882-63-0/14/$31.00 ©2014 IEEE 611

 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

(SINR). We show how this distribution enters into the macro- A. Network model
scopic characterization of the throughput. When considering
SINR we are able to account for the fact that the base stations We consider locations {X1 , X2 , . . .} of base stations (BS)
that are idling, i.e., have no users to serve, do not contribute on the plane R2 as a realisation of a point process, which we
to the interference. This makes the performance of different denote by Φ. 1 We assume that Φ is stationary and ergodic
cells interdependent and we take it into account via a system with positive, finite intensity (mean number of BS per unit of
of cell-load equations. surface) λBS . 2
Finally, we show how to amend the model letting it account In order to simplify the presentation, we shall make first
for the shadowing in the path loss. The latter is known to the following two assumptions, which will be relaxed in Sec-
impact the geometry of the network, in the sense that the serving tions III-E and III-D, respectively.
base station is not necessarily the closest one. It also alters the 1) There is no shadowing. The (time-averaged over fading)
distribution of the SINR. propagation loss depends only on the distance r between
the emitter and the receiver through a path-loss function
A. Related work l(r), which we assume increasing.
2) Full interference. Each base station is always transmitting
The evaluation of user QoS metrics in cellular networks is a
at some fixed power P , common for all stations.
hard problem, but crucial for network operators and equipment
manufacturers. It motivates a lot of engineering and research We will also assume throughout the whole paper that each
studies. The complexity of this problem made many actors user is served by the BS which he or she receives with the
develop complex and time consuming simulation tools such as strongest signal power. The consequence of the assumption 1
those developed by the industrial contributors to 3GPP (3rd above is that each BS X ∈ Φ serves users in a geographic
Generation Partnership Project) [1]. There are many other zone V (X) = {y ∈ R2 : |y − X| ≤ minZ∈Φ |y − Z|} which is
simulation tools such as TelematicsLab LTE-Sim [2], University called Voronoi cell of X in Φ. Both the above assumptions 1
of Vien LTE simulator [3, 4] and LENA tool [5, 6] of CTTC, and 2 allow to represent the signal to interference and noise
which are not necessarily compliant with 3GPP. ratio (SINR) at location y ∈ V (X), X ∈ Φ as
A possible analytical approach to this problem is based on P/l (|y − X|)
the information theoretic characterization of the individual link SINR (y, Φ) := P , (1)
N +P Z∈Φ\{X} 1/l (|y − Z|)
performance; cf e.g. [7, 8], in conjunction with a queueing the-
oretic modeling and analysis of the user traffic; cf. e.g. [9–14]. where N is the noise power. Note that this represents the SINR
All these works consider some particular aspects of the network in the down-link (BS to user) channel. Note also, that we assume
and none of them considers a large, irregular multi-cell network. that each interfering base station always transmits (even if, for
Such a scenario is studied in our approach by using stochastic- example, it has no user to serve). This model, which we call
geometric tools combined with the two aforementioned theories. full interference model, will be improved in Section III-D.
As a result, we propose a global, macroscopic approach to We assume that the peak bit-rate at location y, defined as the
the evaluation of the user QoS metrics in cellular networks, bit-rate of a user located at y when served alone by its BS, is
which we compare and validate with respect to real network some function R (SINR), of the SINR. 3
measurements. We consider variable bit-rate (VBR) traffic; i.e., users arrive
Stochastic geometry has already been shown to give ana- to the network and require to transmit some volume of data at
lytically tractable models of cellular networks, see e.g. [15]. a bit-rate decided by the network. Specifically, we assume that
However, to the best of our knowledge, the prior works in this user channels are intra-cell orthogonal and inter-cell indepen-
context usually do not consider any dynamic user traffic. dent: if BS X serves n users located at y1 , y2 , . . . , yn ∈ V (X)
then the bit-rate of the user located at yj equals to 1/n th of its
B. Paper organization peak bit-rate n1 R (SINR (yj , Φ)), j ∈ {1, 2, . . . , n}. 4
We describe our general cellular network model, comprising The pattern of BS Φ does not evolve in time. We describe
the geometry of base stations, path loss model, space-time now the space-time process of user arrivals (and departures).
traffic demand process, and the service policy in Section II.
1 According to the formalism of the theory of point processes (cf e.g. [16]), a
In Section III we develop two approaches, called respectively P
point process is a random measure Φ = j δXj , where δx denotes the Dirac
typical and mean cell approach, allowing to study the depen- measure at x.
dence of the mean user throughput in the network and other 2 Stationarity means that the distribution of the process is translation invariant,

macroscopic network characteristics on the traffic demand and while ergodicity allows to interpret some mathematical expectations as spatial
averages of some network characteristics.
other model parameters. In Section IV we apply these ap- 3 The theoretical upper-bound of such function characterizing the link level
proaches studying some particular network model and compare performance is typically given by information theory. For example, in the
the obtained numerical results to real field measurements. case of AWGN channel, R (SINR) = W log2 (1 + SINR) where W is
the bandwidth. We shall assume that the fading is already averaged out by
considering the so-called ergodic
h capacity. For example,
i in the case of flat
II. M ODEL DESCRIPTION AND LOCAL fading, R (SINR) = W E log2 1 + |H|2 SINR where H is a random
CHARACTERISTICS variable representing the fading and E [·] is the expectation with respect to H.
4 This can be achieved using various multiple access schemes, e.g. time
We shall now describe our model. division.

