Professional Documents
Culture Documents
,Zwi Altman
1sNBS
U
s
. For a given distribution of users in
the network, the BSs have to solve the optimization problem
described in (1), where g
l
, 1 l N
l
are convex functions
which represents constraints and will be addressed later. We
978-1-61284-231-8/11/$26.00 2011 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2011 proceedings
assume that we do not control the scheduling strategy.
maximize U (1)
subject to 0 P
(b)
s
P
max
, 1 s N
BS
, 1 b N
b
and g
l
((P
(b)
s
)
1bN
b
) 0 , 1 s N
BS
, 1 l N
l
Let us denote by P the subset of (R)
N
b
NBS
which satises
the constraints in (1), and P
s
the subset of R
N
b
which satises
the constraints relative to BS s. It is noted that P = P
1
...
P
NBS
is a product of convex sets, hence convex.
B. Requirements on the algorithm
We rst impose the choice of a distributed algorithm which
ts the at architecture of future radio access networks. At
each step of the algorithm, BS s observes its utility U
s
,
exchanges some information with its neighbors and modies
its transmit powers (P
(b)
s
)
1bN
b
accordingly. Furthermore,
it is noted that nding the global optimum in (1) is NP-
hard ([9]) in most cases, hence we are simply concerned with
nding a local maximum of U in a few iterations with minimal
computing effort.
C. Basic Algorithm
We use the approach introduced in [10] for solving a
constrained optimization problem in a distributed fashion. Let
(t) R
N
b
NBS
, t N denote the power allocation of the
network at time t, and (0) P. We can then apply the
following gradient descent algorithm:
(0) P , (t + 1) =
_
(t) +
U((t))
_
+
(2)
where [.]
+
denotes the projection on P with respect to the
euclidean norm, and a constant step size. It is noted that the
projection is well dened since P is convex.
Furthermore, since P is a product of convex sets, we have
that (2) can be implemented in a distributed way. Let
s
(t)
R
N
b
, t N denote the power allocation of BS s, at time t,
we then obtain the following algorithm, for 1 s N
BS
:
s
(0) P
s
,
s
(t + 1) =
_
s
(t) +
s
U(
s
(t))
_
+
(3)
where
s
is the gradient with respect to (P
(b)
s
)
1bN
b
.
Furthermore, [10](Proposition 3.8, Page 219) gives the
following convergence theorem:
Theorem 1. If we assume that P is a product of convex sets,
that U is bounded below and that
U is Lipschitz continuous,
then there exists
0
such that if 0 <
0
then (3) converges
to a local maximum of U in P.
III. MULTI-CELL OFDMA SYSTEM MODEL
A. SINR
We consider a multi cell OFDMA network, and an FTP ser-
vice. Users arrive randomly according to a Poisson process of
rate , with a le of a given size, and leave the network when
they have completed the transfer. It is noted that some users
might be rejected because of admission control mechanisms.
Let us consider a user i and a BS s. The signal from s received
by i is proportional to the power emitted by s, and we dene
h
s,i
the proportionality coefcient, which is the product of
path loss and shadowing:
h
s,i
=
A
(d
s,i
)
s,i
(4)
where d
s,i
is the distance between i and s,
s,i
= 10
1
+
2
20
with
1
and
2
are two independent normally distributed random
variables with mean 0 and variance
Ns
h
s
,i
P
(b)
s
(5)
with N
2
0
the thermal noise.
B. Proportional Fair (PF) scheduler
We will use the following notations: let r
(p)
i,tm
denote the
instantaneous bitrate of user i at time t
m
on PRB p, and r
i,tm
the average allocated bitrate to user i during [t
0
, t
m
]. Let
denote a small constant averaging parameter. We write T
(p)
tm
=
i if user i was scheduled for transmission on PRB p at time
t
m
. The average bitrate is then calculated by the following
low-pass lter:
r
i,tm+1
= (1 )r
i,tm
+
Np
p=1
i,T
(p)
t
m+1
r
(p)
i,t
m+1
(6)
where is Kroneckers delta.
The PF scheduler chooses the user for transmission on PRB
p at time t
m+1
according to the following rule:
T
(p)
t
m+1
= arg max
i
r
(p)
i,tm+1
r
i,tm
(7)
We denote by r
i
the limit of r
i,tm
when t
m
+ ,
0
+
if it exists. For a proof of convergence of the PF scheduler,
the reader can refer to [11], and to [12] for the more general
case of the -fair scheduler.
