Scheduling and Power Control for In-Band Full

Duplex Communications
Rakshak Agrawal1 , Suresh Kalyanasundaram1 , Krishna Chaitanya A2
1
Nokia Networks, Bangalore Email: [email protected], [email protected]
2
Department of ECE, IIIT Sri City Email: [email protected]

Abstract—In this paper, we study user scheduling and power To make use of IBFD in a cellular network, several chal-
control in a multi-cell scenario where the base station (BS) lenges need to be addressed. The most significant challenge
is capable of operating in in-band full duplex (IBFD) mode. is to reduce the co-channel interference due to both UL
By IBFD we mean that the BS can transmit and receive
simultaneously in the same frequency band through self inter- and DL users. IBFD introduces the following new additional
ference cancellation (SIC). We have studied two categories of interference scenarios: 1) self-interference from transmitter
scheduling schemes in a small cell environment: non-cooperative (Tx) to receiver (Rx) at the BS 2) intra-cell and inter-cell
(NC) and cooperative (CO). In NC scheduling scheme, each BS interference from UL transmissions to DL receptions (UE-UE
chooses an allocation such that the sum of the downlink and interference) 3) inter-cell interference on UL reception at the
uplink proportional fair (PF) metrics computed within the cell
is maximized. In CO scheduling scheme, all BSs co-operate to BS due to DL transmissions from neighboring BSs (BS-BS
choose an allocation that maximizes the total PF metric summed interference). Due to these additional interferences, the full
over downlink and uplink across all the cells in the cooperation duplex communication opens a new range of issues that needs
cluster. We present simulation results comparing the performance detailed study.
of both NC and CO schemes with and without power control. Based on recent studies [6], some of the proposed topologies
While co-operation provides additional gains, the computational
complexity of such a centralized scheme is very large. Therefore, for using IBFD in wireless networks are: 1) relay topology 2)
we recommend the NC schemes which offer a good trade-off bidirectional topology and 3) base station topology. In this
between performance and complexity. paper we focus on the base station topology, where only the
Index Terms—full duplex communication, cellular networks, BS is capable of full duplex transmission. Studies based on BS
5G topology for small cells [8] [9] [10] have already highlighted
the benefits of full duplex. Most of the work is focused on
I. I NTRODUCTION gains in session throughput, sum throughput or delay[11]. In
In cellular networks, traditionally the uplink (UL) and this paper, we have designed scheduling and power control
downlink (DL) transmissions are separated in time or fre- schemes to maximize the geometric mean of UE throughputs.
quency. The reason for using time or frequency division We use a hybrid scheduling scheme in this paper, where the
duplexing (TDD or FDD) as modes of transmission is due BS will schedule both UL and DL simultaneously only if the
to the limitation that a transceiver cannot transmit and receive sum PF metric is greater than that calculated by scheduling
in the same frequency simultaneously [1]. Recent advances HD (DL or UL only). We have proposed new algorithms in
in self interference cancellation (SIC) [2], [3] and [4] has Sections V and VI based on non-cooperative (NC) and co-
overcome this limitation and has drawn considerable interest operative (CO) schemes for scheduling in IBFD. We have
of researchers towards in-band full duplex (IBFD) mode of compared the performance of these NC and CO full duplex
transmission. IBFD is a promising technology as a node can scheduling algorithms with half-duplex mode. We have also
transmit and receive simultaneously over the same frequency evaluated the scheduling algorithms with different ratios of UL
band, which, at best, can double the DL and UL throughputs. and DL UEs per cell.
But, realistically, the achievable UL and DL gains mainly The paper is structured as follows. Section II presents
depend on the SIC capability of the BS and UEs as well as system model and the notations used in this paper. Section
the overall inter and intra cell interference. Practical IBFD III illustrates the mathematical formulation of the scheduling
transceiver is now possible [5], especially for small cells, problem. Sections IV and V explain the proposed algorithms
as the terminals use smaller transmission powers, thereby used for non-cooperative and co-operative scheduling, respec-
reducing the self-interference problem. Even though many tively. System level simulation results are presented in section
design challenges have been overcome, IBFD with small form- VI, and the paper is concluded in Section VI along with a
factor devices still remains difficult. So initial deployments of brief mention of future work.
IBFD terminals are most likely to be at the small-cell BSs
only. This is also the assumption in our paper. Due to the II. S YSTEM M ODEL AND N OTATION