612
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

B. Traffic demand Note that N (X) = ∞ if ρ(X) ≥ ρc (X).

Users arrive uniformly on the plane and require to transmit • The probability that the given BS is not idling in the steady
a random (arbitrarily distributed) volume of data of mean 1/µ state (has at least one user to serve) equals
bits. The duration between the arrivals of two successive users p (X) = min (θ (X) , 1) , (6)
in each geographic zone S of surface |S| is an exponential where θ(X), which we call cell load, is defined as
random variable of parameter λ × |S| . This means that on av- ρ (X)
erage there are λ arrivals per surface unit. The arrival locations, θ (X) := . (7)
ρc (X)
inter-arrival durations as well as the data volumes are assumed
independent across users. The above description corresponds to Note that the cell is stable if and only if θ(X) < 1 and
Z
a space-time Poisson process of arrivals, independently marked θ(X) = ρ 1/R (SINR (y, Φ)) dy . (8)
by data volumes. V (X)
We assume that the users don’t move considerably during Moreover,
their calls. 5 Each user stays in the system for the time θ(X)
necessary to download his data. This takes a random (service) N (X) = , (9)
1 − θ(X)
time because the bit-rate with which he is served depends r(X) = ρ(X)(1/θ(X) − 1) (10)
on the configuration of other users served by the same base
station. Users depart from the system immediately after having provided θ(X) < 1.
downloaded their data. The above expressions allow to express all other character-
The traffic demand per surface unit is then equal to ρ = µλ , istics in terms of the traffic demand per cell ρ(X) and the cell
which will be expressed in bit/s/km2 . The traffic demand in a load θ(X).
given cell equals Remark 1: All the above characteristics are local network
characteristics in the sense that they characterize the service
ρ (X) = ρ |V (X)| , X ∈ Φ. (2) at each BS X and vary over X ∈ Φ. Real data analysis
C. Local quality of service characteristics and simulations for Poisson network models exhibit a lot of
For a fixed configuration of BS Φ, the service of users arriv- variability among these characteristics. In particular, plotting
ing to the cell V (X) of a given BS X ∈ Φ can be modeled by an the mean user throughput r(X) as function of the mean traffic
appropriate (spatial) multi-class processor sharing queue, with demand ρ(X) for different X ∈ Φ does not reveal any apparent
classes corresponding to different peak bit-rates characterized systematic relation between these two local characteristics; cf.
by user locations y ∈ V (X). Note also that a consequence Figure 1.
of our model assumptions (in particular the full interference III. G LOBAL NETWORK CHARACTERISTICS
assumption 2, inter-cell channel independence and space-time
In this section we propose some global characteristics of the
Poisson arrivals) the service processes of different queues are
network allowing to characterize its macroscopic performance.
independent.
We are particularly interested in finding such a relation between
We consider now the steady-state of users served in each cell
the (per surface) traffic demand ρ and the (global) mean user
V (X). 6 The following expressions follow from the queueing-
throughput in the network, with this latter characteristic yet to
theoretic analysis of the processor sharing systems of each BS
be properly defined.
X ∈ Φ, cf [10, 14] for the details.
• The service process of BS X ∈ Φ is stable if and only if A. Typical cell of the network
its traffic demand does not exceed the critical value that is A first, natural idea in this regard is to consider spatial
the harmonic mean of the peak bit-rate over the cell: averages of the local characteristics in an increasing network
|V (X)| window A, say a ball centered at the origin and the radius
ρc (X) := R . (3)
V (X)
1/R (SINR (y, Φ)) dy increasing to infinity. Assuming ergodicity of the point process
Note that ρc (X) depends on X and on Φ. The same Φ of the BS, these averages can be expressed and calculated
observation is valid for the subsequent cell characteristics. as Palm-expectations of the respective characteristics of the so
• The mean user throughput in the given cell, defined as the called “typical cell” V (0). For example
X
ratio of the mean volume of the data request 1/µ to the lim 1/Φ(A) ρ(X) = E0 [ρ(0)] = ρE0 [|V (0)|] . (11)
|A|→∞
average service time of users in this cell, can be expressed X∈A
as follows The typical cell V (0) is the cell of the BS located at the origin
r (X) = max(ρc (X) − ρ (X) , 0) . (4) X = 0 and being part of the network Φ distributed according
to the Palm distribution P0 associated to the original stationary
• The mean number of users in the steady state of the given
distribution P of Φ. In the case of Poisson process, the relation
cell equals to
ρ (X) between the Palm and stationary distribution is particularly
N (X) = . (5) simple and (according to Slivnyak’s theorem) consists just in
r (X)
adding the point X = 0 to the stationary pattern Φ.
5 Small user movements are reflected in channel fading; cf. the remark in
The convergence analogue to (11) holds for each of the
footnote 3.
6 Note that the (mean) QoS characteristics of users in this state correspond previously considered local characteristics E0 [ρc (0)], E0 [r(0)],
to time-averages of user characteristics. E0 [N (0)], E0 [p(0)] and E0 [θ(0)]. The convergence is P almost

