Académique Documents
Professionnel Documents
Culture Documents
Abstract—Device-to-device (D2D) communications underlaying proper mode selection [6] and advanced (network) coding
a cellular infrastructure has recently been proposed as a means schemes [7]. However, to the best of our best knowledge,
of increasing the cellular capacity, improving the user through- prior works have not proposed a global optimization model that
put and extending the battery lifetime of user equipments by
facilitating the reuse of spectrum resources between D2D and takes into account mode selection, OFDM resource assignment
cellular links. In network assisted D2D communications, when and power allocation in an unified framework with the aim
two devices are in the proximity of each other, the network can of minimizing the overall power consumption in a multicell
not only help the devices to set the appropriate transmit power OFDM system.
and schedule time and frequency resources but also to determine Therefore, the purpose of the current paper is to develop
whether communication should take place via the direct D2D
link (D2D mode) or via the cellular base station (cellular mode). a unified optimization problem that minimizes the used sum
In this paper we formulate the joint mode selection, scheduling power in an OFDM system that may reuse PRBs for D2D links.
and power control task as an optimization problem that we Although the proposed centralized mathematical model turns
first solve assuming the availability of a central entity. We out to be strongly NP-hard and the corresponding resolution
also propose a distributed suboptimal joint mode selection and methods might be cumbersome and not-practically useful, it
resource allocation scheme that we benchmark with respect to the
centralized optimal solution. We find that the distributed scheme serves as a benchmark for the development of distributed
performs close to the optimal scheme both in terms of resource heuristic and power control schemes in D2D multicell systems.
efficiency and user fairness. By adopting the proposed framework, we gain insight in the
potential gains of using the direct D2D link as compared to
I. I NTRODUCTION using cellular links between two communicating UEs (Tx UE
Device-to-device (D2D) communications supported by a - Rx UE) when transmitting in both (i.e. cellular and D2D)
cellular infrastructure holds the promise of four types of gains. operational modes.
The proximity of user equipments (UE) may allow for very Although conceptually D2D links may use downlink (DL) as
high bit rates, low delays and low power consumption [1]. well as uplink (UL) cellular resources, UEs transmitting in DL
Secondly, the hop gain refers to using a single link in the spectrum may not be acceptable by regulatory bodies in certain
D2D mode rather than using an uplink and a downlink resource geographical regions. Therefore, in this paper we focus on an
when communicating via the base station (BS) in the cellular OFDM system in which the same uplink PRBs may be used
mode. Thirdly, the reuse gain implies that radio resources simultaneously for a cellular and one (or possibly more) D2D
may be simultaneously used by cellular as well as D2D links link(s) tightening the reuse factor below 1 (as in Figure 1). For
thereby tightening the reuse factor even of a reuse-1 system [1]. a particular UE pair, the sum power minimization problem is
Finally, the pairing gain refers to the degree of freedom of combined with mode selection that determines whether the UE
selecting the UEs communicating with the BS and the UE pair (Tx UE - Rx UE of Figure 1) should use the direct D2D
pairs using a direct D2D link that should use the same time link or they should communicate via the cellular BS. Therefore,
and frequency resources. Additionally, D2D communications we compare the performance of these two communications
may also facilitate new types of wireless peer-to-peer services modes, i.e. cellular mode and D2D mode, and provide a
[1], [2]. simulation analysis of the impact of resource sharing, when
However, D2D communications utilizing cellular spectrum the positions of both the D2D pair and the cellular UE vary
poses new challenges, because relative to cellular communica- within the cells.
tion scenarios, the system needs to cope with new interference
situations. For example, in an orthogonal frequency division II. S YSTEM M ODEL
(OFDM) system in which D2D communication links may We consider the uplink of an OFDM cellular system with
reuse some of the OFDM physical resource blocks (PRB), N cells and a complete reuse of the available frequency
intra-cell interference is no longer negligible. Solution ap- resources among different cells and possibly also within the
proaches to deal with this problem include power control same cell due to D2D communications. In each cell a base
[3], [4], various interference avoiding multiple-input-multiple- station (BS) controls a fixed number of UEs consisting of
output (MIMO) techniques [5] that can be combined with cellular UEs (c-UEs) and D2D transmitters. While all c-UEs
UE2 = Tx UE III. A CENTRALIZED JOINT MODE SELECTION , RESOURCE
ALLOCATION , AND POWER ASSIGNMENT FRAMEWORK
AP1 AP2
According to the underlying D2D infrastructure, we select
UE1
two possible transmission modes, i.e. q = 0 in which a given
UE communicates with the serving BS (cellular mode) and q =
UE3 1 in which a D2D pair communicates directly either sharing
Rx UE
PRBs or with an exclusive assignment of PRBs (D2D mode).
