Wireless Downlink Data Channels:

User Performance and Cell Dimensioning

Thomas Bonald and Alexandre Proutière

France Telecom R&D
38-40 rue du Général Leclerc
92794 Issy-les-Moulineaux, France
{thomas.bonald,alexandre.proutiere}@francetelecom.com

ABSTRACT only one user in each slot. It has been shown that, assum-
We consider wireless downlink data channels where the trans- ing the feasible transmission rate is linear in the signal to
mission power of each base station is time-shared between a interference-plus-noise (SINR) ratio, this maximizes overall
dynamic number of active users as in CDMA/HDR systems. throughput [18, 11, 4]. Note that this optimality principle
We derive analytical results relating user performance, in is not strictly valid when only a discrete set of data rates is
terms of blocking probability and data throughput, to cell available [4].
size and traffic density. These results are used to address a A key component of such a TDMA-like scheme is the
number of practically interesting issues, including the trade- scheduling, i.e., deciding which user should be served in each
off between cell coverage and cell capacity and the choice of time slot. In fact, it is not directly clear what a “good”
efficient scheduling and admission control schemes. scheduling strategy is, as the potential data rate of a user
depends on her/his radio conditions, mainly determined by
the distance to the BS and fading effects. To transmit always
Categories and Subject Descriptors to the user with the highest potential rate maximizes overall
B.8 [Performance and Reliability]: Performance Analy- throughput but typically results in the starvation of distant
sis and Design Aids users. Another strategy, which realizes a reasonable trade-
off between efficiency and fairness, consists in transmitting
General Terms to the user with the highest potential rate proportionally to
her/his current mean data rate [23]. This algorithm, termed
Performance Proportional Fair (PF) and implemented in HDR systems,
has indeed been shown to fairly share the transmission re-
Keywords source [10]. Many other scheduling algorithms have been
proposed and analyzed (see [8] and references therein).
CDMA/HDR systems, flow-level analysis, dimensioning
As a general rule the evaluation of scheduling algorithms is
performed with an assumed static population of users (see,
1. INTRODUCTION e.g., [12]). We maintain that this may lead to misleading
Data services are expected to constitute a significant part conclusions since the actual set of active users is dynamic
of traffic in future CDMA networks. In this paper, we derive and varies as a random process as new data flows are initi-
analytical performance results for downlink data channels, ated and others complete. In particular, while users are gen-
accounting for the random nature of traffic demand, and erally assumed to be uniformly distributed in the cell, the
show the practical interest of these results in dimensioning location of active users in steady state does depend on the
cells and designing radio control algorithms. scheduling employed. This is due to the inherent “elasticity”
To support high data rates, a number of new technologies of data transfers: the resource attributed to any user deter-
have been standardized, such as HDR (High Data Rate) sys- mines how long that user will stay active. Thus a scheduling
tems [6], corresponding to the CDMA2000 1xEV-DO stan- scheme that favors near users results in a large proportion
dard, and their 3GPP equivalent, HSDPA (High Speed Down- of active users being far from the BS.
link Packet Access) systems [1, 17]. Both systems are based A further issue when accounting for the statistical nature
on an intra-cell interference cancellation principle: time is of traffic is that of admission control: situations may arise
slotted and the base station (BS) transmits at full power to where it is preferable to block new demands rather than to
further degrade the performance of ongoing data flows. Ad-
mission control has already been proposed for data services
in a wireline context to preserve network efficiency in over-
Permission to make digital or hard copies of all or part of this work for load [5]. The design and evaluation of admission control
personal or classroom use is granted without fee provided that copies are schemes for the considered wireless network are still largely
not made or distributed for profit or commercial advantage and that copies unexplored areas.
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific In evaluating scheduling and admission control it is im-
permission and/or a fee. portant to understand how offered traffic impacts user per-
MobiCom’03, September 14–19, 2003, San Diego, California, USA. ceived performance, in terms of transfer delays and blocking
Copyright 2003 ACM 1-58113-753-2/03/0009 ...$5.00.
rates. Given the ever changing nature of data applications, 2. MODEL
control mechanisms should be designed so that performance In this section, we first present the model of the radio re-
depends mainly on an appropriately defined load factor, and source and the way this resource is shared. We then describe
not on precise traffic characteristics like the distribution of the characteristics of offered traffic.
a typical transfer volume. This requirement underscores the
interest of analytical modeling tools, in addition to simu- 2.1 Radio resource
lation, in order to gain the necessary insight and identify We consider a cell with a single downlink channel whose
efficient design choices. resource is time-shared between active users. Denote by φbu
Regarding dimensioning issues, it is not straightforward the fraction of time base station (BS) b transmits to user u,
to define the traffic carrying capacity of a CDMA network with u φbu = 1. The data rate of user u is then:
handling data. Resource consumption depends in particular
on the position of the users so that, in addition to the usual Cu = C × φbu , (1)
notion of intensity, traffic also has a spatial component. It
is notably important to know how the capacity of a cell where C is the peak data rate, obtained in the absence of
depends on its size, in analogy with the known trade-off in any other user in the cell, i.e., for φbu = 1.
circuit switched CDMA networks [22, 25]. In practice, the peak data rate depends in a complex way
on the radio environment and varies over time due to user
Related work. To our knowledge, very few papers address mobility, shadowing and multi-path reflections. Unless oth-
the issue of user performance for wireless data channels, erwise specified, we ignore these fading effects, i.e., we as-
accounting for the random nature of traffic and the inher- sume the peak data rate is approximately constant during
ent “elasticity” of data transfers. Most existing models in- data transfer. Section 6 is devoted to the impact of fast fad-
deed represent data transfers as circuit services (see, e.g., ing. To simplify the presentation, we also assume that the
[2]). A notable exception is the recent work of Borst for peak rate C depends on the distance r from BS b to user
CDMA/HDR systems [8]. User performance is explicitly u only. This last assumption is not essential and the the-
evaluated and shown to be insensitive to the flow size distri- oretical results derived in the following still hold in a more
bution, with or without admission control, in a symmetric realistic radio environment, given any peak rate function of
scenario where all users experience the same fast fading and user’s location in the cell. We denote by C0 the maximum
the resource allocation is that realized by the PF scheduler. peak rate (which depends on channel bandwidth and coding
The radio channel is modeled at flow level by a processor- efficiency) and by r0 the maximum distance at which this
sharing queue which is indeed known to have the insensitiv- maximum peak rate is achieved:
ity property [20]. C(r) = C0 for all r ≤ r0 . (2)