613
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

sure and follows from the ergodic theorem for point processes Note that the stationarity of Φ implies the same for the random
(see [17, Theorem 4.2.1], [16, Theorem 13.4.III]). However, as set S. We denote by πS = E[1(0 ∈ S)] the volume fraction
we will explain in what follows, not all of these mean-typical of S and call it the stable fraction of the network. It is equal
cell characteristics have natural interpretations as macroscopic to the average fraction of the plane covered by the stable cells;
network characteristics. cf. [17, Definition 3.4 and the subsequent Remark]. Denote also
First, note that the existence of some (even arbitrarily small) N 0 := E0 [N (0)1(N (0) < ∞)] . (14)
fraction of BS X which are not stable (with ρ(X) ≥ ρc (X),
hence N (X) = ∞) makes E0 [N (0)] = ∞. We are ready now to define the mean user throughput in the
Remark 2: For a well dimensioned network one does not network r0 as the ratio of the average number of bits per data
expect unstable cells. For a perfect hexagonal network model request to the average duration of the data transfer in the stable
Φ all cells are stable or unstable depending on the value of part of the network
the per-surface traffic demand ρ. An artifact of an infinite, 1/µ
r0 := lim . (15)
homogeneous, Poisson model Φ is that for arbitrarily small ρ |A|→∞ (temporal-)mean service time in A ∩ S
there exists a non-zero fraction of BS X ∈ Φ, which are non- Here is the key result of the typical cell approach. Its proof is
stable. This fraction is very small for reasonable ρ, allowing to given at the end of this section.
use Poisson to study QoS metrics which, unlike E0 [N (0)], are Theorem 5: For ergodic network Φ we have
not “sensitive” to this artifact. ρ πS
r0 = . (16)
We will also show in the next section that it is not natural λBS N 0
to interpret E0 [r(0)] (which is not sensitive to the existence of Remark 6: Equation (16) provides a macroscopic relation
a small fraction of unstable cells) as the mean user throughput between the traffic demand and the mean user throughput in
in the network; see Remark 7. Before we give an alternative the network, which we are primarily looking for in this paper.
definition of this latter QoS, let us state the following result, It will be validated by comparison to real data measurements.
which will be useful in what follows. Quantities N 0 and πS do not have explicit analytic expressions
Proposition 3: We have analogous to (13). Nevertheless they can be estimated from
ρ simulations of a given network model Φ. Note that these are
E0 [ρ(0)] = , (12)
λBS static simulations of the network model. No simulation of the
ρ
E0 [θ(0)] = E[1/R (SINR (0, Φ))] . (13) traffic demand process is necessary, which greatly simplifies
λBS the task. For small and moderate values of the traffic demand
Proof: The first equation is quite intuitive: the average cell (observed in real networks) one obtains πS ' 1. Moreover, in
surface is equal to the inverse of the average number of BS per Section III-C we will propose some more explicit approxima-
unit of surface. Formally, both equations follow from the inverse tion of N 0 .
formula of Palm calculus [17, Theorem 4.2.1]. In particular, Remark 7: Assume that there are no unstable cells in the
for (13) one uses representation (8) in conjunction with the network. This is the case e.g. for lattice (say hexagonal) network
inverse formula. models with traffic demand ρ < ρc (X) = ρc , where the value
Remark 4: Note that the expectation in the right-hand-side of the critical traffic is the same for all cells. Then πS = 1,
of (13) is taken with respect to the stationary distribution N 0 = E0 [N (0)] and the relation (16) takes form
of the BS process Φ. It corresponds to the spatial average ρ E0 [ρ(0)]
of the inverse of the peak bit-rate calculated throughout the r0 = 0
= 0 . (17)
λBS E [N (0)] E [N (0)]
network. The (only) random variable in this expression is the
SINR experienced by the typical user. This distribution is Thus, in general r0 6= E0 [r(0)] = E0 [ρ(0)/N (0)]. We want to
usually known in operational networks (estimated from user emphasize that this is not merely a theoretical detail resulting
measurements). It can be also well approximated using Poisson from our (and common) definition of the mean throughput (15).
network model for which its distribution function admits an The expression E0 [r(0)] = E0 [ρ(0)/N (0)], which in principle
explicit expression; cf [18, 19]. can be considered as another global QoS metric, is in practice
difficult to estimate. Indeed, when estimating E0 [r(0)] as the
B. Mean user throughput in the network average of the ratio “traffic demand to the number of users”
from real data measurements, one needs to give a special
Faithful to the usual definition of the mean user throughput treatment to observations which correspond to cells during
as the ratio of the mean volume of the data request to the mean their idling hours (i.e., with no user, and such observations are
service duration (which we retained at the local, cell level) we not rare in operational networks). Neither skipping nor literal
aim to define now the mean user throughput in the (whole) acceptance of these observations captures the right dependence
network as the ratio of these two quantities taken for increasing of the mean user throughput on the traffic demand.
network window A. However in order to “filter out” the impact Proof of Theorem 5: By Little’s law [20, eq. (3.1.14)] the
of cells which are not stable and avoid undesired degeneration temporal mean service time T W of users in any region of the
of this characteristic (e.g. for Poisson process; cf. Remark 2) network W , saySthe union of stable cells with BS in some
let us consider the union of all stable cells region A, W = X∈A∩S V (X), is related to the mean number
[
S := V (X) . N W of the users served in this region W in the steady state
X∈Φ:ρ(X)<ρc (X) by the equation N W = λ|W |T W . Consequently, the mean user

