Академический Документы
Профессиональный Документы
Культура Документы
6, JUNE 2013
Abstract—In this paper we present a novel distributed Inter- employment of discrete [7] rather than continuous [8] power
Cell Interference Coordination (ICIC) scheme for interference- control has become a promising and effective way to control
limited heterogeneous cellular networks (HetNet). We reformu- and mitigate inter-cell interference. However, because a multi-
late our problem in such a way that it can be decomposed into
a number of small sub-problems, which can be solved indepen- level power control technique requires large information ex-
dently through an iterative subgradient method. The proposed change [7], it may not be appropriate in a random deployment
dual decomposition method can also address problems with of small-cells. Another disadvantage of some distributed solu-
binary-valued variables. The proposed algorithm is compared tions [7], [8] in the literature is that they have a non-negligible
with some reference schemes in terms of cell-edge and total cell performance gap compared with a centralized algorithm due to
throughput.
their convergence to a local optimum. Furthermore, they usu-
Index Terms—Inter-cell RRM, interference avoid- ally improve only one performance metric at a time, i.e., either
ance/coordination, ICIC, linear programming optimization, dual
the cell-edge throughput or the total cell throughput [6]. In this
decomposition.
paper, we present a distributed ICIC algorithm applicable to
HetNet cellular deployments, which is solved via a novel dual
I. I NTRODUCTION decomposition approach. Most dual decomposition methods
problem’ and the ‘transmit power problem’. The assignment not transmit (ρi,m=1 = 0). The optimization problem in (4)
problem decides which user is to be scheduled to which RB. will maximize the network utility based on the user selection
The transmit power problem decides the level of transmit in (3b).
power P i and interfering transmit power P j that best satisfies
the given constraint. The complication regarding the P i and B. Linear problem formulation and relaxation
P j is as follows; once an optimal P i is reached, it may no
longer be optimal when the P j has adjusted to the changes In order to reduce the complexity involved with the non-
resulting from inter-cell interference. linear expression in (4c), we may reformulate it to this:
ρ,m = ρj,m + ρk,m . (5)
j,k∈
A. Problem formulation
Now the reformulated problem is a special case (i.e., a problem
In a multi-user interference-limited system, the binary with totally unimodular matrix specifications) whose relaxed
power allocation policy maximizes the total data rate [11]. solution is also the optimal solution to the initial problem.
As the transmit power of each RB can be allocated to either This means that there is a small integrality gap2 between the
maximum transmit power or zero transmit power, we may reformulated problem and its relaxation. Linear programming
substitute the two complex variables P i and P j with one relaxation is a problem that arises when binary variables
joint equivalent variable. Therefore, we consider only these are replaced with real variables belonging to interval [0, 1].
two cases (transmit/not transmit) to calculate the achievable Therefore, the variables in (4d) may be relaxed as:
user data rate. We assume that the served users are able to
ρi,m , ρj,m ,ρk,m , ρm ∈ [0, 1]. (6)
estimate the separate levels for strong sources of interference m
by employing cell-specific orthogonal reference sequences in The small integrality gap of the reformulated problem can be
the LTE standard [1]. In addition, let m ∈ M denote the reduced by introducing a number of tighter cuts3 as follows:
index up to which (0th , 1st , 2nd , ) dominant interfering eNB
ρ ≤ b, ρ ∈ [0, 1], ∈ S, b = || /2 , (7)
is mitigated. Then, let U denote a utility metric which is
formulated as follows: where is the set of eNBs where the cutting plane method is
i,m i,m r̂ applied and b is the lower integral part of the cardinality (| |)