Contribution. In the present paper, we extend the analyt-

ical results of [8] to more general scheduling and admission Rate vs. SINR dependency. For numerical applications, we
control schemes and, accounting for the spatial component use the peak rate function given by the following standard
of offered traffic, apply them to a number of practical issues. model. Let W be the cell chip rate. If Pu denotes the power
We notably define the notion of “cell capacity”, critical for received by user u from her/his BS, η the background noise
dimensioning purposes, and show that: and Iu the interference due to other BS, the user’s signal to
interference-plus-noise ratio (SINR) and energy-per-bit to
• sharing the transmission power in a fair way is efficient noise density ratio are respectively given by [24]:
with respect to user performance;
Pu Eb W
SINRu = , = × SINRu . (3)
η + Iu N0 C
• the waste of radio resources due to the granularity of
feasible rates induced by coding constraints is typically Given a target error probability, it is necessary that Eb /N0 ≥
not significant; δ for some threshold δ, which is assumed to be the same for
all users. The peak data rate of user u is then the mini-
• the so-called “cell breathing” effect arises at very high mum of C0 and W × SINRu /δ. In particular, it is linear in
loads only. In particular, a simple admission control the SINR up to the maximum peak rate C0 . The assump-
independent of user locations is sufficient in practice; tion of a constant Eb /N0 target is generally valid as long as
the same type of modulation is used for all data rates [15].
• the number of active users is typically rather small in In HDR systems for instance, δ is approximately equal to
steady state. In particular, the impact of opportunistic 2.5dB for all data rates except for the three highest, with
schedulers like PF that take advantage of fast fading a maximum value of 6.5dB [6]. The impact of such a rate-
is typically much smaller than one would expect with dependent target δ resulting in a non-linear rate vs. SINR
an assumed static user population. dependency is evaluated in Section 5.

Propagation model. In the first part of the paper, we ne-

Outline. In the next section, we present the model used to glect the interference term Iu , i.e., we consider a single cell
derive the analytical results. In Sections 3 and 4, user per- in isolation. Equivalently, we assume that interference is
formance is evaluated in terms of throughput and blocking constant over the considered cell. The impact of the inter-
rate. The following three sections are devoted to the impact ference generated by other BS is evaluated in Section 7. The
of non-linear data rate vs. SINR dependency, fast fading and power Pu received by user u is equal to P × Γu where P is
interference, respectively. Section 8 concludes the paper. the transmission power of the BS and Γu denotes the path
loss. In practice, path loss varies with respect to the user’s rate. This corresponds to their demand on the radio resource
location and radio conditions. Here we adopt a simple prop- (here the time slot). Thus we define the load of a cell of
agation model where the path loss Γu is a function Γ of the radius R as:
distance r from the BS to user u only: R
dρ(r)
ρ̄ = . (5)
C(r)


1 if r ≤ ε, 0
Γ(r) =

ε α
r
otherwise, Users experience quality of service through the duration of
where ε denotes the maximum distance at which the full data transfers. Note that, in view of (1), this does not only
power P is received and α is the path loss exponent which depend on user characteristics such as her/his peak rate,
characterizes the radio environment (typical values of α are but also on the cell activity, i.e., on the dynamic number of
between 2 and 5). Assuming that the maximum peak rate active users who share the transmission resource. We are
C0 can be achieved (thus r0 > ε), it follows from (2) that interested in the mean flow duration T (r) for a user whose
the peak rate function is: distance to the BS is r (recall that active users don’t move
during their data transfer in our model). Let dx(r) be the
1 if r ≤ r0 , mean number of active users whose distance to the BS is
C(r) = C0 ×

r0 α (4)
r
otherwise. between r and r + dr. Applying Little’s law [16], we get:
dx(r) = T (r)λ × 2πrdr.
Coding constraints. Expression (4) corresponds to an ideal Thus the flow throughput γ(r) of users whose distance to
case where the set of achievable peak rates is continuous. In the BS is r, defined as the ratio of the mean flow size E[σ]
practice, coding constraints result in a discrete set of achiev- to the mean flow duration T (r), is given by:
able peak rates C0 ≡ c0 > c1 > . . . > cn . In view of (4),
these rates define a set of concentric “rings” of external ra- dρ(r)
γ(r) = . (6)
dius r0 < r1 < . . . < rn corresponding to regions where dx(r)
these rates are achievable. Table 1 below gives the rates This quantifies the average performance of data transfers at
defined for HDR channels [6] with the corresponding radius a distance r from the BS.
(normalized so that r0 = 1) evaluated from (4) with two
values of the path loss exponent. Coding constraints. When the set of achievable peak rates
is discrete, the traffic intensity in ring 0 is given by:
Ring k Rate ck Radius rk Radius rk
(Kbit/s) (α = 4) (α = 2) ρ0 = ρπr02 ,
0 2457.6 1 1 while the traffic intensity in ring k, k = 1, . . . , n, is given by:
1 1843.2 1.07 1.15
2 1228.8 1.19 1.41
2
ρk = ρπ(rk2 − rk−1 ).
3 921.6 1.28 1.63 As above, we define the cell load as:
4 614.4 1.41 2.00 n
5 307.2 1.68 2.83


ρ̄ = ρ̄k ,
6 204.8 1.86 3.46 k=0
7 153.6 2.00 4.00
8 102.6 2.21 4.90 where ρ̄k = ρk /ck denotes the load of ring k. The shape and
9 76.8 2.37 5.61 the load contribution of each ring are illustrated in Figure 1
for the values given in Table 1 and n = 4, 7, 10, correspond-
10 38.4 2.82 7.94
ing to cells of radius r4 , r7 , r10 , respectively1 . We observe
that for large cells (corresponding to a large spectrum of
Table 1: Rates and ring radius. possible peak rates), most load is concentrated in the outer
ring.
We are interested in the mean flow duration Tk for a user
2.2 Traffic characteristics in ring k. Denote by xk the mean number of active users in
We assume traffic demand is uniformly distributed in the ring k. By Little’s law:
cell. Data flows arrive as a Poisson process of intensity λ×ds E[x0 ] = T0 λ × πr02 ,
in any area of surface ds. Flow sizes are independent and
identically distributed (i.i.d.). We denote by σ the corre- and for k = 1, . . . , n,
sponding random variable and by ρ = λ × E[σ] the traffic 2
E[xk ] = Tk λ × π(rk2 − rk−1 ).
density (in Kbit/s per surface unit). The traffic intensity
generated by those users whose distance to the BS is be- We deduce the flow throughput for a user in ring k, defined
tween r and r + dr is dρ(r) = ρ × 2πrdr. We consider two as the ratio of the mean flow size E[σ] to the mean flow
cases, depending on whether coding constraints are taken duration Tk :
into account or not. ρk
γk = . (7)
E[xk ]
No coding constraint. In the ideal case where a continuous 1
The maximum cell radius obtained for α = 2 is in fact
set of peak rates is achievable, the load generated by those almost three times larger than that obtained for α = 4; we
users whose distance to the BS is between r and r + dr is changed the scale between the two sets of cells of Figure 1
the ratio dρ(r)/C(r) of their traffic intensity to their peak for sake of readability.
 Path loss exponent α = 4 grows indefinitely in overload (ρ̄ > 1). In the latter case,
the data rate C(r)/x tends to zero for all users, whatever
their distance to the BS: the cell is saturated.