614
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

throughput in this region W can be expressed as 1/(µT W ) = real data measurements. We consider it as an approximation
ρ|W |/N W . Using of (16). It consists in assuming N̄ ' N 0 /πS . This latter
|W | |W | hypothesis will be also separately validated numerically.
W
=P (18) Remark 10: Note that the key characteristic of the mean cell
N X∈A∩S N (X)
P is its load θ̄. In analogy to the load factor of the (classical)
|V (X)|1(N (X) < ∞) |A| M/G/1 processor sharing queue, it characterizes the stability
= X∈A P
|A| X∈A∩S N (X) condition, mean number of users and the mean user throughput.
(19)
D. Cell-load equations
and again the ergodic theorem for point yprocess Φ, we
obtain the that limit in (15) is P-almost surely equal to We have to revoke now the full interference assumption 2
ρE0 [|V (0)|1(N (0) < ∞)]/E0 [N (0)1(N (0) < ∞)]. By the made in Section II-A. An amendment is necessary in this matter
aforementioned inverse formula of Palm calculus we conclude for the model to be able to predict the real network data; cf
E0 [|V (0)|1(N (0) < ∞)] = E[1(0 ∈ S)]/λBS . numerical examples in Section IV. Recall that the consequence
of this assumption is that in the expression (1) of the SINR all
C. Mean cell the interfering BS are always transmitting at a given power P .
It is tempting to look for a synthetic model which would In real networks BS transmit only when they serve at least one
allow to relate main parameters and QoS metrics of a large user. 7 Taking this fact into account in an exact way requires
irregular cellular network in a simple, yet not simplistic way. introducing in the denominator of (1) the indicators that a
The typical cell approach described up to now offers such given station Z ∈ Φ at a given time is not idling. This, in
possibility. In this section we will go a little bit further and consequence, would lead to the probabilistic dependence of the
propose an even simpler model. It consists in considering a service process at different cells and result in a non-tractable
virtual cell, to which we will assign the parameters and QoS model. In particular, we are not aware of any result regarding
metrics inspired by the analysis of the typical cell. In contrast to the stability of such a family of dependent queues. For this
the typical cell, our virtual cell is not random and this is why we reason, we take into account whether Z is idling or not in a
call it the mean cell. Specifically, we define it as a (virtual) cell simpler way, multiplying its powers P by the probability p(Z)
having the same traffic demand ρ̄ and load θ̄ as the typical cell. that it is not idle in the steady state. In other words we modify
Note that these two characteristics admit explicit expressions; the expression of the SINR as follows
cf. Proposition 3 and Remark 4. P/l (|y − X|)
ρ SINR (y, Φ) := P , (25)
ρ̄ := E0 [ρ (0)] = , (20) N +P p (Z) /l (|y − Z|)
λBS Z∈Φ\{X}
ρ
θ̄ := E0 [θ (0)] = E[1/R (SINR (0, Φ))] . (21) for y ∈ V (X), X ∈ Φ where p(Z) are cell non-idling
λBS
probabilities given by (6). We will see in Section IV that this
For the remaining characteristics, we assume that they are
model, called (load-)weighted interference model, fits better to
related to the above two via the relations presented in Sec-
real field measurements than the full interference model. The
tion II-C. Specifically, following (2) we define the surface of
above modification of the model preserves the independence
the mean cell by V̄ = ρ̄/ρ and in analogy to (7) we define the
of the processor-sharing queues at different cells given the
critical load of the mean cell as
ρ̄ realization Φ of the network (thus allowing for the explicit
ρ̄c := . (22) analysis of Section II-C). However the cell loads θ(X) are no
θ̄
We say that the mean cell is stable if ρ̄ < ρ̄c . Inspired by (4) longer functions of the traffic demand and the SINR experienced
we define the user’s throughput in the mean cell by in the respective cells, but are related to each other by the
following equations that replace (8)
r̄ := max (ρ̄c − ρ̄, 0) Z
1
and, as in (5), the mean number of users in the mean cell is θ (X) = ρ ! dy .
V (X)
defined as R P P/l(|y−X|)
ρ̄ N +P min(θ(Z),1)/l(|y−Z|)
N̄ := . Z∈Φ\{X}
r̄ (26)
We observe the following immediate relations. We call this system of equations in the unknown cell loads
Corollary 8: The mean cell is stable if and only if θ̄ < 1. In {θ (X)}X∈Φ the cell-load equations.
this case Remark 11: [Spatial stability] The weighted interference
θ̄
N̄ = , (23) model introduces more “spatial” dependence between the pro-
1 − θ̄ cessor sharing queues of different cells, while preserving their
r̄ = ρ̄(1/θ̄ − 1) , (24) “temporal” (conditionally, given Φ) independence. A natural
which are analogous to (9) and (10), respectively. question regarding the existence and uniqueness of the solution
Remark 9: The equation (24) provides an alternative macro- of the fixed point problem (26) arises. Note that the mapping
scopic relation between the traffic demand and the mean user in the right-hand-side of (26) is increasing in all θ(Z), Z ∈ Φ
throughput in the network. It is purely analytic; no simulations provided function R is increasing. Using this property it is
are required provided one knows the distribution of the SINR of 7 Analysis of more sophisticated power control schemes is beyond the scope
the typical user in (21). It will be validated by comparison to of this paper.

