142 ieee transactions on wireless communications, β¦
TRANSCRIPT
142 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 1, JANUARY 2011
Utility-Optimal Multi-Pattern Reuse inMulti-Cell Networks
Kyuho Son, Member, IEEE, Yung Yi, Member, IEEE, and Song Chong, Member, IEEE
AbstractβAchieving sufficient spatial capacity gain throughthe use of small cells requires careful consideration of inter-cell interference (ICI) management via BS power coordinationcoupled with user scheduling inside cells. Optimal algorithms areknown to be difficult to implement due to high computation andsignaling overhead. This study proposes joint pattern-based ICImanagement and user scheduling algorithms that are practicallyimplementable. The key idea is to decompose the originalproblem into two sub-problems in which ICI management isrun at a slower time scale than user scheduling. We empiricallyshow that even with such a slow tracking of system dynamicsat the ICI management part, the decomposed approach achievesa considerable performance increase compared to conventionaluniversal reuse schemes.
Index TermsβInter-cell interference (ICI), multi-pattern reuse,ICI management, user scheduling, time-scale decomposed algo-rithm, multi-cell network, network utility maximization (NUM).
I. INTRODUCTION
TO achieve high spatial capacity, wireless cellular net-works consider a dense deployment of base stations
(BSs) that cover small cells. As a consequence, inter-cell inter-ference (ICI) from neighboring BSs becomes a major sourceof performance degradation, and the portion of users whosecapacity is limited by ICI grows. To attain the full potentialgain of multi-cell networks, coordinating the transmissionsamong BSs to manage ICI effectively is essential. The keyintuition of BS coordination is that the achievable rates, whichdepend on the amount of ICI, can be increased by adaptivelyturning off some of the neighboring BSs. Thus, there arecases in which the increment of achievable rates preponderatesthe sacrifice of taking away transmission opportunities atneighboring BSs. In particular, this effect of ICI managementis very apparent for users at the edges of cells.
A brute-force approach for mitigating ICI involves the useof a system-wide reuse scheme in the time and/or frequencydomain. However, this may waste precious radio resourcesbecause users at different geographical locations inside cellsprefer different reuse schemes. Several schemes, e.g., frac-tional frequency reuse (FFR) [1] in Mobile WiMAX, have
Manuscript received December 5, 2009; revised May 16, 2010 andSeptember 7, 2010; accepted September 19, 2010. The associate editorcoordinating the review of this paper and approving it for publication wasE. Hossain.
K. Son is with the Department of Electrical Engineering, University ofSouthern California, Los Angeles, CA, 90089 (e-mail: [email protected]).
Y. Yi and S. Chong are with the Department of Electrical Engineering,Korea Advanced Institute of Science and Technology (KAIST), Daejeon 305-701, Korea (e-mail: [email protected]; [email protected]).
This work was supported by the IT R&D program of MKE/KEIT[KI002137, Ultra Small Cell Based Autonomic Wireless Network].
Digital Object Identifier 10.1109/TWC.2010.110310.091778
been proposed to accommodate users in different channelconditions with different reuse factors. However, these a-priorihand-crafted schemes are still far from optimal in the sensethat they do not adapt to dynamic network environments, e.g.,time-varying user loads/locations. In addition, user schedulingworking opportunistically based on perceived time-varyingchannels must be considered in conjunction with ICI man-agement to achieve a high performance gain.
This paper (1) investigates the coupling dynamics of inter-cell ICI management and intra-cell user scheduling and (2)proposes practically implementable joint ICI management anduser scheduling algorithms in multi-cell networks. To thatend, a pattern-based joint optimal algorithm that tracks time-varying channel conditions is initially proposed, where βpat-ternβ corresponds to a combination of BS ON/OFF activities.It is then demonstrated that the proposed optimal algorithmis difficult to implement due to its high complexity. The keybottleneck lies in the ICI management part, which requirescollecting excessive amount of feedback information from allusers and also needs complex operations to make decisionson BS coordination at every time slot. To overcome suchcomplexity, the original optimization problem is decomposedinto two sub-problems (user scheduling and pattern-basedICI management) and these are solved them with differenttime scales. The complexity becomes much lower than thatof the optimal algorithm, yet sustains high efficiency in ICImanagement.
The algorithm based on time-scale decomposition stemsfrom a design rationale in which ICI management may nothave to track fast dynamics, e.g., a fast fading channelcondition. Instead, it may suffice to run the ICI managementscheme following only macroscopic network changes, e.g.,user loads/locations, and their average channel conditions.In spite of such slow tracking of system dynamics in ICImanagement, with the proposed decomposed algorithms, it isempirically shown that the performance increase amounts toabout 6βΌ20% (compared to a conventional universal reusescheme), corresponding to 1/2βΌ2/3 of the optimal algorithm(which is practically impossible to implement).
Research pertaining to mitigating ICI has recently receivedmuch attention [2]β[7]. Optimal binary power control (BPC)for sum rate maximization was considered in [2]. In otherstudies [3], [4], optimal joint ICI management (similar BPC)and user scheduling algorithms that operate in a slot-by-slot manner and require heavy computation overhead wereconsidered. The authors there presented the idea of usingclustering only with neighboring BSs [3] or considering onlyneighboring BSs [4] to reduce complexity. However, these
1536-1276/11$25.00 cβ 2011 IEEE
SON et al.: UTILITY-OPTIMAL MULTI-PATTERN REUSE IN MULTI-CELL NETWORKS 143
schemes continue to require centralized coordination and/orcomplex operations on a per-slot basis, which hinders practicalimplementation.
Several recent approaches [6], [7] have attempted to makealgorithms practical based on a slightly different time-scaleseparation approach from that used in the this paper. Inone recent study [6], the authors abstract users that sharesimilar traffic loads and channel environments into classesand perform ICI management on a very long time scale(e.g., hours) without explicit consideration of intra-cell userscheduling. They basically design an ICI management schemethat tracks system dynamics at a highly macroscopic level.Our approach differs from the aforementioned scheme [6]in that user scheduling is explicitly considered. Moreover,our ICI management runs much faster (e.g., in the orderof seconds) compared to the earlier work [6]. The workwith a time-scale separation similar to ours was recentlyproposed in [7] for different systems, i.e., OFDMA systems,where the transmit power level for different subbands for ICImanagement is periodically updated. Due to the differencesin the system model, a different mechanism is used here thatupdates patterns and not powers, leading to a different stylefor the algorithms and analysis. Additionally, the performancegap between optimal and decomposed algorithms is studied inthis paper.
Related work also includes an examination of the potentialcapacity gains (from the perspective of the flow-level per-formance) by BS coordination [8]. Another important issuein multi-cell networks is to resolve load imbalance problembetween cells. Several investigations [9], [10] explicitly bal-ance the load by changing user associations from the BSin hot-spot cells to an adjacent BS that is less crowded.Sang et al. [9] proposed an integrated framework consistingof a MAC-layer cell breathing technique and load-awarehandover/cell-site selection to deal with load balancing. Bu etal. [10] were the first to consider the formulation of network-wide proportional fairness (PF) [11] rigorously in a multi-cell network where associations between users and BSs aredecision variables. Although it is assumed in this paper thatuser association is fixed, we later argue and empirically showthat ICI management can implicitly resolve the load imbalance.Moreover, it is shown that any performance gain by controllinguser associations may not be significant.
The remainder of this paper is organized as follows. SectionII formally describes our system model and presents a defi-nition of the problem. Section III introduces a joint optimalpattern selection and user scheduling algorithm to solve thisproblem and discuss its implementation difficulties. In order totake into account practical concerns, two algorithms (patternportion change algorithm and user scheduling algorithm) usingtime-scale decomposition schemes that run at different time-scales are designed in Section IV. Section V includes ademonstration of the performance of proposed algorithms, andthe paper is concluded with Section VI.
II. SYSTEM MODEL AND PROBLEM DEFINITION
A. Network Model
We consider a wireless cellular network consisting of mul-tiple cells. Denote by π© .