C. Scheduling gain
We assume a Raleigh fading model: r
(p)
i,t
m
= (S
(p)
i
(p)
i,t
m
)
where
(p)
i,tm
is an exponentially distributed random variable of
mean 1, with
(p)
i,t
m
(p
)
i,t
m
, p = p
,
(p)
i,t
m
(p)
i
,tm
, i = i
and
(p)
i,tm
(p)
i
,t
m
, m = m
_
1
S
(b)
s,i
_
(8)
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2011 proceedings
with L
b=1
_
S
(b)
s,i
_
(9)
where N
u
(s) is the number of users present in s.
For the PF, the mean throughput can also be calculated in
closed form, see [13]:
r
i,PF
=
N
b
b=1
Nu(s)1
k=0
_
N
u
(s) 1
k
_
(1)
k
k + 1
_
S
(b)
s,i
k + 1
_
(10)
One must be careful however, since (10) is only exact when
the mean SINR of a user is the same on all PRBs, that is
S
(b)
s,i
= S
(b
)
s,i
, (b, b
(x) =
(x+d)
1
1
1
if = 1
and f
1
(x) = log(x + d) :
U
s
=
Nu(s)
i=1
f
(r
i
) (11)
We can now calculate the gradient of the network utility
with respect to the BSs power levels for both RR and PF
schedulers.
1) RR scheduler: We rst consider the derivative with
respect to the BSs own power level:
U
s
P
(b)
s
=
1
N
u
(s)
Nu(s)
i=1
1
(r
i,RR
+ d)
_
N
b
b=1
S
(b)
s,i
P
(b)
s,i
_
S
(b)
s,i
_
_
(12)
Now we consider the neighbors power level, if s
1
N
s
:
U
s
P
(b)
s1
=
1
N
u
(s)
Nu(s)
i=1
1
(r
i,RR
+ d)
_
N
b
b=1
h
s1,i
(S
(b)
s,i
)
2
h
s,i
P
(b)
s
_
S
(b)
s,i
_
_
(13)
2) PF scheduler: For a PF scheduler, the same type of
formulas can be obtained:
U
s
P
(b)
s
=
Nu(s)
i=1
1
(r
i,PF
+ d)
N
b
b=1
Nu(s)1
k=0
_
N
u
(s) 1
k
_
(1)
k
S
(b)
s,i
(k + 1)
2
P
(b)
s,i
_
S
(b)
s,i
k + 1
__
(14)
U
s
P
(b)
s
1
=
Nu(s)
i=1
1
(r
i,PF
+ d)
N
b
b=1
N
u(s)1
k=0
_
N
u
(s) 1
k
_
(1)
k+1
h
s1,i
(S
(b)
s,i
)
2
(k + 1)
2
h
s,i
P
(b)
s
_
S
(b)
s,i
k + 1
__
(15)
E. Constraints
We dene constraints on the maximal and minimal total
transmit power: g
1
((P
(b)
s
)
1bN
b
) =
Np
N
b
N
b
b=1
P
(b)
s
P
tot
and g
2
((P
(b)
s
)
1bN
b
) = P
tot
Np
N
b
N
b
b=1
P
(b)
s
where P
tot
is the maximum total transmit power and (0, 1] - the
minimum proportion of total transmit power. It is noted that
those functions are linear hence convex.
The constraint on the minimal transmit power has two
interests: rst, it prevents a numerical instability near 0 when
> 0, since the utility gradient becomes very large if a BS
transmits a total power of zero. The second interest is that
for close to 0, the unconstrained algorithm could result in
certain BSs transmitting very low power, causing a dramatic
increase in their load and BCR.
It is noted that xing a value of is akin to giving a
lower bound on the worst-case BCR. The justication is the
following: consider the case in which BS s transmits a total
power of P
tot
, and all its neighbors transmit at full power.
Then the BCR observed in BS s in this situation is an upper
bound for the BCR that could be observed in other situations
in BS s. Therefore, taking the maximum of this value on all
BSs gives an upper bound for the BCR observable on the
network.
Furthermore, increasing reduces the size of the constraints
set, reducing the maximum possible gains achievable by a
power control algorithm. This is why controls a trade-off
between BCR and Inter-Cell Interference Coordination (ICIC)
gains.
F. Implementation and signaling load
Every 1s, BS s calculates and forwards
Us
P
(b)
s
1
, 1 b N
b
to s
1
if s
1
is a neighbor of s. Hence, assuming N
b
= 3 bands,
6 neighbors for each BS and that each derivative is stored as
a 32-bits oating point number, the signaling load is of 576
bits/s per station, which is extremely small compared with the
expected capacity available on the X2 interface. Furthermore
the delay requirement of 1s is also easily satised, as a delay
between 20ms and 50ms is expected on the X2 interface.