Uk,n = rk,n · dk ; dk = ; m ∈ M, (3a)
rˆk of divided by 2. Finally, a slack variable y may be added to
where r̂ is the average throughput across all UEs1 , rˆk is the the inequality constraint in (7) to transform it to an equality
average throughput of UE k and dk is the user demand to as follows:
provide a network-wide fairness. In this way, the utility metric ρ + y = b, y ∈ [0, 1]. (8)
i,m
Uk,n can describe the additional gain by mitigating up to
the m dominant interfering eNBs. Afterwards, the assignment
IV. P ROPOSED D UAL D ECOMPOSITION M ETHOD
problem for each RB n is solved by selecting the best user
K with the highest utility price as
follows:
One main disadvantage of the semi-centralized problem is
that is still highly dependent upon a centralized processing en-
K = arg max Uki,m . (3b)
k∈K tity which requires a large amount of the network information
Next, the transmit power problem is formulated for each RB exchange. Therefore, we may distribute further the complexity
n as follows: of the relaxed semi-centralized problem in (4) into a number
maximize U i,m · ρi,m ; (4a) of independent sub-problems via dual decomposition [9] as
i∈S m∈M follows: i,m i,m i,m i,m ,m
subject to ρ i,m ,m
+ρ = 1; (4b) U ·ρ −λ ρ +ρ −1
L(ρ, λ)= m∈M ;
,m
ρ = min ρj,m + ρk,m , 1 ; (4c) i∈S −λ ρ + y − b
z i
j,k∈
i
i,m j,m k,m
ρ ,ρ ,ρ , ρm ∈ {0, 1}. (4d) = Li ρi,m , y i , λi,m , λz + λ + λz · b ,
m i∈S i∈S
(9)
i,m
Let ρ denote the binary variable whether the eNB i may where λi , λz > 0 is the Lagrange multipliers or dual variables
transmit (ρi,m = 1) or not (ρi,m = 0) in the case where up associated with the constraints in (4b) and (8), respectively. Let
to the m dominant interfering eNBs are mitigated. Also, let z denote the index of the cutting plane equations. Therefore,
i,m denote the set of different eNBs where the eNB i is in λz is the dual variable associated with the equation z. λi,m is
conflict when they mitigate up to the m dominant interfering calculated using the dual variable λi and λj with respect to
eNBs. For example, i,m = {j, k} infers that both eNB j, k ∈ i,m as follows:
S find the eNB i as their 1st dominant interfering eNB. Note λi m=0
that for simplicity, the i,m is referred to as . Consequently, λi,m = λj m = 0 . (10)
eNB j and eNB k may transmit in the case where m = 1 j∈(i,m)
(ρ,m=1 = 1); when their dominant interfering eNB i does The main network problem is now reformulated as:
1 The user demand may also be defined by using the average throughput 2 The Integrality gap is defined as the maximum ratio between the solution
across served UEs, however this provides only local fairness. In order to avoid quality of the binary/integer problem and of its relaxation
inter-cell signaling, the average throughput across all UEs may be a constant 3 Cuts are a number of additional constraints that may be added to the
network parameter, which can be optimized over time. relaxed problem in order to restrict fractional solutions.
1146 IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 6, JUNE 2013
TABLE I TABLE II
M AIN S YSTEM S IMULATION PARAMETERS P ERFORMANCE IN M ACROCELLS (G AIN /L OSS )
the independent subproblem Li and are unique due to strict REF2 - SEMI ∞ 21.10 (17.0%) 1496 (2177%)
20 21.03 (16.6%) 1470 (2138%)
concavity of the log-transformed U which can be calculated
as:
10 20.91 (15.9%) 1447 (2102%)
i,m i,m U i,m·ρi,m−λi,m ρi,m+ρ,m−1 5 20.7 (14.8%) 1422 (2064%)
ρ∗ λ arg max . PROP
& = (C) −λz ρ + y i − b (12)
3 20.45 (13.4%) 1417 (2057%)
m 2 20.11 (11.5%) 1411 (2047%)
y∗i λi (C) : ρi,m ∈ [0, 1]; y i ∈ [0, 1]; ρ = 1, 1 19.48 (7.99%) 1224 (1763%)
m∈M
REF1- FR1 0 18.04 65.7
To calculate λ, the following subgradient may be used,
λi (t + 1) = λi (t)−β·g i + ; λz (t + 1) = λz (t) − β · g i +, (13)
low-power HeNB (Home eNB) network is implemented to
where [ ]+ is the projection on the non-negative orthant and
simulate a HetNet scenario. Table I gives the main simulation
is a positive iterative step size.
parameters used. We observe two performance metrics, i.e.
