GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 98

MOBILE CONVERGED NETWORKS

DEVICE-TO-DEVICE COMMUNICATION UNDERLAYING

CONVERGED HETEROGENEOUS NETWORKS
AMILA THARAPERIYA GAMAGE, HAO LIANG, RAN ZHANG, AND XUEMIN (SHERMAN) SHEN

ABSTRACT are covered by multiple wireless networks, such

as Long Term Evolution (LTE) cellular net-
To satisfy the ever increasing wireless service works and IEEE 802.11n wireless local area net-
demand, it is effective to form a converged net- works (WLANs). These networks offer different
work by utilizing interworking mechanisms, such advantages based on their diverse radio access
that the resources of heterogeneous wireless technologies. For example, cellular networks
networks can be allocated in a coordinated and support high mobility and guarantee QoS, while
efficient manner. Despite the potential advan- WLANs provide much higher data rates at a
tages of a converged network, its performance lower charge [1]. Therefore, by forming a con-
needs further improvement, especially at cell verged network utilizing interworking mecha-
edges and rural areas where only one network is nisms [2], resources available in these
available. In this article, we investigate how to heterogeneous networks can be jointly allocated
leverage device-to-device, D2D, communication in an optimal manner to improve the data rates
to further improve the performance of a con- and QoS support [1–4]. Furthermore, such joint
verged network which consists of an LTE-A cel- resource allocation widens the coverage area by
lular network and IEEE 802.11n WLANs. Three effectively merging the individual network cover-
main technical challenges that complicate age areas. In a converged network, multi-homing
resource allocation are identified: allocation of capability of user equipments (UEs) allows users
resources capturing diverse radio access tech- to simultaneously communicate over multiple
nologies of the networks, selection of users’ networks using the multiple radio interfaces
communication modes for multiple networks to available at UEs. With UE multi-homing,
maximize hop and reuse gains, and interference resource utilization can be further optimized as
management. To address these challenges, we the user requirements can be satisfied using
propose a resource allocation scheme that per- resources from multiple networks [3].
forms mode selection, allocation of WLAN A converged network may not provide
resources, and allocation of LTE-A network enhanced network performance at areas such as
resources in three different timescales. The cell edges or rural areas where only one network
resource allocation scheme is semi-distributedly is available. Device-to-device (D2D) communi-
implemented in the underlying converged D2D cation can be applied in these areas to improve
communication network, and the achievable per- network performance as it allows direct commu-
formance improvements are demonstrated via nication between source and destination UEs
simulation results. that are in proximity, and by incorporating hop
and reuse gains to the network [5]. Hop gain is a
result of D2D links using either uplink (UL) or
INTRODUCTION downlink (DL) resources only. Reuse gain is
Recent advancements in the mobile industry achieved by simultaneously using the same set of
have dramatically increased the number of smart resources for both traditional (i.e., relaying com-
mobile devices (e.g., smart phones and tablets) munications via base stations) and D2D links.
operating in any geographical region, and the Therefore, to achieve network performance
number of data hungry applications (e.g., video improvements throughout the converged net-
streaming, YouTube and Google Maps) that run work, D2D communication can be enabled in
on these devices. Consequently, the demand for the converged network.
higher data rates with seamless service coverage Enabling D2D communication in a converged
Amila Tharaperiya Gam- and support for various applications’ diverse network provides two more important benefits.
age, Ran Zhang, and quality of service (QoS) requirements is greater First, high-capacity D2D links can be set up
Xuemin (Sherman) Shen than ever before. To satisfy these high service between users who are not within a WLAN cov-
are with the University of demands, it is necessary to efficiently utilize the erage by pairing WLAN radios of the UEs. To
Waterloo. resources available in heterogeneous wireless pair two WLAN radios, control signals and infor-
networks. mation related to authentication are sent via a
Hao Liang is with the Most of the high service demanding areas, cellular network. Although Wi-Fi Direct compat-
University of Alberta. such as office buildings, hotspots, and airports, ible WLAN radios are able to discover neighbor-

98 1536-1284/14/$25.00 © 2014 IEEE IEEE Wireless Communications • December 2014

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 99

ing devices and pair themselves, they require

users to distribute authentication related infor-

k
mation, such as a personal identification num-

in
l
al
ber, via another secure network and manually

on
i ti
enter that information during link setup. There-

ad
UE8

Tr
fore, the D2D link setup process, which takes WLAN AP
advantage of UE multi-homing in a converged
network to send control signals and authentica- UE1 UE3
tion information via a cellular network, provides UE2
users with secure and convenient service. Sec-