IV. SIMULATION
A. Network Simulator
The algorithm is implemented in a large scale network
simulator with 39 BSs. The throughput allocated by the
scheduler is calculated in closed-form using equations (9) and
(10). Every 1s, the transmit power of each BS is adjusted
according to the power control algorithm described above. The
algorithm has been simulated for N
b
= 3. Three algorithms
are compared, using the following nomenclature:
Reuse 1 where all stations transmit at full power
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2011 proceedings
Soft Reuse which is a static power allocation described
in [16]
FFR which is the proposed dynamic algorithm
We choose = 2 since it implies maximizing the harmonic
throughput of BSs, which is a natural metric of capacity for
elastic trafc (see for example [17]) and gives good practical
results.
Because of the nite size of the network, we only calculate
the Key Performance Indicators (KPIs) on the subset of inner
BSs to minimize truncation effects, and the transient period
at the beginning of the simulation is not counted to calculate
KPIs. Simulation parameters are described in Table I.
Simulator parameters
Spatial resolution 25m 25m
Total simulated area 8km 8km
Time resolution 1s
Simulation time 3000s
User speed 5km/h
File size 10Mbytes
Number of sub-bands 3
Number of PRBs 9
Size of a PRB 180kHz
Number of stations 39
Cell layout 13 eNBs 3 sectors
5%
2
Maximum BS transmit power 30W
Service Type FTP
Scheduler Type Proportional Fair
Thermal noise 174dBm/Hz
Path loss 128 + 37.6 log
10
(d) dB, d in km
Shadowing standard deviation 6dB
TABLE I
MODEL PARAMETERS
B. Simulation results
All results are given for = 2, unless specied. Figures
1, 2 and 3 compare the BCR, mean le transfer time and
mean network throughput respectively for the three scenarios,
and we can see a clear improvement for the three KPIs.
The most notable is the BCR improvement from 5.5% to
2.3% in high trafc, demonstrating that the proposed algorithm
effectively reduces congestion in the network. Figure 4 shows
the cumulative distribution function (c.d.f) of the le transfer
time in the network for = 9, and we can see that all users
benet from the reduced congestion, the ones benetting the
most being cell-edge users, namely users with long transfer
times. Figure 5 shows the power allocated to each band
through time by the algorithm. Figure 6 shows the BCR and
the proportion of users whose File Transfer Time (FTT) is
longer than 10s as a function of , for = 0. It illustrates the
trade-off between the FTT reduction and increase in BCR.
Is allows the network operator to set the parameter in
order to enforce some policy, for example to obtain the best
performance while keeping the BCR under a certain threshold.
9 9.5 10 10.5 11 11.5 12
0
1
2
3
4
5
6
Arrival Rate (mobiles/s)
B
l
o
c
k
C
a
l
l
R
a
t
e
(
%
)
Reuse 1
Soft Reuse
FFR
Fig. 1. BCR of the network, = 2
9 9.5 10 10.5 11 11.5 12
10
12
14
16
18
20
22
Arrival Rate (mobiles/s)
M
e
a
n
t
r
a
n
s
f
e
r
t
i
m
e
(
s
)
Reuse 1
Soft Reuse
FFR
Fig. 2. Mean FTT in the network, = 2
9 9.5 10 10.5 11 11.5 12
0.95
1
1.05
1.1
1.15
1.2
1.25
x 10
5
Arrival Rate (mobiles/s)
M
e
a
n
n
e
t
w
o
r
k
t
h
r
o
u
g
h
p
u
t
(
k
b
p
s
)
Reuse 1
Soft Reuse
FFR
Fig. 3. Mean network throughput, = 2
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2011 proceedings
0 20 40 60 80 100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time (s)
F
i
l
e
t
r
a
n
s
f
e
r
t
i
m
e
c
.
d
.