= {1, . . . , π} and π¦ .= {1, . . . ,πΎ}
a set of BSs and MSs (or users), respectively. A user π β π¦is associated with a single BS π β π© , which means that dataintended for the user π is served only by the BS π. Defineπ(β ) : π¦ β π© to be the association function, e.g., π(π) = πif the user π is associated with the BS π. We further denoteby π¦π the set of users associated with the BS π.
We assume that a BS transmits data with either its givenmaximum power or 0, which we simply denote by βONβ orβOFFβ states1. We assume that a same frequency band (orchannel in short) with bandwidth π in all cells, and con-sider only downlink transmissions in the time-slotted systemindexed by π‘ = 0, 1, . . .. At each slot, a BS can select onlyone user for its data transmission. Channels may be time-varying, modeled by some stationary, ergodic random processwith the finite state index set β and the stationary distributionπ½ = (π(π), π β β).
B. Network Resource and Allocation Schemes
The time-varying network resources at slot π‘ are representedby a finite set β(π‘) of the πΎ-dimensional feasible rate(bits/slot) vectors over users. A resource allocation schemethen chooses a feasible rate vector in β(π‘) at each slot andserves a subset of users with the chosen rate vector. A feasiblerate vector in β(π‘) is determined by the following two factors:(i) which BSs are activated and (ii) which users are selectedin cells for data transmission.
To formally discuss (i), we define reuse pattern (or simplypattern) π to be a combination of ON/OFF activities of BSs,which determines inter-cell interference to the correspondingscheduled users in cells. Denote by π« the set of all patterns.A pattern π is said to activate a BS π, if the activity of theBS π is ON under pattern π. Denote by π©π β π© the set ofall BSs activated by the pattern π. In parallel, we denote byπ«π β π« the set of patterns that activate the BS π. Define reusefactor of a pattern π to be ππ
.=
β£π©πβ£π β€ 1, i.e., the ratio of the
number of BSs which use a pattern π to the total number ofBSs. Denote by ππ(π‘) the pattern selection indicator for thepattern π, i.e., ππ(π‘) = 1 when the pattern π is used at slot π‘,and 0 otherwise. Then, since only one pattern is used per oneslot, we should have: β
πβπ«ππ(π‘) = 1. (1)
In regard to (ii), define user scheduling indicator at slot π‘by πΌπ(π‘), i.e., πΌπ(π‘) = 1, when the user π is scheduled in itsassociating cell, and 0 otherwise. Reflecting the constraint thatonly one user can be selected in each cell, we should have:
βπβπ¦π
πΌπ(π‘)
{β€ 1, if ππ(π‘) = 1 and π β π©π ,= 0, otherwise.
(2)
Then, a resource allocation scheme incorporates patternselection and user scheduling that can be regarded as choosinga sequence of
{(ππ(π‘) : π β π«), (πΌπ(π‘) : π β π¦)}β
π‘=0satisfying the constraints (1) and (2).
We now define the transmission rates of users providedby a resource allocation scheme, depending on the choice of
1All discussions in this paper can be readily extended to the case of afinite number of discrete power levels.
144 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 1, JANUARY 2011
patterns. Let πΊπ,π(π‘) represent the time-varying channel gainfrom BS π to user π at slot π‘. The channel gain may takeinto account path loss, log-normal shadowing, fast fading andetc. The received SINR for user π at slot π‘ when pattern π isselected and user π is served by its serving BS, can be writtenas:
Ξππ(π‘)=
{πΊπ(π),π(π‘)π
πππ₯π
π0π +β
πβπ©π,πβ=π(π) πΊπ,π(π‘)ππππ₯π
, if π(π) β π©π,
0, otherwise,
where ππππ₯π is the maximum transmit power of BS π and π0
is the noise spectral density. Following the Shannonβs formula,the data rate for user π on reuse pattern π at slot π‘ is givenby:
πππ(π‘) = π log2 (1 + Ξππ(π‘)) .
Note that πππ(π‘) = 0 for all π(π) /β π©π, i.e., user π cannotreceive any data rate if its serving BS π(π) is not activatedby the pattern π. Also notice that πππ(π‘) is the potential datarate when the user π is scheduled, i.e., its actual data ratemay become 0, when other user, say πβ², associated with theBS π(π), is scheduled for service. We assume that each BS πknows instantaneous achievable data rates for all its associatedusers through channel feedbacks. We further assume that BSshave infinite amount of data to be destined to users.
C. General Problem Statement
In this paper, we aim at proposing the joint pattern selectionand user scheduling that maximizes the long-term network-wide utility whenever possible, i.e., solves the followingoptimization problem Q:
Q: max π =βπβπ©
π (π) =βπβπ¦
ππ(οΏ½οΏ½π)
subject to R β β,
where R = (οΏ½οΏ½π, π β π¦) is the vector of long-termuser throughputs. The network-wide utility π is just thesummation of utilities of all BSs (π (π), π β π© ); π (π) isagain the summation of utilities of all its associated usersπ (π) =
βπβπ¦π
ππ(οΏ½οΏ½π). Assume the standard condition ofdifferentiability and strictly increasing concavity of ππ(β ). Weadopt the generalized (π€,πΌ)-fair utility function introduced in[12]:
ππ(οΏ½οΏ½π) =
{π€π log οΏ½οΏ½π, if πΌ = 1,π€π(1β πΌ)β1οΏ½οΏ½1βπΌ
π , otherwise,(3)
where πΌ and π€π are positive. By varying the πΌ parameter, itencompasses various notions of fairness, in particular, propor-tional fairness (πΌ = 1) and max-min fairness (πΌββ).
The set β β βπΎ+ of all achievable rates of users is
referred to as achievable rate region. First, denote by β(π) theachievable rate when the system is in the π-th channel state.The β(π) is essentially the convex hull of the set of feasiblerates for the π-th channel state, i.e.,
β(π) ={R(π) = (οΏ½οΏ½
(π)π : π β π¦) β£β πππ(π) β Ξ , οΏ½οΏ½
(π)π =
βπβπ«
π(π)ππ π
(π)ππ
},
(4)
where π(π)ππ is the long-term fraction of time that user π is
served under pattern π for the π-th channel state; Ξ is the setof nonnegative vectors πππ(π) = (π
(π)ππ : π β π¦, π β π«) such thatβ
πβπ«π(π)π = 1 and
βπβπ¦π
π(π)ππ β€ π(π)
π , βπ β π© , βπ β π«π.
(5)Then, the β is characterized by β =
βπββ π(π)β(π), where
the addition of sets is defined as follows: π³ + π΄ = {π₯ + π¦ :π₯ β π³ , π¦ β π΄}. Thus, we can characterize the achievable rateregion β by:
β ={R = (οΏ½οΏ½π : π β π¦) β£ β R(π) β β(π),
οΏ½οΏ½π =βπββ
π(π)οΏ½οΏ½(π)π =
βπββ
βπβπ«
π(π)π(π)ππ π
(π)ππ
}.
(6)
III. OPTIMAL ALGORITHM
In this section, the structure of optimal solutions is studiedanalytically for a simple scenario to gain insight and anoptimal pattern selection and user scheduling algorithm thatgenerates the optimal solution is described.
A. Structure of Optimal Solution for Symmetric Networks withStatic Channels
For general networks, it is quite difficult to characterize theoptimal fractions of time for user-patterns (π
(π)ππ : π β π¦, π β
π« , π β β). However, we will show that it is indeed possible toexplicitly characterize them for symmetric networks with staticchannels. A network is said to be symmetric if all BSs havethe same number of users, and their channel characteristicsare identical. Fig. 1 depicts an illustrative example of a lineartwo-cell network having three patterns where (π1, π2, π3) =(1, 0.5, 0.5). Recall that ππ, the reuse factor of pattern π, is theratio of the number of BSs activated by pattern π to the totalnumber of BSs. Since the network is symmetric, it is enoughto analyze the following optimization problem Q-symmetricfor a typical BS, say BS 1:
Q-symmetric:
max(πππ:πβπ¦1,πβπ«1)
π (1) =βπβπ¦1
ππ(οΏ½οΏ½π) (7)
subject toβπβπ«1
βπβπ¦1
πππ
ππβ€ 1, (8)
πππ β₯ 0, βπ β π¦1, βπ β π«1, (9)
οΏ½οΏ½π =βπβπ«1
ππππππ, βπ β π¦1. (10)
Here, the constraint (8) originally comes from the condition(5) on πππ. By additionally applying the symmetric conditionthat the pattern having the same reuse factor should havethe same pattern portion to (5), we can readily derive theconstraint (8). For example, in the two-cell network case, thederivation can be done as follows:
1 =β
πβπ« ππ = π1 + π2 + π3
= π1 + 2π2 = π1/π1 + π2/π2 (β΅ π2 = π3 by symmetry)β₯ β
πβπ¦1
ππ1
π1+β
πβπ¦1
ππ2
π2=
βπβπ«1
βπβπ¦1
πππ
ππ.