Figure 1. Illustration of D2D communications, when a user equipment (UE1) Obviously, for c-UEs, q is by definition always 0, while for
and a D2D pair (Tx UE - Rx UE) may use the same OFDM PRB. Due to D2D pairs mode selection can select q = 0 or q = 1. Let
the D2D link, intracell interference as well as intercell interference between
D2D and cellular links (UE3 to Rx UE) can be very high. (In this example xl,f (q) be a binary variable equal to 1 if user l is assigned
assuming that the D2D link uses cellular UL resources.) PRB f with mode q and 0 otherwise, and let Pl,f (q) ∈ R be a
positive real variable denoting the transmission power allocated
to user l on PRB f when adopting mode q.
Accordingly,
n
G if user l adopts mode q = 0 on PRB f
Gh(l),f (q) = Gn(l),f if user l adopts mode q = 1 on PRB f
d(l),f
communicate directly with the BS, D2D pairs can either exploit A Mixed Integer Linear Programming (MILP) formulation
for jointly optimizing mode selection, resource allocation and
the direct link or communicate via the BS (see Figure 1). We power assignment (JOMSRAP) can be derived as follows
represent the cellular network topology by a graph in which X
BSs belonging to the set B are labeled with n = 1, . . . , N ; minimize
l,f,q
Pl,f (q) (2)
UEs belonging to the set M with l = 1, . . . , L. We denote with subject to Pl,f (q) ≤ Kxl,f (q) ∀l, f, q (C1)
U ⊂ M and D ⊂ M the set of c-UEs and D2D transmitters X
Gh(l),f (q)Pl,f (q) − tgt
γl,f Gjh(l),f (q)Pj,f (q)
respectively. Hence, we define with h(l) the receiver of any j6=l,q
UE l, i.e., the serving BS n(l) for any c-UE l ∈ U and either ≥ γl,f
tgt
σ(1 − K(1 − xl,f (q))) ∀l, f, q (C2)
the serving BS n(l) or the D2D receiver d(l) for any D2D X
xl,f (0) ≤ 1 ∀n, f (C3)
transmitter l ∈ D. X
l∈n(l)
We consider the total frequency bandwidth W divided into xl,f (q) ≥ Rlmin
∀l (C4)
f,q
F PRBs each of bandwidth B = W/F . Each UE l can be tgt
γl,f σ
assigned to any of the f = 1, . . . , F PRBs belonging to the set Pl,f (q) ≥ xl,f (q) ∀l, f, q (C5)
Gh(l),f (q)
F. We model the effective channel gain between the transmitter X
l and its receiver when using PRB f as Gh(l),f . Let σ be the Pl,f (q) ≤ P max ∀l, q (C6)
f,q
thermal noise at the PRB operation bandwidth B = W/F and X
xl,f (q) ≤ 1 ∀l, f (C7)
Pl,f be the actual transmission power of user l on PRB f . q
The perceived signal to noise and interference ratio (SINR) on xl,f (1) = 0 ∀l ∈ U , f (C8)
PRB f for user l is: xl,f (q) ∈ {0, 1} ∀l, f, q (C9)
Gh(l),f Pl,f The objective function of JOMSRAP minimizes the overall
γl,f (pf ) = (1)
Il,f (pf ) + σ transmission power. In Constraints (C1) and (C2), K is a
large positive number. Specifically, constraints (C1) force the
where pf = (P1,f , . . . , PL,f )T is the vector of powers al-
transmission power of user l on a given PRB f to be 0
located to UEs overall the P cellular network and transmitting
j in case xl,f (q) = 0. Constraints (C2) can be derived by
on PRB f , Il,f (pf ) = j6=l h(l),f Pj,f is the interference
G
direct inspection of Equation 1 and they state that if user l
experienced at the receiver h(l) of user l, and Gjh(l),f accounts is assigned to PRB f with mode q = 1, i.e. xl,f (q) = 1,
for the effect of distance-based attenuation, beamforming, its transmission power must guarantee that γl,f (q) ≥ γltgt
fading etc., between h(l) and the interferer j transmitting on (see Equation (1)). Otherwise, if xl,f (q) = 0, the large
PRB f . Neglecting the details of practical coding schemes, we constant K ensures that constraints (C2) are always satisfied.