615
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

easy to see that successive iterations of this mapping started we consider the full interference or the weighted interference
off θ(Z) ≡ 0 on one hand side and off θ(Z) as in (8) (full model, for y ∈ V (X), with V (X) defined by (28). The same
interference model) on the other side, converge to a minimal modification regards the cell-load equations (26) and (27).
and maximal solution of (26), respectively. An interesting All the previous results involving the typical cell remain valid
theoretical question regards the uniqueness of the solution for this modification of the model. In particular, the results of
of (26), in particular for a random, say Poisson, point process Proposition 3 can be extended to the model with shadowing
Φ. Answering this question, which we call “spatial stability” (where the cell associated to each base station is not necessarily
of the model, is unfortunately beyond the scope of this paper.8 the Voronoi cell) provided the origin 0 belongs to a unique cell
The simulation study of the typical cell model, presented in almost surely.
Section IV (where we use Matlab to find a solution of (26) for Note that the mean cell surface E0 [|V (0)|] = 1/λBS , and
any given finite pattern of base stations Φ) is less stable for hence the mean traffic demand per cell E0 [ρ(0)] = ρ/λBS ,
larger values of the traffic demand ρ. do not depend at all on the shadowing. The values of other
In the mean cell approach (cf Section III-C) we take into characteristics of the typical and the mean cell will change
account the weighted interference model by the following depending on the distribution of the random shadowing field
(single) equation in the mean-cell load θ̄ SX (y) (both the marginal distributions and the correlation
across y). An interesting remark in this regard is as follows.
" !#
ρ P/l (|X ∗ |)
θ̄ = E 1/R P , In the full interference model, the mean load of the typical
λBS N + P Z∈Φ\{X ∗ } θ̄/l (|y − Z|)
cell and the load of the mean cell (which are by the definition
(27) equal E0 [θ(0)] = θ̄) depend only on the stationary marginal
∗
where X is the location of the BS whose cell covers the origin. distribution of SINR(0, Φ), cf. (21). Hence, it does not depend
We solve the above equation with θ̄ as unknown. We will show on the (spatial) correlation of SX (y) across y. Moreover, this
in the numerical section that the solution of this equation gives a distribution is known in the case of the Poisson network and
good estimate of the empirical average of the loads {θ (X)}X∈Φ identically distributed marginal shadowing SX (y) ∼ S. As
obtained by solving the system of cell-load equations (26) for explained in [19], in this case SINR(0, Φ) in the model with
the simulated model. shadowing has the same distribution as in the model without
Remark 12: [Pilot channel] The cells which are not idle shadowing and the density of stations equal to λBS × E[S 2/β ].
might still emit some power (e.g. in the pilot channel). This In particular, a specific distribution of S and the correlation of
can be taken into account by replacing p(Z) = min(θ(Z), 1) SX (y) across y play no role. This equivalence of the two models
in (26) by p(Z)(1 − ) + , where  is the fraction of the power (with and without shadowing) is more general, as explained
emitted all the time. Similar modification concerns θ̄ in the in [19, 22], and applies also to the mean-cell-load equation (27).
right-hand-side of (27). Remark 13: The impact of the shadowing on the mean cell
model, both in full and weighted interference scenario in the
E. Shadowing above Poisson model with SX (y) ∼ S can be summarized as
Until now we were assuming that the propagation loss is follows. It modifies only the cell load θ̄ and not the traffic
only induced by the distance between the transmitter and the demand ρ̄. Moreover, multiplying λBS by E[S 2/β ] and dividing
receiver. In this section we will briefly explain how the effect ρ by the same moment one obtains an equivalent (in terms of
of shadowing can be taken into account. all considered characteristics) mean cell without shadowing.
Assume that the shadowing between a given station X ∈ Φ
and all locations y ∈ R2 is modeled by some random field IV. N UMERICAL RESULTS
SX (y − X). That is, we assume the propagation loss between To illustrate the motivation of this work, we present first on
X and y LX (y) = Sl(|y−X|) X (y−X)
. We assume that, given Φ, Figure 1 non-averaged data obtained from the measurements
the random fields SX (·) are independent across X ∈ Φ and performed in an operational network in some zone of some big
identically distributed. In general, we do not need to assume city in Europe. 9
any particular distribution for SX (·) (neither independence nor Different points in this figure correspond to the measurements
the same distribution of SX (y) across y). of the traffic demand and the estimation of the user throughput
The assumption that each user is served by the BS received made by different cells during different hours of the day. No
with the smallest path-loss results in the following modification apparent relation between these two quantities can be observed
of the geographic service zone of X, which we keep calling in this way.
“cell” In order to uderstand and predict the performance of the net-
V (X) = {y ∈ R2 : LX (y) ≤ min LZ (y)} . (28) work for which we have presented the above data, we will now
Z∈Φ 9 More precisely, a dense urban network zone consisting of 382 base
For mathematical consistence we shall assume that, almost stations was selected in a big European city, whose locations loosely satisfy
surely, the origin belongs to a unique cell (i.e., is not located homogeneous spatial Poisson assumption. Ripley’s L-function, cf [23, page 50],
plotted on Figure 9, was used to verify this latter assumption. The density of
on any cell boundary). base stations in this dense urban zone is about 4.62 base stations per km2 . Later,
The SINR at location y can be expressed by (1) or (25) with we will consider also a urban zone of a different European city, where the spatial
l(|y − Y |) replaced by LY (y) for Y ∈ Φ, depending whether homogeneous Poissonianity of the base station locations can also be retained;
cf. Figure 9, with roughly four times smaller density of base stations, more
8 Existence and uniqueness of the solution of a very similar problem (with precisely 1.15 stations per km2 . In both cases the network operates HSDPA
finite number of stations and a discrete traffic demand) is proved in [21]. system with MMSE coding.