The problem Q-symmetric has an interesting structure ofoptimal solution stated by Lemmas 3.1 and 3.2. Let πππππ
SON et al.: UTILITY-OPTIMAL MULTI-PATTERN REUSE IN MULTI-CELL NETWORKS 145
pattern 1
pattern 3
pattern 2interference
no interference
BS 1 BS 2(reference BS) equivalent
P1 = {1, 2} P2 = {1, 3}Pattern 1: (BS1, BS2) = (ON,ON)Pattern 2: (BS1, BS2) = (ON,OFF)Pattern 3: (BS1, BS2) = (OFF,ON)
Fig. 1. Example of a linear two-cell network.
be the effective rate on pattern π for user π, which is thenormalized data rate w.r.t. ππ. Note that there is a trade-offbetween the reuse factor ππ and the data rate πππ. If the userπ chooses the pattern π with the lower value ππ, then the lessBSs are active in the network, and accordingly the higher datarate πππ is expected, and vice versa.
Lemma 3.1: For symmetric networks with static channels,the objective (7) is maximized if and only if
πππ
{β₯ 0, if π = πβ(π),= 0, otherwise,
where πβ(π) = argmaxπ
πππππ.
This implies that each user, if served, only utilizes the patternhaving the largest effective rate. For simplicity, we ignore thecase when more than two patterns have the same largest valuethroughout the analysis in Section III. Accordingly, each userπ can always have the only one optimal pattern π = πβ(π)with πππ > 0.
Lemma 3.2: For the generalized (π€,πΌ)-fair utility function,the optimal fractions of time for user-patterns is given by:
πππβ(π) =(π€πππβ(π)π
1βπΌππβ(π)/π0
)1/πΌand
π0 =( β
πβπ«1
βπβπ¦1π
π€1/πΌπ π
1βπΌπΌ
πβ(π)π1βπΌπΌ
ππβ(π)
)πΌ
,(11)
where π¦1π is the set of users whose most effective patternhaving the highest effective rate is π, i.e., π = πβ(π) =argmaxπ πππππ if π β π¦1π.
Please refer to Appendix for proofs of these lemmas. Now,we give a numerical example to illustrate the property of theoptimal solution.
B. Example: A Linear Two-Cell Symmetric Network
Consider the example of the linear two-cell symmetricnetwork in the Fig. 1. In this example, we have three patternsπ β π« = {1, 2, 3} where π©1 = {1, 2},π©2 = {1}, π©3 = {2}and (π1, π2, π3) = (1, 0.5, 0.5). Suppose that all users havethe same utility function with (π€,πΌ) = (1, 1). Then, wecan obtain π0 = β£π¦1β£ from (11) regardless of the values ofπππβ(π). Let us denote by π¦11 and π¦12 the set of users suchthat ππ1 β₯ 2ππ2 (i.e., the set of center users) and the setof users such that ππ1 < 2ππ2 (i.e., the set of edge users),respectively. Accordingly, the optimal pattern for each userπ β π¦1, i.e., πβ(π) = argmaxπ πππππ, is equal to 1 ifπ β π¦11, and 2 otherwise. Thus, we can obtain the optimal
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Ξ±=0
Ξ±=1
Ξ±=10
The portion of pattern 1: Ο1
The
por
tion
of p
atte
rn 2
and
3: Ο
2 = Ο
3
Fig. 2. Numerical example of the linear two-cell symmetric network whereeach BS has two users; user 1 is in the inner region of the cell and user 2is in the edge of the cell, whose instantaneous data rate vectors are given by(π11, π12) = (10, 11) and (π21, π22) = (3, 8).
time fractions of user-patterns and the optimal portion for eachpattern (π1, π2, π3) as follows:
πππβ(π) =
{β£π¦1β£β1, if π β π¦11,
(2β£π¦1β£)β1, if π β π¦12,
(12)
π1 =β
πβπ¦11
πππβ(π) = β£π¦11β£/β£π¦1β£ and
π2 = π3 =β
πβπ¦12
πππβ(π) = β£π¦12β£/(2β£π¦1β£) .(13)
Note that in the case of proportional fair (πΌ = 1) the optimalportion of each pattern depends only on and is proportionalto the number of users in the sets of center and edge users.However, for general cases (πΌ β= 1), its closed form is verycomplex because the optimal portion of each pattern dependson the data rate πππβ(π) for all users due to (11). Thus, werely on numerical computations for πΌ β= 1.
Fig. 2 depicts the optimal portion of patterns with respect tothe fairness criterion πΌ. We fix the number of users as shownin the Fig. 1, that is, each BS has two users: one is in thecenter and the other in the edge of the cell, β£π¦1β£ = 2, β£π¦11β£ =1, β£π¦12β£ = 1. When πΌ = 1, the optimal portion of patternscan be given by (13): (π1, π2, π3) = (1/2, 1/4, 1/4). Accord-ingly, user throughputs can be easily calculated: (οΏ½οΏ½1, οΏ½οΏ½2) =(π11π11, π22π22) = (π1π11, π2π22) = (5, 2). When we increaseπΌ, i.e., enforcing more fairness, the portions of pattern 2 and3 avoiding ICI increase in order to increase the throughputof edge users. On the other hand, when we decrease πΌ, theportion of pattern 1 increases as expected. In the extreme case,throughput maximization (πΌ goes 0), only user 1 having abetter channel is always served with pattern 1, and user 2cannot be served at all, i.e., (π1, π2, π3) = (1, 0, 0).
C. Joint Optimal Pattern Selection and User Scheduling Al-gorithm
We now present an joint optimal pattern selection anduser scheduling algorithm. To that end, we use a stochasticgradient-based algorithm, e.g., [13] (only considering userscheduling in a single-cell system), that selects the achievablerate vector maximizing the sum of weighted rates where the
146 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 1, JANUARY 2011
weights are marginal utilities at each slot. Then, it sufficesto solve the following problem at each slot, which jointlydetermines the pattern selection X(π‘) = (ππ(π‘) : π β π«)and user scheduling I(π‘) = (πΌπ(π‘) : π β π¦):Q-joint: (14)
maxX(π‘),I(π‘)
Ξπ(π‘) =βπβπ¦
π β²π(οΏ½οΏ½π(π‘β 1))ππ(π‘) (15)
subject toβπβπ«
ππ(π‘) = 1, (16)
βπβπ¦π
πΌπ(π‘)
{β€ 1, if ππ(π‘) = 1 and π β π©π,= 0, otherwise,
(17)
where ππ(π‘) =β
πβπ« ππ(π‘)πΌπ(π‘)πππ(π‘) is the actual data rateassigned to user π at slot π‘ and οΏ½οΏ½π(π‘) = 1
π‘
βπ‘π=1 ππ(π) =
οΏ½οΏ½π(π‘β1)+ ππ‘[ππ(π‘)β οΏ½οΏ½π(π‘β 1)
](by letting ππ‘ = 1/π‘) is the
long-term throughput for user π up to slot π‘.Remark 3.3: If we fix the user scheduling I(π‘) and choose
utility function as ππ(οΏ½οΏ½π) = οΏ½οΏ½π in Q-joint, then this problemis reduced to binary power control (BPC) problem for sum-rate maximization in [2].
The problem Q-joint can be naively solved by an exhaustivesearch. For each pattern π, it needs to compare all possiblecombinations of user scheduling for all BSs. Thus, this naiveapproach requires π(π β πΎπ) complexity, which is compu-tationally intractable. However, Lemma 3.4 tells us the niceproperty of the problem that we need to consider only thecase with the best users selected by intra-cell user schedulingin (18) instead of all possible combination of user schedulingfor each pattern.