The advantage of using the dual decomposition method to
the cell throughput and the 5th percentile (5th -ile) point
distribute the semi-centralized algorithm is to limit the inter-
of Cumulative Distribution Function (CDF) of UE through-
cell signalling required to converge the algorithm.
put. Table II and Table III present the performance of the
Theorem 1: The iterative algorithm will converge to the
proposed algorithm (‘P ROP ’) for macrocells and femtocells,
semi-centralized solution with a large number of iterations.
respectively, where some comparison (reference) schemes are
Proof : The log-linear operating mapping function in (2) is
also included. The ‘R EF 1’ employs full frequency reuse and
i
convex [10], so a unique rk,n ∗ exists for γk,n
i
. In a similar way,
its performance is equivalent when the number of iterations
the solution of the log-transformed U in (12) is unique. Then,
executed by our proposed algorithm is set to zero. Because
if the iteration step size β is sufficiently small, as t → ∞,
its performance is increased if the number of iterations is
the duality gap between the relaxed problem and its dual will
also increased, this relative gain is displayed in the table
converge to zero.
with respect to the ‘R EF 1’. On the other hand, ‘S EMI ’ is
Algorithm 1: To be executed by eNB i for each RB n.
the semi-centralized solution without the dual decomposition.
Initialization: Calculate the U i,m , ρ , b, λi , λi,m and λz .
According to the theorem 1, the gap between ‘S EMI ’ and the
For each iteration 1, ..., Maximum iteration step :
proposed iterative scheme should converge to zero for a large
• Solve the equation in (12).
number of iterations (infinity). Other comparison schemes
• Calculate subgradient step in (13).
include: dynamic FR3 (‘DYN FR3’), the centralized solution
• Update the subgradients ( λi and λz ).
(‘C ENTRAL’) as proposed in [3], the algorithms ‘Z HANG ’
• Exchange subgradients to all eNBs involved.
[8] and ‘A BAII ’ [9] which can also be implemented in a
Termination: Stop if the maximum iteration step is reached.
distributed way. The latter two algorithms employ inter-cell
power control (continuous [7] or discrete [8]) to achieve a
V. S IMULATION S TUDY & R ESULTS maximum radio resource utilization and quality of service
The simulation study is performed in the downlink using (QoS). As mentioned earlier, performance degradation can
a modified version of the LTE-based system-level simulator be seen in dense interference-limited systems using a non-
[12]. Apart from the outdoor eNB network, a closed-access binary power allocation plan. This can be observed particularly
KOSTA et al.: A DISTRIBUTED METHOD OF INTER-CELL INTERFERENCE COORDINATION (ICIC) BASED ON DUAL DECOMPOSITION . . . 1147
TABLE IV
S IGNALING & C OMPLEXITY C OMPARISON (P ER I TERATION ) algorithm solves the problem iteratively by employing a re-
vised dual decomposition method which can address binary
INFORMATION EXCHANGE REQUIRED variables. Simulation activities show the effectiveness of the
ALGORITHM
OVER X2 INTERFACE COMPUTATIONS
ZHANG [7] 4 Variables (d k , Pi, L 1 , L 2 ) |K| + |S|2 algorithm with a small number of iterations especially in the
PROP 2 Variables (Ȝ, ȡȚ,m) |M|+1 case of femtocell, which would otherwise suffer heavily from
inter-cell interference due to their asymmetric deployment.
in the case of femtocells. Additionally, it can be observed
that their performances are proposed for macro deployments. A PPENDIX A
However, we can also observe that the compared distributed D ECOMPOSITION OF THE C ENTRALIZED RRM P ROBLEM
algorithms may converge to a local optimum rather than the The complexity of a centralized problem may be shifted
global optimum. Our algorithm shows its performance both to other network entities (i.e. eNBs) at the cost of inter-
in macro and femto deployments. The relative gain indicates cell signaling. Given that the intra-cell (adjacent-channel)
the adaptability of the algorithm to the current deployment. interference is avoided in OFDMA systems, our problem is
In the case of femtocells, 1 or 2 iterations may be sufficient limited to the inter-cell (co-channel) interference. For sim-
to mitigate the detrimental inter-cell interference. However, in plicity of illustration, the notation of the RB n is omitted.
the case of macrocell performance improvement can be seen Therefore, the centralized problem, which includes all the
with a higher number of iterations which can vary from 3 to 5. network information, is as follows:
The ‘C ENTRAL’ scheme shows the peak performance that can i,m i,m
be achieved by using a centralized algorithm. By assuming a Umax = max U k · ρk
i∈S m∈M k∈K
semi-centralized approach, only a negligible performance gap
can be seen (between ‘C ENTRAL’ and ‘S EMI ’). This is due
to the small correlation of the ‘assignment problem’ with the Umax = max max U i,m
·ρ i,m
Assignment i,m
=Uki,m i∈S m∈M
‘power transmit problem’. Problem→ U
k∈K
Power transmit problem
VI. C OMPLEXITY A NALYSIS The centralized problem is equal to the maximization of
We provide a brief overview of the complexity and signaling the maximized subproblems as shown above. We may reduce
overhead of the evaluated algorithms. For simplicity, we do the signaling overheads with minimum loss of optimality by
not account for the complexity in channel status information selecting the best user (i.e. K = K ) in each power transmit
since all schemes require this information. By default, all problem with the argmax in (3b).