Cell capacity. We may define the cell capacity as the max-

imum traffic intensity for which the cell is not saturated. In
view of (8), this is a function C̄ of the cell radius R:
R −1
< 10% > 10% > 20% > 50% 2rdr
C̄(R) = .
C(r)R2


Note that C̄ is a decreasing function of R, equal to the max-

Path loss exponent α = 2 imum peak rate C0 for R ≤ r0 . Figure 2 gives the cell ca-
pacity with respect to its radius (normalized values so that
C0 = 1, r0 = 1) for the peak rate function (4). We observe
that the cell capacity decreases suddenly when R ≥ r0 . The
capacity of large cells is extremely small.

1
Path loss exponent = 2
< 10% > 10% > 20% > 50% Path loss exponent = 4
0.8

Cell capacity
0.6

Figure 1: Ring shapes and load distribution. 0.4

(n = 4, 7, 10)
0.2

3. THROUGHPUT PERFORMANCE 0
0 1 2 3 4 5 6 7 8
(NO ADMISSION CONTROL) Cell radius
In this section, we evaluate user performance in terms of
flow throughput in the absence of admission control, i.e., Figure 2: Cell capacity – defined as the maximum
when the number of active users is not limited. We first as- traffic intensity without saturation – with respect to
sume that the radio resource is fairly shared between active cell radius, for path loss exponents α = 2 and α = 4.
users, i.e., φbu = 1/x, when x users are active in the cell.
This is the allocation realized by the PF scheduler or a sim-
ple round-robin scheduler, for instance (these are equivalent
in the absence of fast fading). Other power allocations are Flow throughput. To evaluate user performance in under-
considered in §3.3. load, we use a key property of the processor-sharing queue:
the stationary distribution of the number of customers is
3.1 A continuous set of peak rates insensitive to the distribution of service times. This follows
We first consider the ideal case where a continuous set from the reversibility of the underlying Markov process [14,
of peak rates is achievable. As the transmission resource 20]. We deduce that the stationary distribution π of the
is fairly shared between active users, the number of active number of active users is insensitive to the flow size distri-
users x evolves like the number of customers in a processor- bution and given by:
sharing queue with Poisson arrivals of intensity λπR 2 and
i.i.d. service times [16]. Each service time is equal to the flow π(x) = ρ̄x (1 − ρ̄). (9)
duration in the absence of any other user in the cell, i.e., Note that the number of active users is typically rather small
σ/C(r) for a user whose distance to the BS is r. Thus the in steady state: the probability having more than x active
distribution of the random variable σ̄ representing a typical users decreases geometrically with rate ρ̄. In addition, the
service time is given by: probability that the distance of an active user to the BS is
σ 2rdr between r and r+dr is proportional to the load generated by
dσ̄(r) = , r ≤ R. these users, namely dρ(r)/C(r). In particular, the density
for all users in dr C(r) R2
of active users is inversely proportional to their peak rate
In particular, the load of the processor-sharing queue corre- C(r). We deduce:
sponds to the cell load:
dρ(r)
R R dx(r) = × E[x],
E[σ] 2rdr ρ ρ̄C(r)
λπR2 × = 2πrdr ≡ ρ̄. (8)
C(r) R2 C(r)
 

0 0 where, in view of (9), the mean number of active users is:

We conclude that the number of active users x tends to ρ̄
a finite stationary regime in underload (ρ̄ < 1), while it E[x] = .
1 − ρ̄
It then follows from (6) that: Cell capacity. As above, we define the cell capacity as max-
imum traffic intensity for which the cell is not saturated:
γ(r) = C(r)(1 − ρ̄). −1
n


pk
Hence, the flow throughput is equal to the peak rate for C̄(R) = .
ck
k=0

ρ̄ = 0 and decreases linearly in the cell load.

Note that the cell capacity is always smaller than that ob-
3.2 A discrete set of peak rates tained in the ideal case considered in §3.1. Figure 3 gives
We now consider the practically interesting case where the cell capacity obtained for the values of Table 1. We ob-
only a discrete set of peak rates is available. As above, the serve that the difference with the ideal case is very small,
number of active users evolves like the number of customers indicating that resource wastage due to coding constraints
in a processor-sharing queue with Poisson arrivals of rate is limited.
λπR2 and i.i.d. service times. The random variable σ̄ rep-
resenting a typical service time is now equal to σ/ck with Flow throughput. The stationary distribution of the num-
probability pk , k = 0, 1, . . . , n, with: ber of active users in each ring is that of the number of
customers in a processor-sharing queue with classes of re-
r02 rk2 − rk−1
2 spective loads ρ̄0 , . . . , ρ̄n :
p0 = , pk = , k = 1, . . . , n.
R2 R2 (x0 + . . . + xn )! x0
π(x0 , . . . , xn ) = ρ̄0 . . . ρ̄xnn (1 − ρ̄). (10)
Again, the load of the processor-sharing queue corresponds x0 ! . . . xn !
to the cell load: Again, the probability having more than x active users de-
n creases geometrically with rate ρ̄ = ρ̄0 + . . . + ρ̄n . The prob-


E[σ] ability an active user is in ring k is proportional to the load

λπR2 × pk ≡ ρ̄.
ck ρ̄k = ρk /ck generated by these users. In particular, the den-
k=0
sity of active users is inversely proportional to their peak
rate, as in the absence of coding constraint. For instance,
Path loss exponent α = 4 for the values given in Table 1, the mean number of active
users per surface unit in ring 5 is always twice that in ring
1
Without coding constraint 4, whatever the values of the path loss exponent or the cell
With coding constraints load.
0.8 In view of (10), the mean number of active users in ring
k is:
Cell capacity

0.6 ρ̄k
E[xk ] = ,
1 − ρ̄
0.4 and from (7):
γk = ck (1 − ρ̄).
0.2
Again, the flow throughput decreases linearly in the cell
load. This is illustrated in Figure 4 for each ring of Table 1.
0
0 1 2 3 4
Cell radius 2.5
Path loss exponent α = 2
1 2
Without coding constraint
Throughput (Mbit/s)