Lemma 3.4: The problem Q-joint can be decomposed intothe following β£π©πβ£ independent intra-cell user schedulingproblems for a given pattern π:
πβπ(π‘) = arg max
πβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πππ(π‘), βπ β π©π. (18)
Please refer to Appendix for a proof.With the help of Lemma 3.4, we can develop the joint
optimal pattern selection and user scheduling algorithm. Foreach pattern π, we select the best user having the largest valueof π β²
π(οΏ½οΏ½π(π‘ β 1))πππ(π‘) from (18) and then the best patternπβ(π‘) that maximizes the sum of weighted rate π β²
π(οΏ½οΏ½π(π‘ β1))πππβ(π‘) of the scheduled users. Note that it has muchlower complexity2 π(π β βπβπ© πΎπ) = π(π β πΎ) than thatof exhaustive search π(π β πΎπ). The proof of convergenceto the optimal solution is a slight extension to [13], [14] thatstudied only user scheduling for a fixed pattern. We skip theproof.
Joint pattern selection and user scheduling algorithm
At each slot π‘, compute (πβ(π‘), πβπ(π‘), π β π© ) satisfying
πβ(π‘) = argmaxπβπ«
βπβπ©π
[maxπβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πππ(π‘)
],
πβπ(π‘) = arg max
πβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πππβ(π‘), βπ β π©πβ .
(19)
2The maximum operation in the intra-cell user scheduling requires linearcomplexity in the number of users.
This joint optimal pattern selection and user schedulingalgorithm requires instantaneous channel feedbacks from allusers in the network. We assume that at each slot π‘, userπ estimates its own SINR for all patterns π β π«π(π) uponlistening to pilot signals, calculates the instantaneous datarate πππ(π‘) and then reports this information to the centalcoordinator through its serving BS.
However, this joint optimal algorithm still has implementa-tion difficulties. Apart from the computational complexity ofthis algorithm, the central coordinator running the algorithmneeds to collect the following information from each BSπ β π© : instantaneous data rate πππ(π‘) of all its associatedusers π β π¦π on its available patterns π β π«π. The totalamount of feedbacks is quite large, i.e, (
βπβπ© β£π¦πβ£β£π«πβ£),
though they may be delivered along with high speed wiredlinks. Furthermore, a series of tasks, including informationfeedback from BSs to the central coordinator as well asthe computation and the distribution of central coordinatorβsdecision, should be performed in one slot.
IV. TIME-SCALE DECOMPOSED ALGORITHM
A. Algorithm Description
In contrast to the centralized joint pattern selection anduser scheduling algorithm in Section III, user scheduling inpractice is typically undertaken by individual BSs indepen-dently without any coordination or information exchanges withother BSs. In this section, in order to take into account suchautonomous features in user scheduling and to overcome highcomputation and feedback overhead in the optimal algorithm,user scheduling is run at every slot, whereas pattern portionchange less frequently, in this case, every ππ >> 1 slots. Wefirst describe the proposed algorithm (see Fig. 3 for a pictorialdescription) and then explain the rationale behind it.
Pattern portion change algorithm
Initialization: ππ = 1/β£π«β£ for all π β π« .For every ππ slots, each BS π β π© computes the partialderivative π·
(π)π
.= βπ (π)/βππ and sends it to the central
coordinator,
π·(π)π =
βπβπ¦π
π β²π(οΏ½οΏ½π)β
( οΏ½οΏ½ππ
πππππ
), π β π«π. (20)
Then, the central coordinator calculates the gradient vectorD = (π·1, π·2, β β β , π·π ) by collecting π·
(π)π from all BSs,
π·π =βπβπ©
π·(π)π , π β π« , (21)
and updates the pattern portion vector π as follows,
π β ππππβπβπ« ππ=1, (π + πΎD) , (22)
where πππππ΄(β ) denotes an orthogonal projection on a set π΄.
User scheduling algorithm
Initialization: οΏ½οΏ½π(0) = οΏ½οΏ½ππ(0) = πππ(0) = 0.
SON et al.: UTILITY-OPTIMAL MULTI-PATTERN REUSE IN MULTI-CELL NETWORKS 147
BS 1
BS N
BS 2
Predetermine the sequence of patterns
Updates the portion of reuse patters
Userscheduling
Userscheduling
Userscheduling
Userscheduling
Userscheduling
Userscheduling
Userscheduling
Userscheduling
Userscheduling
Central coordinator
(1)pD
(2)pD
( )NpD
( )np p
n N
D D1 2, , , P
1 2 Tp...
Fig. 3. Proposed time-scale decomposed algorithms.
At each slot π‘, each BS π β π©π(π‘) activated by pattern π(π‘)selects the user πβ
π(π‘), i.e., πΌπβππ(π‘) = 1,
πβπ(π‘) = arg max
πβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πππ(π‘),
3 (23)
and updates the following variables for all users π β π¦π withsome constants 0 < π½1, π½2, π½3 < 1:
οΏ½οΏ½π(π‘) = (1β π½1)οΏ½οΏ½π(π‘β 1) + π½1πΌπ(π‘)πππ(π‘) ,οΏ½οΏ½ππ(π‘) = (1β π½2)οΏ½οΏ½ππ(π‘β 1) + π½2πΌπ(π‘) ,
πππ(π‘) =
{(1β π½3)πππ(π‘β 1) + π½3πππ(π‘), if πΌπ(π‘) = 1 ,πππ(π‘β 1), otherwise.
(24)
Two algorithms with different time scales interact with eachother as follows: The pattern portion change algorithm adjuststhe portion of reuse patterns π for every ππ slots, using thevariables οΏ½οΏ½π(π‘), οΏ½οΏ½ππ(π‘) and πππ(π‘). These variables essentiallycorrespond to the long-term averages of πΌπ(π‘)πππ(π‘), πππ(π‘),and πππ(π‘) which are progressively updated at every slot bythe user scheduling algorithm. This time-scale decompositionand the way of interaction between the two algorithms impliesthat the pattern portion algorithm is designed and operated sothat it tracks only the average interference levels and channelconditions rather than instantaneous conditions such as thejoint optimal algorithm in Section III. Considering that a userscheduling algorithm can be carried out autonomously, theactual (amortized) complexity and message passing overheadper slot can be significantly reduced, which makes the pro-posed algorithms much more implementable. Subsection IV-Cincludes a discussion of the value of such complexity reduc-tion, i.e., a utility performance gap compared to the optimalalgorithm.
B. Rationale of Time-scale Decomposed Algorithms
The pattern portion change algorithm can be regarded asa standard gradient projection algorithm for the following
3For completeness, if a tie happens, the BS choose the lower indexeduser.
problem:
Q-pattern:
maxπ
βπβπ¦
ππ(οΏ½οΏ½π) =βπβπ¦
ππ
(βπβπ«
ππππππππ
)
subject toβπβπ«
ππ = 1,
where πππ β [0, 1] is the probability that the user π isscheduled when pattern π is selected, i.e., πππ β ππ = οΏ½οΏ½ππ.For each pattern portion update epoch, i.e., every ππ slots,each BS π calculates the partial derivative π·
(π)π
.=βπ (π)/βππ
of per-cell utility π (π) with respect to the portion of patternπ and sends these information to the central coordinator.
π·(π)π
.=
βπ (π)
βππ=
βπβπ¦π
π β²π(οΏ½οΏ½π) β βοΏ½οΏ½π
βππ, (25)
where
βοΏ½οΏ½π
βππ= ππππππ =
οΏ½οΏ½ππ
πππππ. (26)
Note that three parameters (οΏ½οΏ½π, οΏ½οΏ½ππ and πππ) required torun this pattern portion update algorithm can be attained bythe user scheduling algorithm. Then the central coordinatorgathers information from all BSs and calculates the partialderivative of the network utility π·π
.= βπ/βππ by aggregating
these partial derivatives of the local utility,
π·π.=
βπ
βππ=
βπβπ©
π·(π)π , π β π« , (27)
and updates the portion of reuse patterns following the ascentdirection of network utility.