k
WLAN AP

lin
ond, the interference issue in the underlaying UE10

D
D2
eNB UE9
networks of D2D communication can be relaxed.
In a converged network, there are large amounts UE5 A multi-homing
user
of resources with different channel conditions.
Therefore, co-channel interference (CCI) UE4
between traditional and D2D links can be UE6
reduced by choosing resources with weaker
interference channels between the links.
There are several technical challenges that
make resource allocation for D2D communica- UE7
tion underlaying converged networks complicat-
ed: Figure 1. Underlaying cellular/WLAN converged D2D communication net-
• Involvement of multiple timescales as different work.
networks use different radio access technolo-
gies [3, 6, 7]
• Selection of users’ communication modes (i.e., the Internet service provider (ISP). Synchroniza-
traditional or D2D) for multiple networks to tion of the APs with the LTE-A network is
maximize hop and reuse gains considering the achieved by using synchronization protocols,
resources available in individual networks such as IEEE 1588-2008 and Network Time Pro-
• Interference management [5] tocol version 4 (NTPv4) over the Ethernet back-
Several works in the literature have studied hauls connected to the APs. Users can access
resource allocation for converged networks [1, 3, the services connecting to one network (e.g.,
4] and D2D communication underlaying wireless UE8) or simultaneously connecting to multiple
networks [5, 8–12] separately. However, how to networks using the UE multi-homing capability
improve network performance for D2D commu- (e.g., UE9). In this system, network assisted (or
nication underlaying converged networks needs operator controlled) D2D communication is
further study. considered. Using traditional mode, users com-
In this article, we first present the technical municate with eNB or an AP (e.g., UE8). Using
challenges for allocating resources in an under- D2D mode, source and destination users in
laying converged D2D communication network, proximity directly communicate with each other
and discuss existing and new solutions. Second, a (e.g., UE 4 and UE 5 ). These D2D links can be
resource allocation scheme is proposed to established using multiple networks in the con-
address these challenges, and the related imple- verged network; for example, a D2D link
mentation issues are investigated. The proposed between UE1 and UE2 can be established over
resource allocation scheme operates on three both the LTE-A network and the WLAN. Fur-
timescales and is designed to capture the fea- thermore, when the D2D users are not within an
tures of diverse radio access technologies of dif- AP’s coverage, a high-capacity D2D link can be
ferent networks within the converged network. set up between the users by pairing the UE
Simulation results demonstrate the throughput WLAN radios with the assistance of the LTE-A
and QoS enhancements that can be achieved by network. To pair the WLAN radios, relevant
employing the proposed scheme. Future research control information and authentication
directions are identified at the end of this article. request/response messages are sent through the
LTE-A network.
D2D COMMUNICATION UNDERLAYING
CONVERGED NETWORK CHALLENGES FOR RESOURCE ALLOCATION
AND RELATED WORK
We focus on an underlaying cellular/WLAN con-
verged D2D communication network consisting In this section, we present three main technical
of an LTE-Advanced (LTE-A) cellular network challenges for allocating resources in an under-
and IEEE 802.11n WLANs. Although there are laying cellular/WLAN converged D2D communi-
other types of networks, such as ZigBee and cation network:
Bluetooth, we focus on LTE-A networks and • Allocation of resources capturing diverse radio
IEEE 802.11n WLANs due to their capability to access technologies
provide high data rates over a wide coverage • Selection of users’ communication modes for
area with QoS support. The system model is multiple networks to maximize hop and reuse
shown in Fig. 1. Enhanced NodeBs (eNBs) (or gains
base station) of the LTE-A network and WLAN • Interference management
access points (APs) are interconnected via the Some related works in the literature are also dis-
LTE-A evolved packet core (EPC) network and cussed.