With coding constraints

0.8 1.5
Cell capacity

0.6 1

0.4 0.5

0.2 0
0 0.2 0.4 0.6 0.8 1
Cell load
0
0 1 2 3 4 5 6 7 8
Cell radius
Figure 4: Flow throughput with respect to cell load
for the 11 rings of Table 1.
Figure 3: Cell capacity – defined as the maximum
traffic intensity without saturation – with respect to
cell radius, with and without coding constraints. 3.3 Other power allocations
The allocation so far considered, referred to as “fair power”
sharing, has the practically interesting property that user
performance can be explicitly evaluated and is insensitive These allocations, which are known to outperform the pro-
to the flow size distribution. It may be considered as “un- cessor sharing discipline for a heavy-tailed service distribu-
fair”, however, as the flow throughput experienced by a user tion, have been proposed to be implemented in Web servers
is proportional to her/his peak data rate (refer to Figure 4). in a wireline context [3, 19] and more recently in scheduling
algorithms in a wireless context [13]. They would here exac-
Fair rate sharing. Consider another allocation, referred to erbate the discrimination against far users, whose processing
as “fair rate” sharing, where the data rates of all active users time is typically much higher than that of near users.
are made equal. This is realized if BS b transmits to user The above observations suggest that no significant gain is
u a fraction of time inversely proportional to her/his peak achieved by those allocations which do not share the trans-
rate: mission resource in a fair way and that the so-called “near-
1/C(r(u)) far” unfairness (in terms of flow throughput) is inherent to
φbu = 0
, wireless data systems. The only way to achieve approxi-
u0 1/C(r(u )) mately fair throughput performance is to limit the cell size
where r(u0 ) denotes the distance from BS b to user u0 . The so that all users can achieve high data rates. Finally, it is
number of active users then evolves like the number of cus- worth noting that the key properties satisfied by the “fair
tomers in a discriminatory processor-sharing queue [9] of power” allocation (explicit performance evaluation and in-
load ρ̄. Again, the cell is saturated in overload (ρ̄ > 1). In sensitivity) still hold in the presence of admission control,
underload (ρ̄ < 1), the number of active users x tends to as shown in the next section. The impact of admission con-
a finite stationary regime, but the stationary distribution is trol on user performance is typically extremely difficult to
sensitive to the flow size distribution. evaluate for other power allocations.
For an exponential flow size distribution, we obtain using
[9] the results of Figure 5 for a 5-ring cell, with the values
of Table 1. We observe that the gain in flow throughput for 4. BLOCKING RATE
users in the outer ring is very limited. On the other hand,
the impact on the performance of users in the inner disk is (WITH ADMISSION CONTROL)
significant. This may be explained as follows. First, the fact In the absence of admission control, we have seen that the
that active users have the same data rate does not imply cell is saturated in overload (ρ̄ > 1): the number of ongoing
that users have the same flow throughput. When the cell transfers grows indefinitely and the data rate of each transfer
load ρ̄ is close to 0 for instance, an active user is typically eventually tends to zero. Admission control is necessary to
alone in the cell so that the flow throughput is equal to the guarantee a minimum data rate cmin for all users whatever
peak rate. Second, most load is concentrated in the outer the cell load. For a given target rate cmin , 0 < cmin < C0 , the
rings (refer to Figure 1): the performance experienced by cell radius R cannot exceed Rmax , the maximum distance r
far users is essentially due to their own load and cannot such that C(r) ≥ cmin .
be significantly improved; the performance experienced by In the presence of admission control, users experience
near users, on the other hand, mainly depends on the load quality of service not only through flow throughput but also
generated by far users thus may be significantly worsened. through blocking rate. In this section, cell capacity is eval-
uated in terms of the maximum traffic intensity for given
2.5 minimum data rate cmin and target blocking probability.
Fair power sharing
Fair rate sharing
2 4.1 Admission control based on the number of
active users
Throughput (Mbit/s)

1.5 We first consider a simple admission criterion based on the

number of active users. For a cell of radius R, the minimum
rate cmin is guaranteed if the number of active users does
1
not exceed:
C(R)
0.5 m= .
cmin
0
0 0.2 0.4 0.6 0.8 1 No coding constraint. We first consider the case where the
Cell load set of available peak rates is continuous. As the Markov pro-
cess associated with the unconstrained system considered in
Figure 5: Throughput of users in the inner disk (up- §3.1 is reversible, the restriction of its stationary distribution
per curves) and the outer ring (lower curves) with to admissible states gives the stationary distribution of the
respect to cell load in a 5-ring cell, for two different Markov process associated with the constrained system [14].
power allocations. In addition, the stationary distribution associated with the
constrained system remains insensitive to the service time
distribution. Assuming m is an integer for simplicity, it
then follows from (9) that the stationary distribution of the
Service-dependent allocations. Other possible allocations number of active users is:
are those which depend on the remaining or the processed
service, such as the “shortest remaining processing time” or ρ̄x
the “foreground-background” disciplines, respectively [16]. π(x) = , 0 ≤ x ≤ m. (11)
1 + ρ̄ + . . . + ρ̄m
Thus the blocking rate, which is independent of the user’s 2.5
distance to the BS, is given by:
ρ̄m 2
B= . (12)

Throughput (Mbit/s)
1 + ρ̄ + . . . + ρ̄m
1.5
In addition, the probability that the distance of an active
user to the BS is between r and r + dr is proportional to the
actual load generated by these users, namely: 1

dρ(r)
(1 − B) . 0.5
C(r)
In particular, the density of active users is inversely propor- 0
tional to their peak rate C(r) as in the absence of admission 0 0.5 1 1.5 2
control (refer to §3.1). We also deduce: Cell load

1 dρ(r) Figure 6: Flow throughput with respect to cell load

dx(r) = × E[x],
ρ̄ C(r) in each ring of a 5-ring cell when the number of
where, in view of (11), the mean number of active users is: active users is limited to m = 12.

ρ 1 − (m + 1)ρm + mρm+1
E[x] = × . (13)
1−ρ 1 − ρm+1
The objective here is to evaluate the gain in terms of
As the actual traffic intensity generated by those users whose blocking probability compared to the admission control based
distance to the BS is between r and r + dr is (1 − B)dρ(r), on a maximum number of active users m given by:
expression (6) becomes:
C(R)
dρ(r) m= .
γ(r) = (1 − B) . cmin
dx(r)
It then follows from (12) and (13) that: No coding constraint. For a continuous set of available
(1 − ρ)(1 − ρm ) peak rates, admissible states are those for which
γ(r) = C(r) × .
1 − (m + 1)ρm + mρm+1 min1≤u≤x C(r(u))
≥ cmin ,
Hence, the flow throughput is equal to the peak rate for x
ρ̄ = 0 and decreases to C(r)/m when the cell load ρ̄ tends where r(u) denotes the distance from the BS to active user
to infinity. u. Assume M is an integer for simplicity. Let L = M − m,
and for j = 0, . . . , L, define Rj as the maximum distance r
Coding constraints. The results are similar if the set of such that:
available peak rates is discrete. For a cell of radius R = rn ,
the maximum number of users is m = cn /cmin . The blocking C(r) ≥ (M − j)cmin .
rate is the same in all rings and given by (12). The flow Note that r0 ≡ R0 < R1 < . . . < RL ≡ R. Admissible
throughput in ring k is: states are those for which there are less than m active users
(1 − ρ)(1 − ρm ) or there are x = M −j active users and the distance from the
γk = c k × . BS to any of these users does not exceed Rj , j = 0, . . . , L.
1 − (m + 1)ρm + mρm+1
As in §4.1, the reversibility property implies that the sta-
Figure 6 gives the corresponding flow throughput for a 5- tionary distribution π of the number of active users is insen-
ring cell with the values of Table 1 and a maximum number sitive to the flow size distribution and given by the restric-
of users m = 12, corresponding to a minimum data rate tion to admissible states of the stationary distribution (9)
cmin = 51 Kbit/s (cmin /C0 ≈ 0.02). The flow throughput associated with the unconstrained system. Noting that in
decreases from ck to ck /m for each ring k. this unconstrained system, the probability that the distance
of an active user to the BS is less than Rj is equal to ρ̄→j /ρ̄,
4.2 Admission control based on the minimum where
data rate Rj
We now consider an admission control based on the mini- dρ(r)
ρ̄→j = ,
C(r)


mum data rate: a new data transfer is accepted if and only 0

if its rate and the rate of ongoing transfers would be larger we deduce:
than cmin . In particular, the admission decision now de-
pends on the location of users. The number of users cannot ρ̄x for x < m,
π(x) = π(0) × M −i
exceed: ρ̄→i for x = M − i, i = 0, . . . , L,
C0 where
M = ,
cmin −1
L
which corresponds to the best case where the distance from M −i