π β ππππβπβπ« ππ=1, (π + πΎD) . (28)
Based on the updated portion of patterns, the central co-ordinator predetermines the sequence of patterns for next ππ
slots that satisfies:
(the total number of pattern π) / ππ β ππ , βπ β π« . (29)
While there may be many strategies that leads to (29), anice candidate is a random strategy. The central coordinatorsequentially determines the sequence of patterns by rolling aπ -dimensional die ππ times with probability of the pattern πbeing ππ. Once the sequence of patterns for next ππ slots isdetermined by the pattern portion change algorithm, then bothBSs and users are informed of the sequence.
Now we develop the user scheduling algorithm under thefixed pattern given by the pattern portion change algorithm.From Lemma 3.4, for given pattern, the network-wide userscheduling problem can be decomposed into independentintra-cell user scheduling problems. Therefore, each BS needsto solve the following problem:
Q-scheduling:
max(πΌπ(π‘),πβπ¦π)
βπβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πΌπ(π‘)πππ(π‘)
subject toβπβπ¦π
πΌπ(π‘) β€ 1.
148 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 1, JANUARY 2011
TABLE ICOMPARISON BETWEEN JOINT OPTIMAL ALGORITHM (JOA) AND TIME-SCALE DECOMPOSED ALGORITHM (TDA)
Joint optimal algorithm (JOA) Time-scale decomposed algorithm (TDA)
Time-scale of algorithm every slotevery slot (user scheduling)
every ππ slot (pattern portion change)
Amount of feedback to each BS π at each slot β£π¦πβ£β£π«πβ£ β£π¦πβ£Amount of feedback to the central coordinator
βπβπ© β£π¦πβ£β£π«πβ£
βπβπ© β£π«πβ£
Period of feedback to the central coordinator 1 ππ
The user scheduling algorithm solving Q-scheduling isstraightforward. Each BS π β π©π allowed to use the pat-tern π independently chooses the best user πβ
π(π‘) among itsassociated user set π¦π, i.e., πΌπβ
π(π‘) = 1:
πβπ(π‘) = arg max
πβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πππ(π‘), βπ β π©π, (30)
and updates the following variables for the future purpose ofthe pattern portion change algorithm:
οΏ½οΏ½π(π‘) = (1β π½1)οΏ½οΏ½π(π‘β 1) + π½1πΌπ(π‘)πππ(π‘) ,οΏ½οΏ½ππ(π‘) = (1β π½2)οΏ½οΏ½ππ(π‘β 1) + π½2πΌπ(π‘) ,
πππ(π‘) =
{(1 β π½3)πππ(π‘β 1) + π½3πππ(π‘), if πΌπ(π‘) = 1 ,πππ(π‘β 1), otherwise,
where π½1, π½2, π½3 > 0 are small, averaging parameters; οΏ½οΏ½π(π‘),οΏ½οΏ½ππ(π‘) and πππ(π‘) are the average throughput of user π, theaverage fraction of time that user π is served with pattern π,and the average instantaneous data rate when the user π isserved with pattern π, respectively.
Remark 4.1: For the user scheduling algorithm, it is re-quired for each user to predict its potential data rate for thepattern in the next slot and report to its serving BS. We assumethat each user can estimate the channel gains of different BSsfrom common pilot channels and send channel quality infor-mation to its serving BS through dedicate feedback channels.These functionalities are considered as a basic requirement inlegacy and upcoming standard. Thus, each user can calculateits SINR for the pattern in the next slot by regarding theinterference only from the activated BSs in the next slot aseffective interference. Then, each user reports the predictedSINR (or corresponding potential data rate) to its serving BS.If the serving BS will not be activated in the next slot, theuser does not need to send feedback at all.
Remark 4.2: There are two key differences between a re-cent algorithm [7] and the proposed algorithm. First, thereferenced study additionally introduces a virtual scheduler toobtain the fraction of time in which the scheduler chooses userπ for transmission in sub-band π (their notation: πππ ). In theproposed algorithm, however, the fraction of time that user πis served with pattern π (our notation: οΏ½οΏ½ππ) is simply obtainedusing the actual scheduler without an extra algorithm. Second,the referenced study does not reflect the time-varying nature ofthe data rate available to user π in sub-band π (their notation:π ππ ), as they assume that this rate does not change with time.In the proposed algorithm, the long-term average of the datarate of user π on pattern π (our notation: πππ) is not just theaverage of instantaneous data rate. We take the average ofinstantaneous data rate only if the user π is actually served
by the scheduler. In other words, the value of πππ in theproposed algorithm can reflect the multi-user diversity gainfrom exploiting the channel fluctuation.
C. Complexity Reduction and Its Price
The proposed time-scale decomposed algorithm still in-volves signalings from BSs to a central coordinator. However,it is possible to reduce feedback overhead significantly, asthe periodicity of the feedback is stretched from every slot toevery ππ slot. Moreover, the amount of feedback is reducedfrom
(βπβπ© β£π¦πβ£β£π«πβ£
)to
(βπβπ© β£π«πβ£
); i.e., it requires
only BS-level feedback and not user-level channel feedback.The amount of feedback to each BS from its associated usersat each slot is also reduced from β£π¦πβ£β£π«πβ£ and β£π¦πβ£, as usersneed to send channel information only for a predeterminedpattern. Table I compares the joint pattern selection and userscheduling algorithm with the proposed algorithms based ontime-scale decomposition.
This complexity reduction for implementability comes atthe cost of a performance gap with the joint optimal algorithm.This is due to the fact that the ICI management part inthe decomposed algorithm cannot fully exploit instantaneousinter-cell channel variations; hence, only intra-cell channelvariations are opportunistically utilized. Note that in the jointoptimal algorithm, both pattern selection and user schedulingfully exploit both inter-cell and intra-cell time-varying channelconditions at a fast time scale.
As an example, consider a two cell network where two usersare located at the edge of each cell. Their achievable ratesare limited by severe ICI. The decomposed algorithm willfind the following TDMA-like solution: BS 1 and BS 2 areexclusively active in order to mitigate the ICI, i.e., the portionof the pattern in which both BSs are active is nearly zero.However, suppose that both (time-varying) inter-cell channelgains from BS 1 (or 2) to the user in BS 2 (or 1) experiencedeep fading at some time slot. You can imagine this case as ifthere were a large wall between two cells. Subsequently theuser in cell 1 (or 2) is not interfered by the BS transmission incell 2 (or 1). Therefore, serving two users simultaneously istransiently optimal in this inter-cell deep fading case, whereasa pattern that only one BS is active is the solution for theaverage ICI mitigation case. The joint optimal algorithm canfind this optimal solution by tracking this fast fading conditionwhile the decomposed algorithm cannot. This is the cost forthe complexity reduction, however, as shown in Section V, theperformance gap becomes negligible in then absence of fastfading.
SON et al.: UTILITY-OPTIMAL MULTI-PATTERN REUSE IN MULTI-CELL NETWORKS 149
D. Construction method of the candidate pattern set π«The number of all possible patterns in the network is an ex-
ponential function of the number of BSs, which may increasethe complexity of the algorithm and reduce its practicalityin the large network. However, most of these patterns areactually not used at all, thus it is required to select the essentialcandidate set of patterns out of all possible patterns.
Several papers [8], [15] have investigated on this topic.Bonald et al. [8] investigated an optimal transmit profile in thecellular network based on flow-level analysis. They concludedbased on all their examples that the optimal capacity is attainedby the use of two kinds of transmission patterns only: 1) onepattern is that all BSs are on and 2) the other patterns arethat only the dominant interfering BS is switched off. Ramanet al. [15] proposed a centralized spectrum server that findsan optimal schedule to maximize the average sum rate ingeneral ad-hoc networks. They also conjectured that almostalways only very few active transmission patterns are used ascorroborated by their simulation results.
Encouraged by the observations in [8], [15], we make apractical guideline how to determine the candidate pattern setπ« as follows. The set should contain two kinds of mandatorypatterns: 1) reuse-1 pattern: all BSs are active and 2) dominantpatterns π: all neighboring BSs except the dominant interferingBS are activated by the pattern π. Any other appropriatepatterns may be added in an optional manner if systemdesigners want to increase the performance further, but theincrement might be marginal.