employ the Shannon capacity as a measure of the achievable bit Constraints (C3) guarantee the exclusivity of PRBs assignment
rate. Specifically, given a certain SINR, the spectral efficiency for UEs selecting mode q = 0 and belonging to the same
in bit/s/Hz for user l employing PRB f is ηl,f = log2 (1 + cell. Constraints (C4) impose that at least Rlmin PRBs must be
γl,f (pf )) and the maximum achievable theoretical rate in bit/s allocated to each user l ∈ M . Constraints (C5) are redundant
is Rl,f = Bηl,f . For the sake of tractability, in this paper but enforce Constraints (C2). Constraints (C6) ensure that each
we disregard link adaptation (LA). Thus, we assume that user l ∈ M does not consume more than P max power to satisfy
PRB f allows user l to transmit with a spectral efficiency its rate request. Finally, Constraints (C7) and Constraints (C8)
tgt tgt
ηl,f = log2 (1 + γl,f ) if the experienced γl,f (pf ) exceeds a state that each user l can exploit PRB f by adopting one mode
tgt
fixed target γl,f . Otherwise, user l cannot transmit on PRB f . at most, and each UE l ∈ U is forced to adopt cellular mode
tgt
We impose equal γl,f for each user-PRB pair. q = 0.
2
From JOMSRAP, we note that each PRBs can be assigned payoff function is the utility function in which PRB costs, i.e.
to at most one c-UE within a cell in order to avoid intra- the power vector, are iteratively computed at each iteration (t)
cell interference among c-UEs, while we let D2D users decide from the interference perceived in each PRB at the iteration
whether share PRBs or not. Hence, while in classical OFDMA (t−1) while the allocation strategies of all the other players are
cellular networks the multiple access interference (MAI) can kept fixed. In this framework the strategy profile is the set of
only affect users belonging to different cells, in this scenario all possible mode selection and PRB assignment that each BS
MAI can jeopardize the system performances even within a can adopt. The overall objective is to reach a Nash equilibria
cell. To this end, we note that in classical uplink cellular in which each player, given the strategy to the other players,
systems, interfering users are located at a distance that amounts does not gain any reward by making individual changing in its
at least to the radius of the cell, while in D2D communications own strategy. Unluckily, a Nash equilibrium is not guaranteed
MAI can operate at any distance thus leading to a potentially to exist when equal SINR target are fixed [10], [11], [12]. The
much heavier impairment of performances. natural consequence is that the resource allocation procedure
Proposition 1: JOMSRAP is NP-hard in the strong sense. outputs unstable PRBs assignment.
Proof: The proof follows by reducing the strongly NP- To this end, we propose a load control (LC) policy that pro-
hard 3-partition problem [8] to JOMSRAP. gressively reduces the load of the network (see Algorithm 1).
LC runs for each PRB f and iteration t. In particular LC
IV. A DISTRIBUTED SUBOPTIMAL MODE SELECTION AND reduces the SINR target of those users that consume the most
RESOURCE ALLOCATION SCHEME and in turn interfere the most and cause higher instability.
As showed in Proposition 1, the resolution of JOMSRAP Hence, LC tends to penalize some users leading to unfair
by means of a brute-force approach can be demanding in traffic configuration. To prevent this effect, the LC mechanism
terms of computational complexity that renders it infeasible is activated by taking into account 1) the amount of instability
in most scenarios in practice. This calls for the development that the cell induces and 2) the user to be penalized is the one
of more practical distributed algorithms that can be applied that can potentially experience the highest benefit in terms of
in a real-world system. To this end, we reduce JOMSRAP to SINR reduction.
a single-cell mode selection and resource allocation problem The resource allocation and LC strategy are repeated until a
(S-MSRAP) in which each BS is in charge of selecting the steady PRB assignment and power allocation are reached. The
proper mode and PRB to each mobile. derivation of ∆Pl and ∆ηl are provided in the Appendix.
X
minimize Pl,f (q)xl,f (q) (3) Algorithm 1 Load control for S-MSRAP.