π(0) = 1 + ρ̄ + . . . + ρ̄m−1 + ρ̄→i .

the BS to any active user does not exceed r0 . i=0


We also deduce the blocking rate Bj for users whose distance Note that 0 ≡ k0 ≤ k1 ≤ . . . ≤ kL ≡ n. Admissible states
to the BS is Rj , j = 0, . . . , L: are those for which there are less than m active users or
L L
there are x = M − j active users and the distance from the


M −i


M −i BS to any of these users does not exceed rkj , j = 0, . . . , L.

ρ̄→i − ρ̄→i−1
Again, it follows from the reversibility of the Markov pro-
i=0 i=j+1
Bj = L
. cess associated with the unconstrained system considered in


M −i §3.2 that the stationary distribution of the number of active

1 + ρ̄ + . . . + ρ̄m−1 + ρ̄→i
users is:
i=0
ρ̄x for x < m,
Noting that the blocking rate B(r) for users whose distance π(x) = π(0) × M −i
ρ̄→k for x = M − i, i = 0, . . . , L,
to the BS is r is i

B(r) = B0 if r ≤ R0 , where
L −1
and for j = 1, . . . , L, M −i


π(0) = 1 + ρ̄ + . . . + ρ̄m−1 + ρ̄→k i

,
B(r) = Bj if Rj−1 < r ≤ Rj , i=0

and for k = 0, . . . , n,
we get the mean blocking rate:
k
L
Rj2 − Rj−1
2


R2 

ρ̄→k = ρ̄j .
B = B0 02 + Bj . (14)
R j=1
R 2 j=0

The blocking rate in ring k is given by:

Figure 7 gives the corresponding cell capacity (normalized
values so that C0 = 1, r0 = 1) compared to that obtained L L
M −i M −i
 

with the admission control based on a maximum number of ρ̄→k i

− ρ̄→k i−1
users m. We observe that no significant gain is achieved. Bk =
i=0 i=jk +1
,
This is actually true for any reasonable values of the mini- 
L
M −i
mum data rate (cmin < 0.1, say, corresponding to 246 Kbit/s 1 + ρ̄ + . . . + ρ̄m−1 + ρ̄→k i
for HDR channels) and target blocking rates (B < 10%, i=0

say). We explain this result in §4.3 below. where jk denotes the minimum integer j 0 such that kj 0 = k,
and the mean blocking rate is
n
Number of users r02 

rk2 − rk−1
2
1 Minimum data rate B = B0 + Bk . (15)
R2 R 2
k=1
0.8
Figure 8 is the analog of Figure 7 for the values of Table
Cell capacity

1. Again, we observe that no significant gain is achieved

0.6
with the admission control based on the minimum data rate.
Furthermore, comparing Figures 7 and 8 indicates as in §3.2
0.4 that resource wastage due to coding constraints is limited.

0.2
Number of users
1 Minimum data rate
0
0 0.5 1 1.5 2 2.5
Cell radius 0.8
Cell capacity

Figure 7: Comparison of the cell capacity – defined 0.6

as the maximum traffic intensity for a minimum
data rate cmin = 0.02 and target blocking rates 1% 0.4
(lower curves) and 5% (upper curves) – obtained for
two different admission criteria, without coding con- 0.2
straints (path loss exponent α = 4).
0
0 0.5 1 1.5 2 2.5
Cell radius

Coding constraints. For a discrete set of peak rates, ad-

missible states are those for which Figure 8: Comparison of the cell capacity – defined
as the maximum traffic intensity for a minimum data
min ck {xk>0} rate cmin = 0.02 and target blocking rates 1% (lower
0≤k≤n
≥ cmin , curves) and 5% (upper curves) – obtained for two
x
different admission criteria, with coding contraints
where xk denotes the number of active users in ring k. For
(path loss exponent α = 4).
j = 0, . . . , L, define kj as the largest integer k such that:
ck ≥ (M − j)cmin .
4.3 Cell breathing all users whatever their location: the so-called “near-far”
Previous results suggest that a simple admission control unfairness (in terms of blocking rates) does not hold.
based on the number of active users leads to a cell capacity
similar to that obtained with more complex schemes such as 5. IMPACT OF NON-LINEAR
that based on the minimum data rate where the admission
decision depends on user locations. As the blocking rate
RATE VS. SINR DEPENDENCY
B(r) obtained with the latter is an increasing function of We have so far assumed that the target energy-per-bit to
the distance r from the BS b to user u, one would expect noise density ratio δ is constant, resulting in a linear rate
the cell to “breath”, i.e., the active users to be near the BS vs. SINR dependency (up to the maximum peak rate C0 ).
when the cell load increases (refer to Figure 9). As noted in §2.1, this assumption is not necessarily valid in
real systems. In this section, we study the impact of the
Low load High load rate-dependent target δ of HDR channels.

5.1 Cell shape and load distribution

< 5% Table 2 below gives the target SINR for each data rate
> 5%
defined in HDR systems [6]. As the chip rate is W = 1228.8
Kchip/s, we get from (3) the corresponding Eb/N 0 target:
> 10% δ0 ≈ 6.5dB, δ1 ≈ 5.4dB, δ2 ≈ 3dB, and δk ≈ 2.5dB for all
> 50% rings k = 3, . . . , 10. Thus the Eb/N 0 target is approximately
constant except for the highest peak rates. This can notably
be explained by the specific modulation used for these rates.
Figure 9: Cell breathing: repartition of active users
Ring Rate SINR Radius Radius
with respect to cell load (α = 4).
(Kbit/s) (dB) (α = 4) (α = 2)
0 2457.6 9.5 0.79 0.63
This “cell breathing” effect indeed arises as for circuit ser- 1 1843.2 7.2 0.91 0.82
vices [25], but at very high loads only (the results of Figure 2 1228.8 3.0 1.16 1.34
9 were obtained for loads ρ̄ = 0.5 and ρ̄ = 100). This is 3 921.6 1.3 1.27 1.62
illustrated by Figure 10 which gives the mean blocking rate 4 614.4 −1.0 1.45 2.12
and the blocking rate in each ring with respect to the cell 5 307.2 −4.0 1.73 2.99
load for the values of Table 1 (α = 4): for a target mean 6 204.8 −5.7 1.91 3.64
blocking rate smaller than 10%, the blocking rate is approx- 7 153.6 −6.5 2.00 3.99
imately the same in all rings, and very well approximated 8 102.6 −8.5 2.24 5.02
by expression (12). 9 76.8 −9.5 2.37 5.63
10 38.4 −12.5 2.82 7.95
1
Table 2: Rates, target SINRs and ring radius for
0.8 HDR channels.
Blocking probability