V. PERFORMANCE EVALUATION
A. Simulation Setup
Two network configurations are considered for simulation-based performance evaluations: (i) a linear two-cell networkand (ii) a two-tier multi-cell network with 19 cells. In bothcases, the distance between BSs is set at 2km.
β Linear two-cell network (see Fig. 1): There are threepatterns π« = {1, 2, 3}. Under pattern 1, both BSs areON, and under pattern 2 (resp. 3), only BS 1 (resp. 2) isON.
β Two-tier multi-cell network (see Fig. 4): 11 patterns (8mandatory + 3 optional reuse-3 patterns) are considered.Under pattern 1, all BSs are ON. Under patterns 2βΌ4(only one BS reuses the pattern among adjacent threeBSs) or 5βΌ11 (six BSs reuse the pattern among adjacentseven BSs), a BS using these two types of patterns canexpect ICI mitigation from the first-tier and from one ofits neighboring cells, respectively.
In modeling the propagation environment, a path lossβ130β35 log10(πππ), log-normal shadowing with a standarddeviation ππ =8dB and Jakesβ Rayleigh fading (3km/h) forfast fading are adopted. In subsection V-B, we evaluate theperformance of the proposed algorithm with and withoutfast fading cases. In the case with fast fading, the channelvaries over the time since all the above mentioned modelsare considered. On the other hand, in the case without fastfading, we consider only the path loss and shadowing modelsexcept Jakesβ fading. Therefore, the channels remain stableduring the simulation time. The channel bandwidth and the
BS1
1,2,6,7,8,9,10,11
BS2
1,3,5,7,8,9,10,11
BS3
1,4,5,6,8,9,10,11
BS4
1,3,5,6,7,9,10,11
BS5
1,4,5,6,7,8,10,11
BS6
1,3,5,6,7,8,9,11
BS7
1,4,5,6,7,8,9,10
BS8
1,4,5,6,7,8,9,11
BS9
1,2,5,6,7,8,10,11
BS10
1,3,5,6,7,8,9,10
BS11
1,2,5,6,7,8,9,11
BS12
1,4,5,7,8,9,10,11
BS13
1,2,5,6,7,8,9,10
BS14
1,3,5,6,8,9,10,11
BS15
1,2,5,7,8,9,10,11
BS16
1,4,5,6,7,9,10,11
BS17
1,2,5,6,8,9,10,11
BS18
1,3,5,6,7,8,10,11
BS19
1,2,5,6,7,9,10,11
Fig. 4. Two-tier multi-cell network composed of 19 cells.
time-slot length are set at 10MHz and 5ms, respectively. Themaximum transmission powers of the BSs are all identicalto 20W. The other parameters for the simulations follow thesuggestions in the IEEE 802.16m evaluation methodologydocument [16]. All users have a logarithmic utility function,i.e., (π€,πΌ) = (1,1). The pattern update period ππ is set at 500and the step size is chosen to be a typically small value, i.e.,π½1 = π½2 = π½3 = πΎ = 0.001. We have tested other valuesof ππ, π½1, π½2, and π½3, and the similar results were obtained.Simulations were run of over 50000 slots.
To evaluate the performance under various user distributionscenarios, we introduce a variable, so-called, βuser distributionoffsetβ π β [0, 1]. It adjusts the minimum distance between theBS and the user to π Γ (cell radius). Users in each cell arerandomly distributed with this minimum distance restriction.For example, if π = 0, users are uniformly generated over theentire cell. On the other hand, if π becomes 1, users swarmthe edge areas of cells.
Simulation results of the following three algorithms arepresented:(i) a conventional universal reuse scheme (UNI), inwhich all BSs in the network are always active without anyICI management, (ii) the joint optimal algorithm (JOA) in(19), and (iii) the algorithm based on time-scale decomposition(TDA) in (20)βΌ(24). As performance metrics, the geometricaverage of user throughputs (GAT) and the average of edgeuser throughputs (AET) are used. We use GAT becausemaximizing this metric is equivalent to the system objective(sum of log throughputs). AET is a measure of cell edgeperformance that is defined as the average throughput of theusers located at the cell edges. In simulation, we treat βedgeusersβ as those who are more than 800m away from theirserving BSs, otherwise termed βcenter usersβ.
B. Two-Tier 19-Cell Network Case
We first consider the two-tier multi-cell network with 19cells, where each cell has ten users. Without fast fading, asshown in Fig. 5(a), both JOA and TDA perform similarly.Compared to UNI, GAT and AET of JOA and TDA increasesby 10βΌ33% (depending on the user distribution) and by
150 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 1, JANUARY 2011
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90
1
2
3
4
5
6
7
8
User distribution offset, Ο
Thr
ough
put p
erfo
rman
ce [M
bps]
GAT: JOAGAT: TDAGAT: UNIAET: JOAAET: TDAAET: UNIGeometric average of user
throughputs(GAT)
Average of edge user throughputs(AET)
~33%
(a) Without fast fading
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90
1
2
3
4
5
6
7
8
User distribution offset, Ο
Thr
ough
put p
erfo
rman
ce [M
bps]
GAT: JOAGAT: TDAGAT: UNIAET: JOAAET: TDAAET: UNI
~25%
Average of edge user throughputs(AET)
Geometric average of user throughputs(GAT)
(b) With fast fading
Fig. 5. Throughput performances of three algorithms: joint optimal algorithm(JOA), time-scale decomposed algorithm (TDA) and universal reuse (UNI).
33%. A higher performance gain was observed when the userdistribution offset is larger (i.e., more users are located at thecell edges), which is due to the fact that ICI managementmainly targets for performance improvement of edge users.With fast fading, as shown in Fig. 5(b) however, as discussedin subsection IV-C, a performance gap between JOA and TDAexists due to the loss in opportunism in TDA. However, TDAstill outperforms UNI in terms of both GAT (5βΌ25% depend-ing on user distribution) and AET (25%). It is noteworthy thatTDA can attain more than 1/2 (at π = 0) and up to 2/3 (atπ = 0.9) of the GAT performance gain that can be achievedby JOA.
Fig. 6 shows the convergence of 11 pattern portions whenoffset π is equal to 0.3. As you can see, the pattern portionsconverge quickly in 10βΌ15 iterations. Even in different net-work configurations (with a different number of BSs and adifferent values of offset), we could see similar convergencetrends, e.g., typically within 10βΌ20 iterations.
In our simulation, the proportional fairness (πΌ = 1) isconsidered as our system objective. Although additional sim-
0 10 20 30 40 50 600
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Number of iterations
Por
tions
of p
atte
rns
Ο2 ~ Ο
4
Ο5 ~ Ο
11
Ο1
Fig. 6. Rate of convergence.
ulation results with other values of πΌ are not included in thispaper due to space limitations, higher performance gains wereobtained for higher values of πΌ (i.e., higher fairness objec-tives), as increasing the throughput of edge users is essentialin order to achieve higher fairness. Thus, ICI managementcan offer conspicuous improvement in a network pursuingfairness-oriented system objectives.
C. Adaptation Test
To test adaptation to dynamic changes in user loads, thelinear two-cell network was used. The user load distributionwas varied every 2Γ104 slots (=100 sec) as shown in Fig. 7(a).The TDA was compared with other three static schemes: RF1,RF2 and MIXED. RF1 and RF2 correspond to those with reusefactors 1 and 2, respectively. RF1 and RF2 are incorporated inthe MIXED scheme, which operates as a RF1 during the halfportion of the time and RF2 during the other half portion ofthe time. The RF1, RF2 and MIXED schemes can be regardedas those with static pattern portions (π1, π2, π3) = (1, 0, 0),(0, 0.5, 0.5) and (0.5, 0.25, 0.25), respectively.