increasing use of small cells in the next generation wireless Let B = B (1) , . . . , B (N ) be the set of base stations,
architecture, IBFD has become more relevant for 5G cellular where N is the total number of cells we consider in the
networks. system. The set of uplink UEs in cell n is denoted by
 (n) (n)
interference as noise. Hence, assuming that um and dl are
scheduled for uplink and downlink transmission in tth TTI, the
rate of uplink transmission can be estimated using Shannon’s
capacity equation as
!
gu
(n) (n) (t)P (n)
B u
Ru(n) (t) = log 1 + m2 m
, (1)
σn + IB (n) (t)
where Pu(i) is the transmit power of uplink UE and IB (i) (t)
m
is the interference experienced at the BS in the given TTI.
Similarly, rate of downlink transmission in cell n in TTI t is
 
gB (n) d(n) (t)PB (n)
(n)
Rd (t) = log 1 + 2 l , (2)
σ (n) + Id(n) (t)
dl l

Fig. 1. Illustration of different types of inter-cell and intra-cell channel gains where PB (n) is the transmit power of BS and Id(n) is the inter-
to be considered in a Full Duplex System (n)
l

ference at the downlink UE dl . We discuss the interference

n o terms in equations (1) and (2) in detail in Section III.
(n) (n)
U (n) = u1 , . . . , uM (n) , n = 1, . . . , N . Similarly, the
III. P ROBLEM F ORMULATION
set of downlinko UEs in cell n is denoted by D(n) =
n
(n) (n) The most important part of problem formulation is to
d1 , . . . , dL(n) , n = 1, . . . , N . We denote the channel gain identify the metric used for measuring the performance of the
from transmitter i to receiver j by hij , for example, hu(n) B (k) algorithms. In this paper, we use the geometric mean (GM) of
i
(n) UE throughputs as the metric to compare different schemes.
denotes channel gain from uplink UE ui in cell n to BS in
cell k, B (k) . We denote the power of the channel gain hij by GM is the objective that the standard proportional fairness
gij , i.e., gij = |hij |2 . We assume that all base stations have metric (PF) attempts to maximize as explained in [13], [14]
knowledge of channel states of links to itself from all UEs and and [15]. The PF scheduler isP obtained as the solution to sum
all base stations. Similarly, downlink UEs have knowledge of utility maximization problem i log(R̄i ), where the logarithm
channel states of the links to itself from all base stations and of user-achieved throughput is used as the utility function. The
uplink UEs. We indicate the peak power of BS B (n) by P̄B , term R̄i is user i’s achieved throughput and the summation is
(n) over all the users across all cells. The gradient algorithm given
similarly the peak power of uplink user um by P̄u . Please
refer to Figure 1 and Table I for all notations. in [14] implies that the scheduling and resource P allocation

at each sub-frame should be done to maximize i Ri /R̄i ,
TABLE I
where the summation is over the scheduled users and subject to
N OTATIONS the available resource constraints. Here Ri is the instantaneous
data rate that can be achieved by UE i and the maximization
Notation Description is over all the users in each cell. The ratio Ri /R̄i is called
hBk Bl channel gain from BS in cell k to BS in cell l.
the PF metric of UE i. The schemes discussed in this paper
h (n) channel gain from BS in cell k to ith DL UE in cell n attempt to maximize the aggregate PF metric realized on a
Bk di
h (n) channel gain from UL UEi cell n to BS in cell k per sub-frame basis. Hence, comparison of these scheduling
Bk ui
schemes should be made on the basis of the geometric mean
h (l) (k) channel gain from lth UL UE in cell l to kth DL UE in
ui dj
cell j of UE throughputs realized by them.
γ fraction of self interference after SIC. We consider the problem of scheduling users in each TTI
P (n) transmit power used by the mth UL UE in cell n for uplink and downlink transmission in each cell. We denote
um
PB (n) transmit power used by BS in cell n.