616
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

12000
Throughput versus traffic per cell Figure 3 shows the mean number of users per cell in the
stable part of the network N 0 /πS (obtained from simulations)
10000
and the analytically calculated number of users in the mean cell
User throughput [kbps]

N̄ versus mean traffic demand per cell. We have two remarks.

8000
For the traffic demand smaller than 500kbps per cell (for which
all the simulated cells are stable; πS = 1, cf. Figure 2), both
6000
models predict the same mean number of users per cell. Beyond
4000
this value of the traffic demand per cell the estimators of the
number of users in the typical cell become not accurate due very
2000 rapidly increasing fraction of the unstable region. (Error bars
on all figures represent the standard deviation in the averaging
0 over 30 realizations of the Poisson network).
0 500 1000 1500 2000 2500 3000 3500 4000
Traffic demand per cell [kbps] Finally, Figure 4 presents the dependence of the mean user
throughput in the network on the mean traffic demand per cell
Fig. 1: Local user throughput versus local traffic demand for some zone (selected to satisfy
a spatial homogeneity of the base stations) of an operational cellular network deployed obtained using the two approaches: r0 and for the typical cell
in a big city in Europe. 9288 different points correspond to the measurements made by and r̄ for the mean cell. Again, both models predict the same
different sectors of different base stations during 24 different hours of some given day.
performance up to roughly 500kbps.
specify correspondingly our general model and study it using
the proposed approach. The obtained results will be compared C. Weighted interference
to the appropriately averaged real field measurements.
We consider now the load-weighted interference model tak-
A. Model specification ing into account idling cells. We see in Figures 5, 6 and 7
Consider the following numerical setup. Assume Poisson that the consequence of this (more realistic) assumption is that
process of BS with intenisty λBS = 4.62km−2 (which cor- the cell loads are smaller, a larger fraction πS of the network
responds to an average distance between two neighbouring BS remains stable, and the two approaches (by the typical cell and
of 0.5km). We assume the path-loss function l(r) = (Kr)β , by the mean cell) predict similar values of the QoS metrics
with K = 7117km−1 , and the path loss exponent β = 3.8. The up to a larger value of the traffic demand per cell, roughly
propagation model comprises the log-normal shadowing with 700kbps. Note that it is in this region that the real network
the logarithmic standard deviation 10dB; cf [18], and the mean operates for which we present the measurements, and that its
spatial correlation distance 0.05km. performance coincides with the performance metrics calculated
The transmision power is P = 58dBm, with the fraction  = using the typical and mean cell approach. More precisely, the
10% used in the pilot channel. The antenna pattern is described field measurements on Figures 5, 6 and 7 correspond to the
in [1, Table A.2.1.1-2]. The noise power is N = −96dBm. same day and network zone considered on Figure 1.
We assume the peak bit-rate equal to 30% of the ergodic Remark 14: [Measurement methodology] Measurement
capacity of the AWGN channel; cf. Footnote 3, with the points on Figure 5 show the fraction of time, within a given
frequency bandwidth W = 5MHz and the Rayleigh fading with hour, when the considered base stations were idle, averaged
mean power E[|H|2 ] = 1. over the base stations, as function of the average traffic
Estimations of the typical cell are performed by the simu- demand during this hour. Similarly, measurement points on
lation of 30 realizations of the Poisson model within a finite Figure 6 show the spatial average of the mean number of users
observation window, which is taken to be the disc of radius reported by the considered base stations within a given hour,
2.63km. We first average over all BS in this window and then as function of the average traffic demand during this hour.
over the model realizations. The empirical standard deviation Finally, measurements on Figure 7 give the ratio of the total
form the obtained averages will be presented via error-bars. number of bits transmitted by all the base stations during a
We shall study now our model using the typical and mean given hour, to the total number of users they served during
cell approach, assuming first the full interference model and this hour in function of the average traffic demand during this
then the weighted one. hour.
Remark that Figure 7 makes evident a macroscopic relation
B. Full interference between the traffic demand and the mean user throughput in the
We consider first the full interference model (i.e. all BS emit network zone already considered on Figure 1. This relation, we
the signal all the time, regardless of whether or not they serve are primarily looking for in this paper, is not visible without
users). Figure 2 shows the mean cell load of the typical cell the spatial averaging of the network measurements described in
E0 [θ(0)] and the stable fraction of the network πS obtained Remark 14. In order to ensure the reader that a relatively good
from simulations, as well as the analytically calculated load of matching between the measurements and the analytic prediction
the mean cell θ̄, versus mean traffic demand per cell ρ/λBS . We is not a coincidence we present on Figure 8 similar results for
confirm that the typical cell and the mean cell models have the a urban zone of a different European city, where the spatial
same load. Note that for the traffic demand up to 500kbps per homogeneous Poissonianity of the base station locations can
cell we do not observe unstable cells in our simulation window also be retained; cf. Figure 9. The only engineering difference
(πS = 1). of this network zone with respect to the previously considered