Fig. 7(b) shows the pattern portion adaptation characteristicsof TDA to a dynamic user load distribution. In the period 1(0 βΌ 2Γ 104 slots), as there are relatively many users aroundthe cell center, the portion of pattern 1 increases up to 80%,which is nearly identical to that of RF1. On the other hand, inthe another (8 βΌ 10Γ 104 slots), as there are relatively manyusers around the cell edge, the portion of pattern 1 decreasesup to 20%, which is nearly identical to that of RF2. Users areequally divided into center and edge regions in the period 2(2 βΌ 4 Γ 104 slots) so that the pattern portion becomes (0.5,0.25, 0.25) like as MIXED. In periods 3 and 4, there is thesame number of center users in each cell, but the number ofedge users are different; i.e., heterogeneous load distribution.In period 3 (resp. 4), BS A (resp. B) exceeds the numberof edge users. As expected, π2 (resp. π3) increases while π3
(resp. π2) decreases.Fig. 7(c) shows the GAT performance comparison period
by period. Static schemes perform well only if the user loaddistribution fits their patterns, for example, RF1 in period 1,RF2 in period 5 and MIXED in period 2. In the cases in-
SON et al.: UTILITY-OPTIMAL MULTI-PATTERN REUSE IN MULTI-CELL NETWORKS 151
1 2 3 4 5 6 7 8 9 10
x 104
0
5
10
15
20
Slots
Num
ber
of u
sers
center users of BS Aedge users of BS Acenter users of BS Bedge users of BS B
(a) Dynamic change of user load distribution
1 2 3 4 5 6 7 8 9 10
x 104
0
0.2
0.4
0.6
0.8
1
Slots
Por
tion
of p
atte
rns
Ο1
Ο2
Ο3
(b) Pattern portion adaption to the dynamic user load distribution
period 1 period 2 period 3 period 4 period 50
2
4
6
8
10
12
GA
T [M
bps]
RF1: UNIRF2MIXED (RF1+RF2)Proposed: TDA
(c) GAT performance comparison period by period
Fig. 7. Adaptiveness to dynamic user load.
volving other distributions, however, performance degradationis inevitable. Compared to these static schemes, the proposedTDA scheme can adapt to the dynamic user load distributionand find the optimal portion of patterns. Therefore, it alwaysachieves the best performance in all cases.
D. Imbalance Load Scenario: Association Control vs. Inter-ference Control
We also test the performance in the linear two-cell networkfor imbalanced loads. Two users were located 900m away fromBS 1 and (2ΓπΏπΌ) users were located 900m away from BS 2,respectively, where πΏπΌ quantifies the load imbalance. Under
1 2 3 4 5 6 7 8 9 100.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
Load imbalance
Nor
mal
ized
per
form
ance
Interference controlAssociation control
Fig. 8. Association control vs. interference control under imbalanced load.
this load imbalance scenario, the performance of the follow-ing two different approaches are compared: (i) Associationcontrol: a load-aware handover in [9] and (ii) Interferencecontrol: the proposed TDA. As a baseline, we considered theGAT performance of UNI and normalized the performance ofthe two approaches above by UNI.
Originally, users are associated with the closest BS offeringthe best signal strength. In the case of the association controlapproach, however, if the expected throughput measure in [9]from the other BS is greater than that from the current BS, thenthe user changes its association. When the πΏπΌ value is small,users do not change their associations. When πΏπΌ is increasedto more than 6, association changes from the hot-spot cell (BS2) and the under-loaded cell (BS1) occur (moving one, twoand three users at πΏπΌ=6, 8 and 10, respectively) according tothe load-aware handover in [9]. As shown in Fig. 8, however,the gain from the association control is marginal.
On the other hand, using the interference control approach,we can implicitly resolve the load imbalance by preventing thehot-spot cell (BS 2) from being turned off, i.e., provide moreinterference-free transmission opportunities compared to BS 1.In brief, the interference control approach originally developedfor ICI mitigation can also resolve the load imbalance as well,and the improvement of interference control is superior to thatof the association control.
VI. CONCLUSION
In this paper, we have focused on the problem of joint ICImanagement and user scheduling in multi-cell networks. It wasshown that the joint optimal algorithm is too complex in termsof computational and signaling overhead to be implemented inpractical systems. To overcome this complexity and make thealgorithm practical, we decomposed the original optimizationproblem into two sub-problems, where ICI management isrun at a slower time scale compared to user scheduling. Thistime-scale decomposition stems from a design rationale inwhich ICI management may not have to track fast changingdynamics, and it may suffice to attain much gain simply byrunning it based only on macroscopic network changes. We
152 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 1, JANUARY 2011
empirically show that even with such a slow tracking of thesystem dynamics at the ICI management, our algorithm canachieve high performance gain compared to the conventionaluniversal reuse scheme. Moreover, it is practically imple-mentable compared to the very complex optimal algorithm.
APPENDIX
A. Proof of Lemma 3.1
Let π0 and πππ denote the Lagrangian multiplier associatedwith the constraint (8) and constraint (9), respectively. Thenthe Lagrangian function is given by:
β(π, π) =βπβπ¦1
ππ(οΏ½οΏ½π)βπ0
( βπβπ«1
βπβπ¦1
πππ
ππβ 1
)
+βπβπ¦1
βπβπ«1
ππππππ.
The problem Q-symmetric is a convex optimization so thenecessary and sufficient conditions for optimality are givenby Karush-Kuhn-Tucker (KKT) conditions [17]:
i) π0 β₯ 0 and πππ β₯ 0, βπ β π¦1, βπ, (31)
ii) π0
(βπβπ«1
βπβπ¦1
πππ
ππβ 1
)= 0 and ππππππ = 0, βπ, βπ, (32)
iii)βββπππ
= π β²π(οΏ½οΏ½π)πππ β π0
ππ+ πππ = 0, βπ β π¦1, βπ. (33)
Substituting (33) into (32) and (31) yields the followingconditions, (34) and (35), respectively.[
π0
ππβ π β²
π(οΏ½οΏ½π)πππ
]πππ = 0, βπ β π¦1, βπ β π«1, (34)
π0 β₯ π β²π(οΏ½οΏ½π)πππππ, βπ β π¦1, βπ β π«1. (35)
Let πβ(π) = argmaxπ
πππππ denote the most efficient pattern
of user π. By (35), there are two cases.Case 1: If π0 > π β²
π(οΏ½οΏ½π)ππβ(π)πππβ(π), then[π0
ππβ π β²
π(οΏ½οΏ½π)πππ
]> π β²
π(οΏ½οΏ½π)πππ
(ππβ(π)πππβ(π)
πππππβ 1
)> 0.
Consequently, (34) holds only if πππ = 0, βπ, βπ, however,this is not a reasonable solution.Case 2: If π0 = π β²
π(οΏ½οΏ½π)ππβ(π)πππβ(π), then[π0
ππβ π β²
π(οΏ½οΏ½π)πππ
]= π β²
π(οΏ½οΏ½π)πππ
(ππβ(π)πππβ(π)
πππππβ 1
).
Consequently, (34) holds only if πππ β₯ 0 if π = πβ(π), and 0otherwise. This completes the proof of Lemma 3.1. β
B. Proof of Lemma 3.2
By the Lemma 3.1, we can rewrite the condition (33) forπ = πβ(π) as follows:
π β²π(οΏ½οΏ½π)πππβ(π) β π0
ππβ(π)= 0. (36)
The derivative for the generalized (π€,πΌ)-fair utility is givenby π β²
π(οΏ½οΏ½π) = π€π/οΏ½οΏ½πΌπ = π€π/(πππβ(π)πππβ(π))
πΌ. Substitutingthis derivative into (36) yields the following optimal fractionof time for user-patterns:
πππβ(π) =(π€πππβ(π)π
1βπΌππβ(π)/π0
)1/πΌ. (37)
Since πππ = 0 for π β= πβ(π), we can rewrite the condition(32) as follows:β
πβπ«1
βπβπ¦1
πππ
ππβ 1 =
βπβπ«1
βπβπ¦1π
πππ
ππβ 1 = 0, (38)
where π¦1π is the set of users whose most effective patternhaving the highest effective rate is π, i.e., π = πβ(π) =argmaxπ πππππ if π β π¦1π. By plugging (37) into (38), wecan have the optimal value of π0:
π0 =( βπβπ«1
βπβπ¦1π
π€1/πΌπ π
1βπΌπΌ
πβ(π)π1βπΌπΌ
ππβ(π)
)πΌ. (39)
This completes the proof of Lemma 3.2. β
C. Proof of Lemma 3.4
For the given pattern π, i.e., ππ(π‘) = 1, we can rewrite (15)as follows,
Ξπ(π‘) =βπβπ©
βπβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πΌπ(π‘)πππ(π‘)
=βπβπ©π
βπβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πΌπ(π‘)πππ(π‘)
=βπβπ©π
Ξππ(π‘),
where the second equality holds from (2), πΌπ(π‘) = 0, π /βπ©π and Ξππ(π‘) =
βπβπ¦π
π β²π(οΏ½οΏ½π(π‘ β 1))πΌπ(π‘)πππ(π‘). As
π β²π(οΏ½οΏ½π(π‘ β 1)) and πππ(π‘) are given parameters at slot π‘, we
only have to investigate decencies among πΌπ(π‘). Since theconstraints (2) on πΌπ(π‘) do not play a role across different BSs,Ξππ(π‘) are mutually independent. Therefore, solving the orig-inal problem is equivalent to maximize Ξππ(π‘) independentlyfor each BS π β π©π.