a schedule in cell n by Sn = (u(n) (t), d(n) (t)) , where t
P̄ (n) peak transmit power used by the mth UL UE in cell n
um is the TTI index. A schedule of all the cells S is denoted by
P̄B (n) peak transmit power used by BS in cell n.
wk weight of the kth DL UE
(S1 , . . . , SN ). The average throughput Ru(n) and Rd(n) for
k l
vl weight of the lth UL UE uplink and downlink user respectively can be calculated using
L(n) number of DL UEs in cell n an IIR filter:
M (n) number of UL UEs in cell n
(n) (n) (n)
R̄m (t) = αR̄m (t − 1) + (1 − α)1R(n) (t)Rm (t), (3)
m
We consider a communication system in which time is (n)
divided into slots. In LTE, one time slot is referred to as where Rm (t) is the instantaneous rate and 1R(n) (t) is an
m
(n)
transmission time interval (TTI). One TTI is the smallest indicator function that is equal to 1 if the user Rm is
time frame in which a BS can schedule its UEs for uplink scheduled in TTI t, else it is equal to 0.
and downlink transmission. In each TTI, the receivers decode In this paper we consider non-cooperative as well as coop-
their messages from the corresponding transmitters by treating erative scheduling schemes. In case of non-cooperation, BS in
each cell takes decision on scheduling of its users without de- such that the weighted sum rate in TTI t solves the following
pending on the schedule in other cells. In cooperative schemes, optimization
decision on scheduling of users in all cells is taken together
so that the sum throughput is maximized. The transmission   
power of scheduled user and BS may be fixed or variable in
both co-operative and non-cooperative schemes.

  gu(n) B (n) (t)P̄u(n) 
k k
max  w (t)log 1 +
 
k
From the above discussion, we see that for a non- σn2 + I¯B (n) (t − 1) + γ(l6=φ) P̄B (n) 
 
(k,l) 
 
cooperative scheme the objective of the proportional fair | {z } | {z }
avg. inter-cell UL intf. self-intf.
scheduler for an in-band full duplex system should be to  
choose k and l in a cell so that

 gB (n) d(n) (t)P̄B (n) 

l
h i
1 + σ 2 + I (n) (n) (t) + I¯ (n) (t − 1)
+vl (t)log   .
(k ∗ , l∗ ) = arg max wk (t)Ru(n) (t) + vl (t)Rd(n) (t) , (4) (n) u d d