617
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

Stable fraction 1 1

Stable fraction
0.8 0.8

0.6 0.6

0.4 0.4
Cell load

Cell load
0.2 0.2
Typical cell load
Typical cell load Stable fraction
Stable fraction Mean cell
Mean cell load Field measurements
0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Traffic demand per cell [kbps] Traffic demand per cell [kbps]
Fig. 2: Cell load and the stable fraction of the network versus traffic demand per cell in Fig. 5: Load and the stable fraction of the network versus traffic demand in the weighted
the full interference model. interference model. Also, load estimated from real field measurements.

20 10
Typical cell
Mean cell
Typical cell
8 Mean cell
Field measurements
15
Number of users

Number of users
6

10

5
2

0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Traffic demand per cell [kbps] Traffic demand per cell [kbps]
Fig. 3: Number of users per cell versus traffic demand per cell in the full interference Fig. 6: Number of users versus traffic demand per cell in the weighted interference model.
model. Also, the same characteristic estimated from the real field measurements.

1400 5000
Typical cell Typical cell
Mean cell Mean cell
4500 Field measurements
1200
4000
User throughput [kbps]

User throughput [kbps]

1000 3500

3000
800
2500
600
2000

400 1500

1000
200
500

0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Traffic demand per cell [kbps] Traffic demand per cell [kbps]
Fig. 4: Mean user throughput in the network versus traffic demand per cell in the full Fig. 7: Mean user throughput in the network versus traffic demand per cell in the
interference model. weighted interference model. Also, the same characteristic estimated from the real field
measurements.

dense urban zone is roughly four times smaller density of base

stations, more precisely 1.15 stations per km2 .
Remark 15: [Day and night hours] Let us make a final mean curve, indicated on Figure 8 and visible on both presented
remark regarding the empirical relation between the mean user plots of the throughput, seems to be a more general rule,
throughput on the mean traffic demand revealed on Figures 7 which escapes from the analysis presented in this paper and
and 8. Recall that different points on these plots correspond to remains an open question. A possible explanation can lay in a
different hours of some given day. In fact, the points laying different space-time structure of the traffic during the day and
below the mean curve correspond to day hours while the points night, with the former one being much more clustered (fewer
laying above the mean curve correspond to night hours. This users, requesting larger volumes, generating less interference
“circulation” of the measured values around the theoretical and overhead traffic).

618
 10th International Workshop on Spatial Stochastic Models for Wireless Networks 2014