l,f,q
X 1: ∆ < 1
subject to xl,f (0) ≤ 1 ∀f (C10) 2: if allocation is still changing after Nini slots and
n(l)
X redSlots(cell) = 0 then
xl,f (q) ≥ Rlmin ∀l (C11) unst ← Set of users in cell still
f,q 3:
P oscillating
P
X redP rob ←
(cell) (cell)
xl,f (q) ≤ 1 ∀l, f (C12)
4: l∈unst Pl / l Pl
q 5: ← rand
xl,f (q) ∈ {0, 1} ∀l, f, q (C13) 6: if < redP rob then
7: Compute for each l ∈ cell ∆Pl ← |Pl0 − Pl | and ∆ηl ←
Similarly to related constraints in Problem (2), con- |ηl0 − ηl |
straints (C10) impose the exclusivity of PRB assignment for 8: Compute for each l ∈ cell the benefit hl ← ∆Pl /∆ηl
9: Select h? ← max {h}
UEs, constraints (C11) impose that at least Rlmin PRBs must
? ← γh? ∆
γhtgt tgt
10:
be allocated to each user in the cell, constraints C12 state that 11: redSlots(cell) ← NredSlots
each user l can exploit PRB f by adopting at most one mode. 12: end if
The computation burden of JOMSRAP is distributed among 13: else
the BSs that have now to solve a simple polynomial algorithm. 14: redSlots(cell) ← max {redSlots(cell) − 1, 0}
15: end if
Proposition 2: The linear programming (LP) problem ob-
tained by relaxing constraints (C13) has an integer optimal
solution
Proof: Constraints (C10) - (C12) can be grouped as V. N UMERICAL RESULTS
3
0.08 0.08
100
0.06 0.06
Power tot. [W]
0.02 0.02
−50
0 0 −100
400 400 400
Loss≈ 137% 300
300 200 300 200
150 150 −150 200
200 200
100 100 0 100
100 100 50
50 50 100
150 0
0 0 0 0 200 RUE [m]
RUE [m] RD2D [m] RUE [m] RD2D [m]
RD2D [m]
(a) Power consumption in cellular mode. (b) Power consumption in D2D mode. (c) Gain of D2D over cellular mode.
Figure 2. Comparison between cellular and D2D mode: the percentage gain of D2D against cellular mode is about 100% for RD2D = 20m and RUE = 50m,
while progressively reduces as RD2D and RUE increase. The advantage becomes negligible and not effective for RD2D ≥ 120m.
100
0.08 100
80
Mode Selection Gain [%]
60
0.04
40 40
Gain ≈ 99.8% Gain ≈ 86%
0.02 20
20
0
0 0 Gain ≈ 3.6% 400
400 0 Gain ≈ 3.5%
300 200 300 200
50
200 150 100 200 150
100 350 400 100
100 150 250 300 100
50 200 50
100 150
RUE [m] 0 0 200 50 RUE [m] 0 0
R [m] R [m]
D2D RD2D [m] RUE [m] D2D
(a) Power consumption. (b) Mode selection vs. Cellular Mode. (c) Mode selection vs. D2D Mode.
Figure 3. Power consumption when both cellular and D2D mode are available: compared with Figure 2, mode selection allows to drastically improve the
system performances.
22.2
22
Tot. Spectral efficiency [b/s/Hz]
0.25
1
21.8
0.2
21.6
Power tot. [W]
0.15 0.995
Jain´s Fairness Index
21.4
0.1
21.2
0.99
0.05 21
0 20.8
400 0 0.985 0
300 200 50
200 150 100 100
100 0.98
100 150 350 400
50 250 300
200 150 200 50 100 150 200 200
RUE [m] 0 0 50 100 250 300 350 D2D radius [m]
RD2D [m] 400
D2D radius [m] RUE [m]
R [m]
(a) Power consumption. (b) Total spectral efficiency. (c) Jain’s fairness index [13].
Figure 4. Performances of S-MSRAP with LC.
when the optimal centralized JOMSRAP (2) is used. To solve an accurate exploitation of the multi-user diversity and mode
JOMSRAP, RUNE has been integrated with the optimization selection. Specifically, the gain of mode selection strategy
solver I LOG C PLEX 12.2. Hence, we perform extensive sim- over a conventional cellular communication can be signifi-
ulations varying the distance RUE between each c-UE and cant for RD2D ≤ 150m and become less effective only for
the serving BS, as well as the proximity RD2D between higher distances. The impact of mode selection is analyzed
a D2D pair. For each {RUE , RD2D }, we average over 500 in Figure 5 and Figure 6. In particular, Figure 5 illustrates
feasible Monte Carlo experiments and we consider the channel the percentage of selecting cellular mode by D2D pairs, while
gains constant during the JOMSRAP procedure. In Figure 2, Figure 6 indicates the occurrence of resource sharing in D2D
we solve JOMSRAP by forcing cellular mode, i.e. q = 0 communications.
∀l (see Figure 2(a)), and D2D mode, i.e. q = 1 ∀l (see The performances of the distributed S-MSRAP with LC heuris-
Figure 2(b)), and compare the gain of D2D over conventional tic are depicted in Figure 4. Compared with the optimal results
cellular communications (see Figure 2(c)). Figure 3 depicts the provided by JOMSRAP (Figure 3(a)), the distributed scheme
advantage of the jointly mode selection, PRB assignment and achieves suboptimal performances in terms of power consump-
power allocation scheme. Such an advantage is triggered by tion (Figure 4(a)), but rather than relying on a central entity
4
Table I
S YSTEM PARAMETERS .
Parameters Summary 80
≈ 75.9
RcellU E 50m,100m,. . . ,400m 70
40
30
20