f
Reuse 1
Soft Reuse
FFR
Fig. 4. c.d.f of FTT of all users in the network, = 2
0 200 400 600 800 1000
0
0.5
1
1.5
2
2.5
3
3.5
4
Time (s)
B
S
p
o
w
e
r
(
W
)
Band 1
Band 2
Band 3
Fig. 5. Transmit power of a BS, = 2
10 20 30 40 50 60 70 80 90 100
0
5
10
15
20
25
30
35
40
45
Gamma(%)
B
C
R
a
n
d
F
T
T
(
%
)
BCR
FTT > 10
target BCR
Fig. 6. Trade-off between FTT and BCR for = 0
V. CONCLUSION
This work has presented a distributed SON algorithm for
interference coordination in OFDMA networks. The algorithm
uses information available from neighboring cells and closed
form formulas, making it both computationally light and
suitable for practical implementation. It has been applied
to a large-scale network simulator, showing important gains
over a full power allocation, for cell-edge users, while not
degrading other KPIs. The trade-off between gains for cell-
edge users and increase in the BCR has been shown, and a
straightforward method for the network operator to manage it
has been provided.
REFERENCES
[1] R. Cendrillon, J. Huang, M. Chiang, and M. Moonen, Autonomous
spectrum balancing for digital subscriber lines, Signal Processing, IEEE
Transactions on, vol. 55, no. 8, pp. 4241 4257, aug. 2007.
[2] A. Gjendemsj, D. Gesbert, G. Oien, and S. Kiani, Binary power control
for sum rate maximization over multiple interfering links, Wireless
Communications, IEEE Transactions on, vol. 7, no. 8, pp. 3164 3173,
aug. 2008.
[3] A. Stolyar and H. Viswanathan, Self-organizing dynamic fractional
frequency reuse for best-effort trafc through distributed inter-cell
coordination, in INFOCOM 2009, IEEE, apr. 2009, pp. 1287 1295.
[4] J. woo Cho, J. Mo, and S. Chong, Joint network-wide opportunistic
scheduling and power control in multi-cell networks, Wireless Commu-
nications, IEEE Transactions on, vol. 8, no. 3, pp. 1520 1531, mar.
2009.
[5] T. Bonald, S. Borst, and A. Proutiere, Inter-cell scheduling in wireless
data networks, in in Proc. European Wireless, 2005, pp. 566572.
[6] K. Son, S. Chong, and G. Veciana, Dynamic association for load
balancing and interference avoidance in multi-cell networks, Wireless
Communications, IEEE Transactions on, vol. 8, no. 7, pp. 3566 3576,
jul. 2009.
[7] G. Wunder, M. Kasparick, A. Stolyar, and H. Viswanathan, Self-
organizing distributed inter-cell beam coordination in cellular networks
with best effort trafc, in Modeling and Optimization in Mobile, Ad
Hoc and Wireless Networks (WiOpt), 2010 Proceedings of the 8th
International Symposium on, may. 2010, pp. 295 302.
[8] ITU-RM.2134, Requirements related to technical performance for imt-
advanced radio interface(s), Tech. Rep., nov 2008.
[9] J. Papandriopoulos and J. Evans, Low-complexity distributed algo-
rithms for spectrum balancing in multi-user dsl networks, Communi-
cations, 2006. ICC 06. IEEE International Conference on, vol. 7, pp.
3270 3275, jun. 2006.
[10] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and distributed computation:
numerical methods. Upper Saddle River, NJ, USA: Prentice-Hall, Inc.,
1989.
[11] H. Kushner and P. Whiting, Convergence of proportional-fair sharing
algorithms under general conditions, IEEE transactions on wireless
communications, vol. 3, pp. 12501259, July 2004.
[12] R. Combes, Z. Altman, and E. Altman, On the use of packet scheduling
in self-optimization processes: application to coverage-capacity opti-
mization, in WiOpt 2010, Avignon, France, Jun. 2010.
[13] B. Blaszcyszyn and M. Karray, Fading effect on the dynamic per-
formance evaluation of ofdma cellular networks, in 1st International
Conference on Communications and Networking, 2009.
[14] R. Combes, Z. Altman, and E. Altman, A self-optimization method
for coverage-capacity optimization in ofdma networks with mimo, in
Accepted in Value Tools 2011.
[15] J. Mo and J. Warland, Fair end-to-end window based congestion
control, IEEE transactions networking, vol. 8, pp. 556566, October
2000.
[16] T. Bonald and N. Hegde, Capacity gains of some frequency reuse
schemes in ofdma networks, in Global Telecommunications Conference,
2009. GLOBECOM 2009. IEEE, dec 2009, pp. 16.
[17] T. Bonald and A. Prouti` ere, Wireless downlink data channels: user
performance and cell dimensioning, in Proceedings of the 9th annual
international conference on Mobile computing and networking, ser.
MobiCom 03. New York, NY, USA: ACM, 2003, pp. 339352.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2011 proceedings