0.6 In view of the propagation model of §2.1, the SINR is in-

versely proportional to r α (for r > ε). Table 2 gives the
corresponding external ring radius, normalized so that r0
0.4
would be equal to 1 if the Eb/N 0 target were constant and
equal to 2.5dB. Comparing with the values of Table 1, we
0.2 verify that the rate-dependent target δ of HDR channels
Individual rings
Mean blocking rate essentially impacts the inner rings. In particular, the load
0 distribution remains approximately the same, most load be-
0.001 0.01 0.1 1 10 100 1000 ing concentrated in outer rings.
Cell load
5.2 Cell capacity
Figure 10: Blocking rate with respect to cell load for As the load distribution is not significantly affected by the
the 11 rings of Table 1 with a minimum data rate of rate-dependent target δ, we expect the qualitative results of
19.2 Kbit/s (cmin = c10 /2). Section 4 to hold. In particular, the “cell breathing” ef-
fect arises at very high loads only and an admission decision
based on the number of active users is sufficient. This is
The above observation can be explained by the fact that confirmed by the results of Figure 11 for instance, which
a significant part of the traffic intensity is generated in the show that no significant gain is achieved by an admission
outer rings: if the blocking rate in these rings is close to control based on the minimum data rate. We also observe
1, the mean blocking rate is necessarily high (larger than that the impact of the rate-dependent target δ on cell ca-
the fraction of traffic intensity generated in these rings). At pacity is mainly due to the reduction of inner rings (the cell
nominal loads corresponding to mean blocking rates smaller capacity for a constant Eb/N 0 target is here defined for an
than 10%, the blocking rate is approximately the same for admission control based on the number of users).
 its negative effects for those users whose distance to the
Constant target Eb/No
HDR - Number of users BS is around r0 . The nearest users almost always get the
1
HDR - Minimum rate maximum peak rate C0 while the farest users almost never
get the maximum peak rate C0 : the former are insensitive
0.8
to fast fading while for the latter, the negative and positive
Cell capacity

effects of fast fading cancel out.

0.6

0.4 Opportunistic scheduling

1 Round-robin scheduling
0.2 No fast-fading

0.8

Cell capacity
0
0 0.5 1 1.5 2 2.5
0.6
Cell radius

0.4
Figure 11: Comparison of the cell capacity – defined
as the maximum traffic intensity for a minimum data
0.2
rate cmin = 0.02 and a target blocking rate 1% – ob-
tained for a constant target energy-per-bit to noise
0
ratio and for HDR channels with two admission cri- 0 0.5 1 1.5 2 2.5
teria (path loss exponent α = 4). Cell radius

Figure 12: Impact of fast fading on the cell capac-

ity, defined as the maximum traffic intensity for a
6. IMPACT OF FAST FADING minimum data rate cmin = 0.02 and a target blocking
We now study the impact of fast fading on cell capacity. rate 1% (Rayleigh fading, path loss exponent α = 4).
Fast fading is an extremely complex physical phenomenon
involving multi-path reflections [21]. Here we assume that
all users experience the same fast fading in the sense that
the SINR of user u at time t is ξu (t)×SINRu where ξu (t) are 6.2 Opportunistic scheduling
i.i.d. copies of some stationary process ξ(t) with unit mean A number of so-called “opportunistic” schedulers have
and SINRu denotes the SINR user u would get in the ab- been proposed to take advantage of fast fading [23, 7]. The
sence of fast fading. For numerical applications, we take the principle is to transmit to the various users when their radio
standard Rayleigh fading corresponding to an exponential conditions are relatively favorable, while ensuring fair access
marginal distribution of the process ξ(t). to the transmission resource. A typical example is the PF
In the rest of the paper, we consider the ideal case where scheduler mentioned in Section 1. The impact of oppor-
a continuous set of peak rates is available; we have verified tunistic scheduling on user performance has been evaluated
as above that resource wastage due to coding contraints is in [8] in a symmetric scenario where fast fading equally im-
limited. We also consider an admission control based on the pacts the data rates of active users. As observed above, fast
number of users only; again, the impact on cell capacity of fading does not affect all users in the same way, however,
an admission decision based on the minimum data rate is due notably to the maximum peak rate C0 .
limited. To simplify the analysis, we consider three user classes:
6.1 Round-robin scheduling • near users, whose distance to the BS is r < r 0 and for
We first consider a round-robin scheduling, that does not which the peak rate is almost always equal to C0 (with
take advantage of the time-varying radio conditions of each probability > 95%, say);
user. Averaging over the fast fading variations, it follows
from (4) that the data rate a user would get at distance r if • far users, whose distance to the BS is r > r 00 and for
she/he were alone in the cell is given by: which the peak rate is almost never equal to C0 (also
with probability > 95%);
r0 α
C(r) = E[C0 × min(ξ , 1)],
r

• the other users, whose distance to the BS is r ∈ [r 0 , r00 ]
where ξ ≡ ξ(0) corresponds to the marginal distribution of and for which the peak rate is equal to C0 with non-
the fading process. In particular, the blocking rate is given negligible probability.
by (12), for the corresponding cell load: For instance, the impact of Rayleigh fading on SINR is less
R
dρ(r) than 4.8dB (resp. larger than −13dB) with probability >
ρ̄ = . 95%: for a path loss exponent α = 4, we deduce the corre-
C(r)


0
sponding distances: r 0 ≈ 0.47, r00 ≈ 1.3.
As illustrated in Figure 12 for Rayleigh fading, the impact We denote by x the number of near users, x00 the number
of fast fading on cell capacity is significant for a cell radius of far users, and x0 the number of other users. For near
R ≈ 1 only (recall the convention r0 = 1). This can be users, we have C(r) ≈ C0 and there is no scheduling gain.
explained by the fact that, due to the maximum peak rate For far users, we have C(r) ≈ C0 ×( rr0 )α and the distribution
C0 , the positive effects of fast fading do not compensate of the feasible rate is approximately that of ξ × C(r). The
transmission rate of any user at distance r > r 00 from the 7. IMPACT OF INTERFERENCE
BS is then: In order to assess the impact of interference on cell capac-
C(r) ity, we consider two types of homogeneous networks: linear
G00 × ,
x + x0 + x00 networks, where BS are equidistant and placed on a common
infinite line; hexagonal networks, where cells are hexagons of
where G00 denotes the scheduling gain. In the presence of
the same size and cover the entire plane. In both cases, de-
i far users, a conservative approximation of the scheduling
note by 2 × R the distance between two BS. We assume that
gain G00 (i) is what one would obtain in the absence of any
the BS are always active and transmit at the same power P .
other users, corresponding to a symmetric scenario as con-
The interference suffered by user u served by BS b is then:
sidered in [8]. Denoting by ξ1 , ξ2 , . . . , ξi i.i.d. copies of ξ, we
deduce: 0