centralized algorithms (i.e. ‘C ENTRAL’ & ‘DYN FR3’) must R EFERENCES
have a common processing entity in order to reach the final
[1] Sesia, Toufik, Baker, LTE - The UMTS Long Term Evolution: From
solution and distribute this to all parts of the network. The Theory to Practice. Wiley, 2009.
distributed algorithms in [7], [8] are implemented without a [2] S. Martello and P. Toth, Knapsack Problems: Algorithms and Computer
centralized entity at the cost of excessive inter-cell signaling. Implementation. John Wiley and Sons, 1990.
[3] M. Rahman and H. Yanikomeroglu, “Inter-cell interference coordination
As a result, it is required to compute multiple sub-versions in OFDMA networks: a novel approach based on integer programming,”
of the centralized problem (own cell & neighboring cells). in Proc. 2010 IEEE VTC – Spring, pp. 1–5.
For example, the distributed algorithm in [7] exchanges four [4] M. Rahman and H. Yanikomeroglu, “Enhancing cell-edge performance:
variables with all eNBs over X2 interface (i.e. user demand a downlink dynamic interference avoidance scheme with inter-cell
coordination,” IEEE Trans. Wireless Commun., vol. 9, no. 4, pp. 1414–
dk , eNB transmit power P i and two SINR components (L1 1425, Apr. 2010.
& L2 )) in order to reconstruct the weighted sum rate (utility [5] M. Pischella and J.-C. Belfiore, “Distributed weighted sum throughput
metric). Furthermore, it is required to solve two sub-problems maximization in multi-cell wireless networks,” in Proc. 2008 IEEE
PIMRC, pp. 1–5.
at each iteration (i.e. user assignment and transmit power [6] S. G. Kiani, G. E. ien, and D. Gesbert, “Maximizing multicell capacity
(on/off) which requires |K| and |S|·2 computations, respec- using distributed power allocation and scheduling,” in Proc. 2007 IEEE
tively). Table IV gives a signalling and complexity comparison WCNC, pp. 1690–1694.
[7] H. Zhang, L. Venturino, N. Prasad, L. Peilong, S. Rangarajan, and
of the proposed dual algorithm and Zhang’s algorithm in X. Wang, “Weighted sum-rate maximization in multi-cell networks via
terms of information exchange and required computations (per coordinated scheduling and discrete power control,” IEEE J. Sel. Areas
iteration). The complexity and the signaling overhead of our Commun., vol. 29, no. 6, pp. 1214–1224, Jun. 2011.
[8] M. Abaii, Y. Liu, and R. Tafazolli, “An efficient resource allocation
distributed algorithm is reduced to a minimum by collecting strategy for future wireless cellular systems,” IEEE Trans. Wireless
only the necessary information in order to converge to the Commun., vol. 7, no. 8, pp. 2940–2949, Aug. 2008.
semi-centralized solution. Note, that the network-wide user [9] W. Yu and R. Lui, “Dual methods for nonconvex spectrum optimization
of multicarrier systems,” IEEE Trans. Commun., vol. 54, no. 7, pp.
demand dk is incorporated into the dual λ variable. Each eNB 1310–1322, Jul. 2006.
solves its own problem with respect to this exchanged variable [10] D. Leith, V. Subramanian, and K. Duffy, “Log-convexity of rate region
which requires |M | + 1 computations (including the case ‘do in 802.11e WLANs,” IEEE Commun. Lett., vol. 14, no. 1, pp. 57–59,
Jan. 2010.
not transmit’). [11] A. Gjendemsjoe, D. Gesbert, G. Oien, and S. Kiani, “Binary power
VII. C ONCLUSION control for sum rate maximization over multiple interfering links,” IEEE
Trans. Wireless Commun., vol. 7, no. 8, pp. 3164–3173, Aug. 2008.
In this paper, a novel distributed ICIC has been presented [12] J. C. Ikuno, M. Wrulich, and M. Rupp, “System level simulation of LTE
for heterogeneous multi-cell cellular networks. The proposed networks,” in Proc. 2010 IEEE VTC – Spring, pp. 1–5.