maxI(π‘)
Ξππ(π‘) =βπβπ¦π
π β²π(οΏ½οΏ½π(π‘β 1))πΌπ(π‘)πππ(π‘)
subject toβπβπ¦π
πΌπ(π‘) β€ 1,
This problem can be further simplified as (18), which com-pletes the proof of Lemma 3.4. β
REFERENCES
[1] WiMAX Forum, βMobile WiMAXβpart I: a technical overview andperformance evaluation,β Aug. 2006.
[2] A. Gjendemsj, D. Gesbert, G. E. Oien, and S. G. Kiani, βBinary powercontrol for sum rate maximization over multiple interfering links,β IEEETrans. Wireless Commun., vol. 7, no. 8, pp. 3164β3173, Aug. 2008.
[3] S. Das, H. Viswanathan, and G. Rittenhouse, βDynamic load balancingthrough coordinated scheduling in packet data systems,β in Proc. IEEEINFOCOM, San Francisco, CA, Mar. 2003.
[4] J. Cho, J. Mo, and S. Chong, βJoint network-wide opportunistic schedul-ing and power control in multi-cell networks,β in Proc. IEEE WoWMoM,San Francisco, CA, June 2007.
[5] K. Son, S. Chong, and G. de Veciana, βDynamic association for loadbalancing and interference avoidance in multi-cell networks,β IEEETrans. Wireless Commun., vol. 8, no. 7, pp. 3566β3576, July 2009.
[6] B. Rengarajan and G. de Veciana, βNetwork architecture and abstrac-tions for environment and traffic aware system-level coordination ofwireless networks: the downlink case,β in Proc. IEEE INFOCOM,Phoeniz, AZ, Apr. 2008.
[7] A. L. Stolyar and H. Viswanathan, βSelf-organizing dynamic fractionalfrequency reuse for best-effort traffic through distributed inter-cellcoordination,β in Proc. IEEE INFOCOM, Rio de Janeiro, Brazil, Apr.2009.
SON et al.: UTILITY-OPTIMAL MULTI-PATTERN REUSE IN MULTI-CELL NETWORKS 153
[8] T. Bonald, S. Borst, and A. Proutiere, βInter-cell scheduling in wirelessdata networks,β in Proc. European Wireless, Cyprus, Greece, Apr. 2005.
[9] A. Sang, X. Wang, M. Madihian, and R. D. Gitlin, βCoordinated loadbalancing, handoff/cell-site selection, and scheduling in multi-cell packetdata systems,β in Proc. ACM MobiCom, Philadelphia, PA, Sep. 2004,pp. 302β314.
[10] T. Bu, L. Li, and R. Ramjee, βGeneralized proportional fair schedulingin third generation wireless data networks,β in Proc. IEEE INFOCOM,Barcelona, Spain, Apr. 2006.
[11] F. Kelly, A. Maullo, and D. Tan, βRate control in communication net-works: shadow prices, proportional fairness and stability,β J. OperationalResearch Society, vol. 49, pp. 237β252, July 1998.
[12] J. Mo and J. Walrand, βFair end-to-end window-based congestioncontrol,β IEEE/ACM Trans. Networking, vol. 8, no. 5, pp. 556β567,Oct. 2000.
[13] A. L. Stolyar, βOn the asymptotic optimality of the gradient schedulingalgorithm for multiuser throughput allocation,β Operations Research,vol. 53, no. 1, pp. 12β25, Jan. 2005.
[14] H. J. Kushner and P. A. Whiting, βConvergence of proportional-fairsharing algorithms under general conditions,β IEEE Trans. WirelessCommun., vol. 3, no. 4, pp. 1250β1259, July 2004.
[15] C. Raman, R. D. Yates, and N. B. Mandayam, βScheduling variable ratelinks via a spectrum server,β in Proc. IEEE DySPAN, Baltimore, MD,Nov. 2005.
[16] IEEE 802.16m-08/004r5, βIEEE 802.16m Evaluation Methodology Doc-ument (EMD),β Jan. 2009.
[17] S. Boyd and L. Vandenberghe, Convex Optimization, 1st edition. Cam-birdge University Press, 2004.
Kyuho Son (Sβ03-Mβ10) received his B.S., M.S. andPh.D. degrees all in the Department of Electrical En-gineering from Korea Advanced Institute of Scienceand Technology (KAIST), Daejeon, Korea, in 2002,2004 and 2010, respectively. He is currently a post-doctoral research associate in the Department ofElectrical Engineering at the University of SouthernCalifornia. His current research interests includeinterference management in heterogeneous cellularnetworks, green wireless networking and networkeconomics. He served as the Web Chair of the 7th
International Symposium on Modeling and Optimization in Mobile, Ad Hoc,and Wireless Networks (WiOpt 2009).
Yung Yi (Sβ04-Mβ06) received his B.S. and the M.S.in the School of Computer Science and Engineer-ing from Seoul National University, South Koreain 1997 and 1999, respectively, and his Ph.D. inthe Department of Electrical and Computer Engi-neering at the University of Texas at Austin in2006. From 2006 to 2008, he was a post-doctoralresearch associate in the Department of ElectricalEngineering at Princeton University. Now, he is anassistant professor at the Department of ElectricalEngineering at KAIST, South Korea. His current
research interests include the design and analysis of computer networking andwireless communication systems, especially congestion control, scheduling,and interference management, with applications in wireless ad hoc networks,broadband access networks, economic aspects of communication networks(aka network economics), and green networking systems. He has beenserving as a TPC member at various conferences such as ACM Mobihoc,Wicon, WiOpt, IEEE Infocom, ICC, Globecom, ACM CFI, ITC, the localarrangement chair of WiOpt 2009 and CFI 2010, and the networking areatrack chair of TENCON 2010.
Song Chong (Sβ93-Mβ95) received the B.S. andM.S. degrees in Control and Instrumentation En-gineering from Seoul National University, Seoul,Korea, in 1988 and 1990, respectively, and the Ph.D.degree in Electrical and Computer Engineering fromthe University of Texas at Austin in 1995. SinceMarch 2000, he has been with the Department ofElectrical Engineering, Korea Advanced Institute ofScience and Technology (KAIST), Daejeon, Korea,where he is a Professor and the Head of the Commu-nications and Computing Group of the department.
Prior to joining KAIST, he was with the Performance Analysis Department,AT&T Bell Laboratories, New Jersey, as a Member of Technical Staff. Hiscurrent research interests include wireless networks, future Internet, andhuman mobility characterization and its applications to mobile networking. Hehas published more than 90 papers in international journals and conferences.
He is an Editor of Computer Communications journal and Journal ofCommunications and Networks. He has served on the Technical ProgramCommittee of a number of leading international conferences including IEEEINFOCOM and ACM CoNEXT. He serves on the Steering Committee ofWiOpt and was the General Chair of WiOpt β09. He is currently the Chairof Wireless Working Group of the Future Internet Forum of Korea and theVice President of the Information and Communication Society of Korea.