Iu = P × Γ(rub ),
G00 (i) = E[max(ξ1 , ξ2 , . . . , ξi )]. b0 6=b

For Rayleigh fading, we obtain [7]: 0

where rub denotes the distance from BS b0 to user u. Let
1 1 η̄ = η/P and I¯ = Iu /P , and denote by r the distance from
G00 (i) = 1 +
+ ... + . BS b to user u. The peak data rate of user u is then the
2 i
minimum of C0 and:
For the other users, we assume that the distribution of
the feasible rate is approximately the same, given by: W Γ(r)
× .
r 00 δ η + I¯
r0 α 2rdr
ξ0 × C 0 = C0 × min(ξ , 1) , To compare the results with those obtained for isolated cells,
r r00 2− r0 2


r0

0
we still denote by r0 the maximum distance at which this
where ξ is a unit mean random variable representing the maximum peak rate is achieved in the absence of interfer-
variations around the mean rate C 0 . The transmission rate ence, i.e.,
of any user at distance r ∈ [r 0 , r00 ] from the BS is then:
W Γ(r0 )
C0 C0 = × . (17)
G0 × , δ η̄
x + x0 + x00
where G0 denotes the scheduling gain. In the presence of i 7.1 Linear networks
such users, a conservative approximation of the scheduling We first consider a linear network. A reasonable approxi-
gain G0 (i) is what one would obtain in the absence of any mation consists in considering that interference is generated
other users, corresponding again to a symmetric scenario as by the 2 closest BS only. The interference term I¯ is then:
considered in [8]. Denoting by ξ10 , ξ20 , . . . , ξi0 i.i.d. copies of ξ 0 , ¯ = Γ(2R − r) + Γ(2R + r).
we deduce: I(r)

G0 (i) = E[max(ξ10 , ξ20 , . . . , ξi0 )]. We deduce the peak rate function:

The stationary distribution of the number of active users W Γ(r)

C(r) = min C0 , × ¯ .
of each class is insensitive to the flow size distribution and δ η + I(r)

given by:
x0 x00
(x + x0 + x00 )! x ρ̄0 ρ̄00
π(x, x0 , x00 ) = π(0) 0 00
ρ̄ .
x!x !x ! i=1
G0 (i) i=1
G00 (i)

where π(0) follows from the usual normalizing condition and

ρ̄, ρ̄0 and ρ̄00 denote the corresponding class loads:
r0 r 00 R Figure 13: A linear network.
dρ(r) 0 dρ(r) 00 dρ(r)
ρ̄ = , ρ̄ = , ρ̄ = .
C0 C0 C0 × ( rr0 )α
  

0 r0 r 00

Denoting by m the maximum number of users, the blocking Note that, in view of (17), this function is entirely deter-
rate is independent of user location and given by: mined by the maximum peak rate C0 , the distance r0 and
the ratio W/δ. Previous results still hold, with the cell load
x+x0 +x00 =m π(x, x0 , x00 ) given by:
B= 0 00
(16)
x+x0 +x00 ≤m π(x, x , x ). R
ρ2dr
ρ̄ = . (18)
As illustrated in Figure 12 for Rayleigh fading, the impact C(r)


0
of opportunistic scheduling on cell capacity is relatively lim-
The cell capacity, defined as the maximum traffic intensity
ited, especially for large cells. This is a rather counter-
without saturation, is given by:
intuitive result in view of the high scheduling gains (e.g.,
G0 (i) ≈ 1.5 and G00 (i) ≈ 2.9 for i = 10). This may notably R
dr
−1

be explained by the fact that, even in the absence of schedul- C̄(R) = .

C(r)R


0
ing gain and admission control, the number of active users
is typically rather small in steady state (cf. Section 3). The Figure 14 gives the cell capacity with respect to the cell
number of active users is here further limited by admission radius (normalized values C0 = 1, r0 = 1) for HDR param-
control, especially for large cells. eters (W = C0 /2, δ = 2.5dB) and negligible ε. This is a
decreasing function of the cell radius, with a maximum cell 7.2 Hexagonal networks
capacity equal to: Now consider a hexagonal network. The interference suf-
fered by a user u served by BS b is almost entirely generated
0.86 for α = 4,
C(0) ≈ by the 6 surrounding BS. A conservative approximation of
0.70 for α = 2.
the interference term I¯ is given by the following function of
Thus the impact of interference on cell capacity is significant, the distance r from BS b to user u:
especially for small cells. This is notably due to the fact ¯
I(r)= Γ(2R − r) + 2Γ( (R − r)2 + 3R2 )
that, for HDR parameters, the maximum peak rate C0 is
achievable at a distance r strictly smaller than R, whatever +2Γ( (R + r)2 + 3R2 ) + Γ(2R + r).
the cell radius R. This approximation is obtained assuming that user u is on
a segment from BS b to a neighbor BS b0 . We deduce the
1 peak rate function:
With interference
Without interference W Γ(r)
0.8 C(r) = min C0 , × ¯ .
δ η + I(r)

Cell capacity

0.6

0.4

0.2

0
0 2 4 6 8 10 12
Cell radius

Figure 14: Cell capacity in a linear network – de-

fined as the maximum traffic intensity without sat-
uration for path loss exponent α = 4 (lower curves)
and α = 2 (upper curves) – with and without inter- Figure 16: A hexagonal network.
ference.

1
With interference
Without interference
With interference 0.8
1 Without interference
Cell capacity

0.8 0.6
Cell capacity

0.6 0.4

0.4 0.2

0.2 0
0 2 4 6 8 10 12
0 Cell radius
0 0.5 1 1.5 2 2.5 3
Cell radius Figure 17: Cell capacity in a hexagonal network
– defined as the maximum traffic intensity with-
Figure 15: Cell capacity in a linear network – de- out saturation for path loss exponent α = 4 (lower
fined as the maximum traffic intensity for a mini- curves) and α = 2 (lower curves) – with and without
mum data rate cmin = 0.02 and target blocking rates interference.
1% (lower curves) and 5% (upper curves) – with and
without interference.
For simplicity, we approximate the hexagonal cells by cir-
cular cells of radius R. Figure 17 compares the cell capacity
In the presence of an admission control based on a max- – defined as the maximum traffic intensity without satura-
imum number of active users m = C(R)/cmin , the blocking tion – with and without interference. This is a decreasing
rate is given by expression (12) for the cell load (18). Fig- function of the cell radius, with a maximum cell capacity
ure 15 gives the corresponding cell capacity for a path loss equal to:
exponent α = 4. Again, we observe that the impact of in- 0.65 for α = 4,
C(0) ≈
terference on cell capacity is significant. 0.29 for α = 2.
 With interference
1 Without interference Some issues need to be further explored. For instance, we
assumed in Section 7 that BS are always active and observed
0.8 an important decrease of the cell capacity due to interfer-
ence. It would be useful to study how idle periods of BS
Cell capacity

0.6 could improve capacity. Other interesting issues include the

impact of mobility and the integration of data services with
0.4 voice and video services.