 dl 
(k,l) k l
| k {zl } | l {z }
intra-cell DL intf. avg. inter-cell DL intf.
n n
where UE pair (k, l) ∈ {φ} ∪ U × {φ} ∪ D . A choice (8)
of k or l equal to φ indicates that no user is chosen in
uplink or downlink, respectively. Therefore, when pair (k, l)
To find the schedule that maximizes the weighted sum rate
is selected such that an uplink and a downlink user chosen
in (8), base station B (n) needs to know the scheduling of
in each direction, it can be termed as full duplex (FD) mode.
users in other cells. But B (n) does not have this information
Similarly, if pair (k, l) is selected such that either only uplink
as that schedule is also being decided in the same TTI. For
or a downlink user is chosen, it can be termed as half duplex
this reason base station B (n) uses average interference terms
(HD) scheduling. The weights wk (t) and vl (t) in (4) can be
such as I¯B (n) and I¯d(n) which can be calculated based on
calculated from average rate in previous TTI as l
instantaneous interference by applying IIR filter as
1
wk (t) = , (5)
(n)
R̄k (t − 1) I¯B (n) (t) = β I¯B (n) (t − 1) + (1 − β)IB (n) (t), (9)
1
vl (t) = (n)
. (6)
R̄l (t − 1) where IB (n) (t) is the instantaneous uplink interference at
BS B (n) caused by uplink transmission by UEs as well as
downlink transmission by BSs in other cells. This inter-cell
For co-operative schemes, we can obtain a proportionally
interference at BS B (n) can be computed as
fair schedule S by maximizing the objective in (4) but summed
over n = 1, . . . , N . In other words, the scheduling of users in
all cells (k1∗ , l1∗ ), . . . , (kN
∗ ∗
, lN ) is N
X M (j)
X
IB (n) (t) = 1u(j) (t)gu(j) B (n) (t)P̄u(j)
k k k
j=1,j6=n k=1
(k1∗ , l1∗ ), . . . , (kN
∗ ∗
, lN )=
N L(j)
N h X X
+ 1d(j) (t)gB (j) B (n) (t)P̄B (j) , (10)
X i
arg max wkn (t)Ru(n) (t) + vln (t)Rd(n) (t) , k
(k1 ,l1 ),...,(kN ,lN ) n=1 kn ln j=1,j6=n k=1
(7)
where 1u(j) (t) is an uplink schedule indicator function that is
n n k
where UE pair (kn , ln ) ∈ {φ} ∪ U × {φ} ∪ D is selected (n)
equal to 1 if uk is scheduled in TTI t, else it is equal to 0.
for each cell appropriately. Please note that a selection of Similarly, 1d(j) (t) is an indicator function for downlink sched-
both kn and ln equal to φ is possible when cell muting is k
ule. The first term in (10) corresponds to the interference due
enabled, i.e., there is no uplink or downlink user scheduled in
to the uplink transmission in the neighboring cells. The second
a particular cell. Enabling cell muting may help in reducing
term is the interference due to the downlink transmissions in
inter-cell interference. The following sections dig deeper into
the neighbouring cells.
the schemes used for simulation in this paper.
In the second part of (8), the term Iu(n) d(n) (t) corresponds
k l
to downlink intra-cell interference from UL UE k to DL UE
IV. N ON - COOPERATIVE SCHEDULING l and its value is equal to gu(n) d(n) (t)P̄u(n) . The term γ P¯B
k l k
is the residual interference after self interference cancellation,
In this section, we present non-cooperative schemes to find i.e., 1 − γ is the factor by which self interference is can-
a proportionally fair schedule Sn for each n = 1, . . . , N. In a celed. Average interference at downlink user I¯d(n) (t) can be
l
fixed power based scheme, the BS B (n) in cell n chooses users calculated from instantaneous interference Id(n) (t) in similar
l
manner as (9). This instantaneous inter-cell interference at the the optimum threshold is chosen, it can compensate for the
(n)
downlink user dl can be computed as wrong decision of choosing FD over HD in non-cooperative
scheduling.
N M (j)
X X
Id(n) (t) = 1u(j) (t)gu(j) d(n) (t)P̄u(j) Algorithm 1 Non-cooperative scheduling in cell n
l k k l k
j=1,j6=n k=1
N L(j) Compute weighted rate P F (n) (t) for every pair of uplink
+
X X
1d(j) (t)gB (j) d(n) (t)P̄B (n) , (11) and downlink users (k, l) from (8)
j=1,j6=n k=1
k l Choose the pair (k, l) that maximizes P F (n) (t) and denote
it as (k̂, ˆl) and term the objective function value for that
(n)
where 1d(n) (t) is an indicator function to indicate whether dl pair as RF∗ D .
k
is scheduled or not. The first term in (11) is interference due to Compute P F (n) (t) for each uplink user k with no downlink