R EFERENCES
[1] 3GPP, “TR 36.814-V900 Further advancements for E-UTRA - Physical
Layer Aspects,” in 3GPP Ftp Server, 2010.
[2] G. Piro, L. A. Grieco, G. Boggia, F. Capozzi, and P. Camarda, “Simulating
LTE cellular systems: an open source framework,” IEEE Trans. Veh.
Technol., vol. 60, pp. 498–513, 2011.
night [3] C. Mehlführer, J. C. Ikuno, M. Simko, S. Schwarz, M. Wrulich, and
M. Rupp, “The Vienna LTE simulators - Enabling reproducibility in
wireless communications research,” EURASIP Journal on Advances in
Signal Processing, vol. 2011:29, pp. 1–14, 2011.
day [4] M. Simko, Q. Wang, and M. Rupp, “Optimal pilot symbol power
allocation under time-variant channels,” EURASIP Journal on Wireless
Communications and Networking, vol. 2012:225, no. 25, pp. 1–11, 2012.
[5] N. Baldo, M. Miozzo, M. Requena, and J. N. Guerrero, “An open source
product-oriented LTE network simulator based on ns-3,” in Proc. of
MSWIM, 2011.
[6] N. Baldo, M. Requena, J. Nin, and M. Miozzo, “A new model for the
simulation of the LTE-EPC data plane,” in Proc. of WNS3, 2012.
Fig. 8: Mean user throughput in the network versus traffic demand per surface for an urban [7] A. J. Goldsmith and S.-G. Chua, “Variable-rate variable-power MQAM
zone of a big city in Europe. (The density of base stations is 4 times smaller than in the for fading channels,” IEEE Trans. Commun., vol. 45, pp. 1218–1230,
dense urban zone considered on Figure 7). 1997.
[8] P. E. Mogensen, W. Na, I. Z. Kovács, F. Frederiksen, A. Pokhariyal,
K. I. Pedersen, T. E. Kolding, K. Hugl, and M. Kuusela, “LTE Capacity
Compared to the Shannon Bound,” in Proc. of VTC Spring, 2007, pp.
1234–1238.
[9] S. Borst, “User-level performance of channel-aware scheduling algorithms
in wireless data networks,” in Proc. of IEEE Infocom, 2003.
[10] T. Bonald and A. Proutière, “Wireless downlink data channels: user
performance and cell dimensioning,” in Proc. of Mobicom, Sept. 2003.
[11] N. Hegde and E. Altman, “Capacity of multiservice WCDMA Networks
with variable GoS,” in Proc. of IEEE WCNC, 2003.
[12] T. Bonald, S. C. Borst, N. Hegde, M. Jonckheere, and A. Proutière, “Flow-
level performance and capacity of wireless networks with user mobility,”
Queueing Systems, vol. 63, no. 1-4, pp. 131–164, 2009.
[13] L. Rong, S. E. Elayoubi, and O. B. Haddada, “Performance evaluation
of cellular networks offering TV services,” IEEE Trans. Veh. Technol.,
vol. 60, no. 2, pp. 644 –655, feb. 2011.
[14] M. K. Karray and M. Jovanovic, “A queueing theoretic approach to the
dimensioning of wireless cellular networks serving variable bit-rate calls,”
Fig. 9: Ripley’s L-function calculated for the considered dense urban and urban network IEEE Trans. Veh. Technol., vol. 62, no. 6, July 2013.
zones. (L function is the square root of the sample-based estimator of the expected number [15] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage
of neigbours of the typical point within a given distance, normalized by mean number of
points in the disk of the same radius. Slinvyak’s theorem allows to calculate the theoretical
and rate in cellular networks,” IEEE Trans. Commun., vol. 59, no. 11, pp.
value of this function for a homogeneous Poisson process, which is L(r) = r.) In fact, in 3122 –3134, november 2011.
large cities spatial, homogeneous “Poissonianity” of base-station locations is often satisfied [16] D. J. Daley and D. VereJones, An introduction to the theory of point
“per zone” (city center, residential zone, suburbs, etc.). Moreover, log-normal shadowing processes. Volume I, 2nd ed. New York: Springer, 2003.
further justifies Poisson assumption, cf. [18, 24]. [17] F. Baccelli and B. Błaszczyszyn, Stochastic Geometry and Wireless Net-
works, Volume I — Theory, ser. Foundations and Trends in Networking.
V. C ONCLUSION NoW Publishers, 2009, vol. 3, No 3–4.
[18] B. Błaszczyszyn, M. K. Karray, and H. P. Keeler, “Using Poisson
In order to evaluate user’s QoS metrics, in particular the mean processes to model lattice cellular networks,” in Proc. of IEEE Infocom,
2013.
user throughput, in large irregular multi-cellular networks, two [19] H. P. Keeler, B. Błaszczyszyn, and M. K. Karray, “SINR-based k-coverage
approaches based on stochastic geometry in conjunction with probability in cellular networks with arbitrary shadowing,” in Proc. of
queueing and information theory are developed. The typical IEEE ISIT, 2013.
[20] F. Baccelli and P. Brémaud, Elements of queueing theory: Palm Martin-
cell approach consists in considering true spatial averages of gale calculus and stochastic recurrences. Springer, 2003, vol. 26.
local network characteristics and thus capturing the global [21] I. Siomina and D. Yuan, “Analysis of cell load coupling for lte network
network performance. A simpler, approximate but fully analytic planning and optimization,” Wireless Communications, IEEE Transactions
on, vol. 11, no. 6, pp. 2287–2297, 2012.
approach, called the mean cell approach, is inspired by the anal- [22] B. Błaszczyszyn and H. P. Keeler, “Equivalence and comparison of
ysis of the typical cell. The key quantity explicitly calculated in heterogeneous cellular networks,” in Proc. of PIMRC’13 – WDN-CN2013,
this approach is the cell load. In analogy to the load factor of 2013.
[23] D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic geometry and its
the (classical) M/G/1 processor sharing queue, it characterizes applications. Wiley Chichester, 1995.
the stability condition, mean number of users and the mean [24] T. X. Brown, “Cellular performance bounds via shotgun cellular systems,”
user throughput. Considering some real field measurement, we Selected Areas in Communications, IEEE Journal on, vol. 18, no. 11, pp.
2443–2455, 2000.
validate the proposed approach by showing that it allows to
predict the performance of a real network.
The present work raised open theoretical questions regarding
the stability of spatially and, more difficult, space-time depen-
dent processor sharing queues modeling the performance of
individual network cells (cf Section III-D). More work is also
required to figure out the problem of different performance of
the network during day and night hours (cf Figure 8).

619