0.2 Acknowledgment. We are grateful to Sem Borst and Jim

Roberts for fruitful discussions on the work presented in this
0 paper.
0 0.5 1 1.5 2 2.5
Cell radius 9. REFERENCES
[1] 3GPP TS 25.848, Physical layer aspects of UTRA
Figure 18: Cell capacity in a hexagonal network – High-Speed Downlink Packet Access, Release 4, 2001.
defined as the maximum traffic intensity for a mini- [2] E. Altman, Capacity of Multi-service Cellular
mum data rate cmin = 0.02 and target blocking rates Networks with Transmission-Rate Control: A
1% (lower curves) and 5% (upper curves) – with and Queueing Analysis, in: Proc. of ACM Mobicom, 2002.
without interference. [3] N. Bansal and M. Harchol-Balter, Analysis of SRPT
Scheduling: Investigating Unfairness, in: Proc. of
ACM Sigmetrics Conference on Measurement and
In the presence of an admission control based on the num-
Modeling of Computer Systems, 2001.
ber of active users, the blocking rate is given by expression
(12). Figure 18 gives the corresponding cell capacity for [4] A. Bedekar, S. Borst, K. Ramanan, P. Whiting, E.M.
a path loss exponent α = 4. Again, we observe that the Yeh, Downlink scheduling in CDMA data networks,
impact of interference on cell capacity is significant. in: Proc. of IEEE Globecom, 1999.
[5] S. Ben Fredj, T. Bonald, A. Proutière, G. Régnié and
J. Roberts, Statistical Bandwidth Sharing: A Study of
8. CONCLUSION congestion at flow level, in: Proc. of ACM Sigcomm,
We have derived analytical results relating user perfor- 2001.
mance, in terms of blocking probability and data through- [6] P. Bender, P. Black, M. Grob, R. Padovani, N.
put, to cell size and traffic density, accounting for the ran- Sindhushayana and A. Viterbi, CDMA/HDR: A
dom nature of traffic and the way the radio resource is Bandwidth-Efficient High-Speed Wireless Data Service
shared. We have notably shown that the performance of the for Nomadic Users, IEEE Communications Magazine,
“fair power” sharing allocation (realized by a round-robin or 70–77, July 2000.
a PF scheduler) cannot be significantly improved. Further- [7] F. Berggren and R. Jntti, Asymptotically fair
more, this allocation has the practically interesting property scheduling in fading channels, in: Proc. of IEEE VTC
that user performance can be explicitly evaluated, with or Fall, 2002.
without admission control, and is insensitive to detailed traf-
[8] S. Borst, User-level performance of channel-aware
fic characteristics such as the flow size distribution. scheduling algorithms in wireless data networks, in:
We have observed that the impact of HDR feasible rate Proc. of IEEE Infocom, 2003.
constraints on cell capacity is generally negligible. We con-
[9] G. Fayolle, I. Mitrani and R. Iasnogorodski, Sharing a
clude that the optimality principle of a TDMA-like strategy
processor among many classes, Journal of the ACM
proved for a continuous set of available data rates remains
(27) 519–532, 1980.
valid in a more realistic situation where only a discrete set
of data rates is available. [10] J.M. Holtzman, Asymptotic Analysis of Proportional
Another key observation is that, due to the relatively Fair Algorithm, in: Proc. of 12th IEEE International
small number of active users in steady state, the impact Symposium on Personal, Indoor and Mobile Radio
of opportunistic schedulers that take advantage of fast fad- Communications, Sept-Oct 2001.
ing is much more limited than one would expect from the [11] J.M. Holtzman and S. Ramakrishna, A Scheme for
analysis of a static scenario with a fixed user population. In throughput maximization in a dual class CDMA
particular, the cell capacity does not much differ from that system, IEEE Journal on Selected Areas of
obtained with a simple round-robin scheduler, especially for Communications (40,2) 830–841, 1998.
large cells. [12] A. Jalali, R. Padovani and B. Pankaj, Data
Concerning admission control, we have studied the differ- throughput of CDMA-HDR a High Efficiency - High
ence in terms of cell capacity between a scheme based on Data Rate Personal Communication Wireless System,
the number of active users and another scheme based on in: Proc. IEEE VTC Spring, 2000.
the minimum data rate. The former leads to a blocking rate [13] N. Joshi, S.R. Kadaba, S. Patel and G.S. Sundaram,
which does not depend on user locations, unlike the latter. Downlink scheduling in CDMA Data Networks, in:
We have shown that provided the target blocking rate is not Proc. of ACM Mobicom, 2000.
too high (less than 10 % typically), these two schemes are [14] F. Kelly, Reversibility and Stochastic Networks, Wiley
in fact equivalent: it is not necessary to base the admission and Sons, 1979.
decision on user locations.
[15] S-L. Kim, Z. Rosberg, J. Zander, Combined Power
Control and Transmission Rate Selection In Cellular
Networks, in: Proc. of IEEE VTC Fall, 1999.
[16] L. Kleinrock, Queueing systems, Wiley and Sons, 1976.
[17] R. Love, A. Gosh, R. Nikides, L. Jalloul, M. Cudak
and B. Classon, High-Speed Downlink Packet Access
Performance, in: Proc. of IEEE VTC Spring, 2001.
[18] J.G. Pottie, System design choices in personal
communications, IEEE Personal Communications,
50–67, 1995.
[19] I.A. Rai, G. Urvoy-Keller and E. W. Biersack,
Size-based scheduling with differentiated services to
improve response time of highly varying flows, in:
Proc. of 15th ITC Specialist Seminar, Internet Traffic
Engineering and Traffic Management, Würzburg,
2002.
[20] R.F. Serfozo, Introduction to Stochastic Networks,
Springer Verlag, 1999.
[21] B. Sklar, Rayleigh Fading Channels in Mobile Digital
Communication Systems, Part I: Characterization,
IEEE Communications Magazine, 90–100, July 1997.
[22] V.V. Veeravalli, A. Sendonaris, The
Coverage-Capacity Tradeoff in Cellular CDMA
Systems, IEEE Trans. on Vehicular Technology,
1443–1451, 1999.
[23] P. Viswanath, D. Tse and R. Laroia, Opportunistic
Beamforming Using Dumb Antennas, IEEE Trans. on
Information Theory (48) 1277–1294, June 2002.
[24] A.J. Viterbi, CDMA: Principles of Spread Spectrum
Communication, Addison-Wesley, 1995.
[25] S-T Yang and A. Ephremides, Resolving the CDMA
Cell breathing Effect and Near-Far Unfair Access
Problem by Bandwidth-Space Partitioning, in:
Proc. of IEEE VTC Spring, 2001.