the uplink transmissions in neighboring cells. The interference user active. Let k̃ be the uplink user that maximizes the PF
from downlink transmissions in neighboring cells is captured metric and denote the maximum value by R̃U .
in the second term of (11). Compute P F (n) (t) for each downlink user l with no uplink
The average interference at a downlink user is updated even user active. Let ˜l be the downlink user that maximizes the
when it is not scheduled. This way we can obtain a better PF metric and denote the maximum value by R̃D .
estimate of the interference in the current TTI. Note that in if RF∗ D > η × max(R̃U , R̃D ) then
the non-cooperative scheme, while an average value of the select full duplex mode
inter-cell interference is used, the intra-cell interference due else
to the scheduling of an uplink user on the downlink user if R̃D > R̃U then
(Iu(n) d(n) (t)) and the self-interference (γ P̄B ) are taken into select DL only mode
k l
account while determining the schedule. This method ensures else
that there is a good spatial separation between co-scheduled select UL only mode
downlink and uplink users in the same cell. end if
We note that equations (8), (10) and (11) consider that the end if
(n) (n)
transmit power of all the UEs and BSs is fixed. If power Update R̄k (t), R̄l (t), I¯d(n) (t) and I¯d(n) (t) for all users
k l
control is enabled corresponding equations can be derived by in cell n.
1 1
replacing peak powers by instantaneous powers. The optimiza- Compute wk (t + 1) = (n) , vl (t + 1) = (n) .
R̄k (t) R̄l (t)
tion problem to be solved with variable powers is
 !
gu(n) B (n) (t)Pu(n) In our study we have simulated the following non-
max wk (t)log 1 + 2 ¯ k k
cooperative schemes:
σn + IB (n) (t − 1) + γ(l6=φ) PB (n)
  1) NCHD (Non-Cooperative Half Duplex scheme): In this
gB (n) d(n) (t)PB (n) scheme, each cell is operating in half duplex mode
l
+ vl (t)log 1 + 2  ,
and is independently scheduling an UL or a DL UE
σ (n) + I (n) (n) (t) + I¯ (n) (t − 1)
uk dl dl
dl depending on the direction that has the larger PF metric
(12) as given in (8). This scheme is treated as the baseline
where the maximization is over Pu(n) , PB (n) and (k, l). for comparison of gains of full duplex over half duplex.
k
In Algorithm 1, we present the detailed steps to compute the 2) NCFD (Non-cooperative Full Duplex scheme): In this
schedule that maximizes the PF metric in a given TTI. In this scheme we find the schedule based on Algorithm (1)
algorithm, we follow a hybrid scheduling scheme by compar- where BS and UE powers are fixed as in (8).
ing the PF metric realized by the full duplex and half duplex 3) NCFD-PC (Non-cooperative Full Duplex scheme with
modes. In order to reduce the effect of inter-cell interference, Power Control): In this scheme we find the schedule
we have introduced the threshold η which compensates for the based on Algorithm (1) where uplink and downlink
average inter-cell interference while choosing among FD and powers are variable as in (12). The power values are
HD mode. If the PF metric of half duplex mode is larger than chosen from a range of six equidistant discreet points in
that of full duplex mode by a factor of eta, then either DL [0, P̄u(n) ] for uplink in [0, P̄B (n) ] for downlink.
k
or UL UE is scheduled depending on whether the DL PF or
UL PF is larger. Otherwise, the full duplex mode is selected. V. C OOPERATIVE SCHEDULING
The performance of the algorithm depends on the appropriate
choice of η. If the threshold is too low, then FD mode is In this section, we discuss cooperative scheduling in which
scheduled more often than optimally required which means users are scheduled to maximize the sum PF of all the cells
due to inter-cell interference, the sum rate may be lesser than in a cooperation cluster. In this scheduling, we also consider
optimal. If the threshold is too high, then FD is scheduled power allocation that depends on the channel states and the
too less and the gain over HD remains quite low. Hence, if TTI in which scheduling is being decided.
 For co-operative scheduling without power control the TABLE II
schedule in TTI t can be obtained by solving S IMULATION PARAMETERS
 
N
! Parameter Value
X gu(n) B (n) (t)P̄u(n)
k k
max  wkn (t)log 1 + System Bandwidth 10 MHz
n=1
σn2 + IB (n) (t) + γ(ln 6=φ) P̄B (n)
  Carrier Frequency 2 GHz
gB (n) d(n) (t)P̄B (n) Inter-site distance (ISD) 50 m
l
+ vln (t)log 1 + 2  ,
Uplink UE Tx power 0.2 W
σ (n) + Iu(n) d(n) (t) + Id(n) (t)
dl k l l
Downlink UE Tx power range 0.2 W
(13)
Noise power 7.166 ×10−16
where the maximization is over (k1 , l1 ), . . . , (kN , lN ) as de- BS and UE Antenna Gain 10 dBi
fined in (7). To admit power control we model the combina-
BS and UE Noise Figure 9 dB
torial optimization (13) as a continuous optimization problem
Traffic Full Buffer
by writing power variables in place of peak powers. We add
constraints on power variables of uplink and downlink users Scheduling Wideband
so that only one uplink and/or one downlink user is scheduled. Pathloss between BS-UE (dB) RRH pathloss model
Thus the optimization problem of cooperative scheduling with Minimum distance between BS to UEs 3m
power control is UE Drop method Random
  !
XN gu(n) B (n) (t)Pu(n)
k k
max  wkn (t)log 1 +
2
n=1
σ n + I B (n) (t) + γ(l 6=φ) PB (n)
n is variable in this scheme. Cell muting is enabled in this
  scheme.
gB (n) d(n) (t)PB (n) The number of cells in a co-operation cluster is 7 in all
l
+ vln (t)log 1 + 2 
σ (n) + Iu(n) d(n) (t) + Id(n) (t) the above schemes. In the next section we compare the
dl k l l
results obtained for different Cooperative and Non-cooperative
scheduling schemes for full duplex.
subject to 0 ≤ Pu(n) ≤ P̄U , for all n = 1, . . . , N,
m

0 ≤ PB (n) ≤ P̄B , for all n = 1, . . . , N, VI. S IMULATION R ESULTS

Pu(n) ≥ 0, for all n, m. (14) In this section we summarize the results of our system level
m
simulations. We have considered the indoor RRH path-loss
where the interference terms IB (n) (t) and Id(n) (t) are given
l model, which is detailed in [16] and [17]. To ensure statistical
in (10)oand
n n (11),orespectively
n and the maximization
o is over reliability, we have simulated 3000 independent drops for
Pu(n) , PB (n) and (k1 , l1 ), . . . (kN , lN ) . each scenario. While the UEs are stationary with fixed path
k
In our study we have simulated the following co-operative loss values for a given drop, the impact of fast fading is
schemes: accounted for in our simulations. The parameters we used in
1) DTHD (Dynamic TDD based co-operative Half Duplex the simulations are summarized in Table II. In the following
scheme): This is a co-operative half duplex scheduling paragraphs, we study the impact of different parameters on the
scheme used as a baseline. In this scheme we maximize performance gains from IBFD.
the sum PF metric as in (13) where in any particular 1) Impact of ratio of number of DL and UL UEs: We
TTI, either all cells select an uplink user k (k ∈ U n ) compare the geometric mean of UE throughputs obtained
or downlink user (l ∈ Dn ). Since BSs co-operate with in full duplex scheme with that of the baseline half duplex
each other, the selection of uplink or downlink direction scheme. With the NCFD scheme, we observe gains in the
depends on the condition that sum PF metric of all the range of 15 − 35% in the geometric mean of UE throughputs
cells in that direction is better than that in the opposite when compared to NCHD. These gains are much lower than
direction. In this scheme, we assume that the power of the expected 100% because the performance is limited by
all BS and users are fixed. large intra-cell UE to UE interference. Another reason for
2) CoFD (Co-operative Full Duplex scheme): This is a the low gains with NCFD is the absence of cooperation
joint scheduling scheme where we maximize the sum among cells. Each cell’s scheduling decisions are independent
PF of all the cells as in (13). The BS and users are of the knowledge of possible interference from other cells.
assumed to have fixed power in this scheme. Therefore, as explained in Algorithm 1, we use a threshold η to
3) CoFD-PC (Co-operative Full Duplex Scheme with compensate for this lack of knowledge about the interference
Power Control): In this scheme we solve the constrained from other cells. Optimal threshold η is chosen by sweeping
optimization problem in (14). The aim is to find a joint across a discrete set of values in the range from 0 to 2. The
schedule that can maximize the combined sumPF metric value that gives the maximum gains, which is 1.6 in our
of all the cells. The power of each user and base station simulations, is chosen for all comparisons. Figure 2 shows
the GM gains for 105 dB SIC as a function of the threshold
η. We see that the largest gains in GM is obtained at around a
threshold value of 1.6 across all values of the number of DL
to UL users simulated.

Fig. 4. Scheduling percentage for different DL:UL ratio and different SIC
values.

Fig. 2. Maximum gains in geometric mean for NCFD vs NCHD are seen
when for η = 1.6. when FD is scheduled. To compensate for this increase in UL
throughput, DL Only mode is scheduled more often when SIC
The results in Figure 3 and Figure 4 are generated by is infinity.
choosing the optimal threshold (η = 1.6). We can see from
Figure 3 that the overall gain obtained in geometric mean by
using full fuplex scheduling varies depending on the number of
DL and UL UEs in each cell. The gains in the NCFD scheme
over NCHD is largest when the number of DL and UL UEs
is approximately the same. We also notice that, for the same
DL to UL ratio, the gains are more with larger values of SIC.
This is as expected because the larger the self interference
cancellation the lower the UL interference leading to higher
data rates. From Figure 4 we see that the percentage of times

Fig. 5. Comparison of relative geometric mean of different scheduling

schemes.

2) Impact of co-operation between BSs: Figure 5 gives the

relative GM obtained by using different co-ordinated and non-
coordinated scheduling schemes. These results are taken for
SIC level of 105dB with 1DL and 1UL UE per cell. We can
see that with co-operative scheme (CoFD) we can get much
better gains over half duplex schemes. It can be also confirmed
that gains for co-operative scheme increases as number of
DL and UL UEs increase but the complexity also increases
Fig. 3. Percentage gain in the geometric mean by using NCFD over NCHD
for different number of DL and UL users and for different SIC values. simultaneous. We can see from Figure 6 that percentage of
times when FD mode was selected is higher for co-ordinated
FD mode is selected is more when the number of UEs in scheduling schemes which also explains better gains.
DL and UL are approximately the same. Although the trend 3) Impact due to power control: With power control we
is clearly visible for a given SIC level, when we compare the observe high gains in geometric mean of UE throughput in full
values across different SIC levels, we notice that there is slight Duplex based schemes as compared to half duplex schemes.
reduction in FD schedule % for higher SIC values. The reason Most of the benefit of using power control is seen in UL
for this behavior is that due to the larger SIC, the interference direction as UL UE transmission causes maximum interference
in UL is lower, which helps to increase the UL throughput to the UE receiving DL data from BS. The impact of using
 observed that power control in uplink direction contributes
to majority of improvements in power control scheme as
compared to downlink direction.
In future we would like to explore a distributed algorithm
for co-operative scheduling rather than the current centralized
scheme which is highly complex in nature.
R EFERENCES
[1] A. Goldsmith, Wireless Communications. New York, NY, USA: Cam-
bridge University Press, 2005.
[2] D. Bharadia, E. McMilin, and S. Katti, “Full Duplex Radios,” ACM
SIGCOMM 2013, Hong Kong, 2013.
[3] J. Choi et al., “Achieving Single Channel, Full Duplex Wireless Com-
munication,” ACM MOBICOM 2010,Chicago, IL, 2010.
[4] M. Jain et al., “Practical, Real Time Full Duplex Wireless,” ACM
MOBICOM 2011, Las Vegas, NV, 2011.
Fig. 6. Comparison of percentage of scheduling in different scheduling
[5] M. Heino et al., “Recent advances in antenna design and interference
schemes.
cancellation algorithms for in-band full duplex relays,” IEEE Commu-
nications Magazine, vol. 53, no. 5, pp. 91101, 2015.
[6] A. Sabharwal et al. “In-band full-duplex wireless: Challenges and
power control in downlink direction is very limited. From opportunities,” IEEE Journal on Selected Areas in Communications 32.9
Figure 7 we can see that with power control in non-cooperative (2014): 1637-1652.
[7] S. Goyal et al., “Full duplex cellular systems: will doubling interference
scheme (NCFD-PC), we could observe higher gains in GM prevent doubling capacity?,” IEEE Communications Magazine, vol. 53,
over NCHD. With cooperative scheduling using power control no. 5, pp. 121127, 2015.
(CoFD-PC) we could see 60 − 70% gains over half duplex [8] N. Mahmood, L. Garcia, P. Popovski, and P. Mogensen, “On the
performance of successive interference cancellation in 5G small cell
scheme for 2 DL and 2 UL users per cell. The gains are networks,” in Proc. IEEE WCNC, Istanbul, Turkey, Apr. 2014.
expected to increase with higher number of users per cell. [9] M. Gatnau et al., “Full duplex communication under traffic constraints
for 5G small cells,” in IEEE 82nd Vehicular Tech. Conference, 2015.
[10] N. H. Mahmood et al., “Throughput analysis of full duplex communica-
tion with asymmetric traffic in small cell systems,” in 11th International
Conference on Wireless and Mobile Communications, 2015.
[11] M. Gatnau et al. “Can Full Duplex Boost Throughput and Delay of 5G
Ultra-Dense Small Cell Networks?,” IEEE 83rd Vehicular Technology
Conference (vtc Spring), 2016.
[12] A. Cirik et al. “A subcarrier and power allocation algorithm for OFDMA
full-duplex systems,” Networks and Communications (EuCNC), 2015
European Conference on. IEEE, 2015.
[13] F. P. Kelly, A. K. Maulloo and D. K. H. Tan, “Rate Control for
Communication Networks: Shadow Prices, Proportional Fairness and
Stability,” Journal on Operations Research Society, Vol. 49, pp. 237–
252, April, 1998.
[14] R. Agrawal and V. Subramanian, “Optimality of Certain Channel Aware
Scheduling Policies,” Proceedings of Allerton Conference on Communi-
cation, Control and Computing, October, 2002.
[15] R. Agrawal, A. Bedekar, R. La and V. Subramanian, “A Class and
Channel-Condition based Weighted Proportionally Fair Scheduler,” 17th
International Teletraffic Congress, December 2001.
[16] 3GPP, “Physical layer aspect for evolved universal terrestrial radio access
Fig. 7. Gain in geometric mean when comparing NCFD-PC with NCHD. (UTRA),” TR 25.814, Sep. 2006.
[17] 3GPP, “Further advancements for E-UTRA physical layer aspects,” TR
36.814, V9.0.0, 2010.
VII. C ONCLUSIONS AND F UTURE WORK
In this paper we present different co-operative and non-
cooperative scheduling schemes in full duplex enabled BS in
small cell scenario. We compare gains in geometric mean
in each of these schemes over baseline half duplex com-
munication schemes. We see that co-operative scheme gives
higher gains but non-cooperative scheduling schemes give a
better trade-off between complexity and gains. It is also seen
that the gains in geometric mean are highest when number
of downlink and uplink users in each cell are equal. We
notice that percentage of times full duplex mode is selected
by the scheduler is proportional to gains in geometric mean.
We observe higher gains with power control based schemes
as compared to that of fixed power based scheme. We also