IEEE Wireless Communications • December 2014 99

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 100

The objective of CHALLENGE 1: ALLOCATION OF of all the channels. Then the resource allocation
problem in the two timescales can be formulated
enabling D2D communi- RESOURCES CAPTURING as a multiple timescale Markov decision process
cation in the converged MULTIPLE RADIO ACCESS TECHNOLOGIES (MMDP) [15]. The optimal resource allocation
decisions can be determined by solving the
network is to provide Different types of networks use different medi- MMDP problem. However, such a method is
um access control layer (MAC) and physical highly complex due to the large number of states
higher data rates with layer (PHY) technologies: available in practical systems [3].
enhanced QoS for users • LTE-A network: orthogonal frequency-division
CHALLENGE 2: EFFICIENT MODE SELECTION
multiple access (OFDMA)-based PHY and
throughout the con- centrally coordinated MAC [6]. The objective of enabling D2D communication
• IEEE 802.11n WLAN: orthogonal frequency- in the converged network is to provide higher
verged network. To division multiplexing (OFDM)-based PHY, data rates with enhanced QoS for users through-
achieve this objective, it and MAC based on hybrid coordination func- out the converged network. To achieve this
tion (HCF), which consists of contention- objective, it is crucial to select the best commu-
is crucial to select the based channel access and contention-free nication modes that take advantage of user prox-
best communication polling-based channel access mechanisms [7] imity, and fully realize hop and reuse gains. In
The existence of multiple PHY and MAC this section, we discuss the two key challenges
modes that take advan- technologies within a converged network poses a for efficient mode selection:
challenge for resource allocation by complicating • High complexity and communication overhead
tage of user proximity, the resource allocation process. The set of possi- • Realization of hop and reuse gains
and fully realize hop ble resource allocation decisions and the achiev- In a converged network, the mode selection
able user throughputs over a network depends process has high complexity as it requires esti-
and reuse gains. on the PHY and MAC technologies of the net- mation of a large number of channels due to the
work [3, 4, 13]. For example, in an OFDMA- availability of a large number of potential D2D
based network, bandwidth allocation decisions and traditional links over multiple networks. It
should be in multiples of a subcarrier bandwidth, also causes a large communication overhead due
while during the contention-based channel access to transmission of a large volume of CSI [5].
of the WLAN, user throughputs should be calcu- Therefore, repeating mode selection on a very
lated considering the transmission collisions that fast timescale (e.g., at every resource allocation
occur due to the MAC scheme. Therefore, effi- interval in LTE-A networks with a duration of 1
cient resource allocation schemes should be ms [6]) to determine the best communication
designed based on the underlying PHY and modes based on instantaneous channel condi-
MAC technologies in order to make feasible tions is not practical. To reduce the complexity
resource allocation decisions while properly esti- and overhead, mode selection can be performed
mating the throughputs that users will achieve. on a slower timescale based on the channel
The challenge in a converged network is that statistics. However, the timescale should not be
there are diverse PHY and MAC technologies in too slow as D2D links may become very weak
the different types of networks to be considered. over time due to user mobility. This issue can be
In [1, 3, 4], resources of converged networks are relaxed in the converged network by forming
allocated by estimating the throughputs via D2D links for high-mobility users via the cellular
OFDMA-based networks using the Shannon network only while forming D2D links for low-
capacity formula and via WLANs by calculating mobility users via the cellular network and
the average user throughputs considering the WLANs, as cellular network coverage is much
effect of collisions. wider than WLAN coverage.
The resource allocation interval (i.e., interval Realization of hop gain is a challenge as the
between two successive resource allocations) of user modes that provide the highest throughputs
cellular networks is much shorter than that of should be selected, while calculation of D2D
WLANs as cellular networks and WLANs are mode throughput is complicated. Throughput of
designed to support high and low mobility users D2D mode is the sum of the D2D link through-
with speeds upto 350 kmh–1 and 3 kmh–1, respec- put and the additional throughput that can be
tively [6, 7]. The resource allocation intervals are achieved utilizing the saved resources. When
calculated based on the channel coherence times. D2D mode is used, resources are saved as D2D
Therefore, resources for these two networks are links use either UL or DL resources only. The
allocated in a fast and a slow timescale, respec- additional throughputs will be in the DL if the
tively [3]. Existence of multiple timescales in a UL resources are used for the D2D links, and
converged network poses a challenge for vice versa. In a time-division duplexing (TDD)
resource allocation as prediction of the through- system, the throughput of D2D mode can be cal-
puts that will be achieved over future time slots culated by allocating all the available resources
is required to optimally allocate the resources. for the D2D link, while that of traditional mode
That is, when resources of the converged net- can be calculated by allocating a part of avail-
work are allocated at the beginning of a slow able resources for the UL and the remaining
timescale time slot, it is required to predict the resources for the DL [8]. This method provides
throughputs that will be achieved over the future accurate results as the UL and DL share the
fast timescale time slots that lie within the cur- same set of resources in a TDD system. Howev-
rent slow timescale time slot. One approach to er, in frequency-division duplexing (FDD) sys-
tackle this challenge is as follows. First, the wire- tems, joint allocation of UL and DL resources is
less channels can be modeled as finite-state necessary as the UL and DL use two dedicated
Markov channels [14], where the state space cor- sets of resources on two different carrier fre-
responds to the channel state information (CSI) quencies.

100 IEEE Wireless Communications • December 2014

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 101

Selecting modes to realize the highest reuse D2D link should be determined based on the An orthogonal resource
gain in a network is challenging due to two main achievable throughputs with each resource type,
reasons: considering CCI. An orthogonal resource is uti- is utilized by only one
• Calculation of optimal power levels for D2D lized by only one link, whereas a non-orthogonal
and traditional links over a shared resource is resource is shared/reused by a D2D and a tradi-
link, whereas a non-
complicated due to the existence of CCI tional link. When a D2D link is far away from a orthogonal resource is
between the links. traditional link and the two D2D communicating
• Finding the optimal pair of D2D and tradi- users are in proximity, allocating non-orthogonal shared/reused by a
tional links to share a particular resource is resources for these two links is beneficial due to D2D and a traditional
tedious as there are a large number of differ- limited CCI between the links [8, 10]. In this sce-
ent link pairs to be considered for each nario, the total achievable throughput via the two link. When a D2D link is
resource. links reusing an RB of the LTE-A network is
In [8–10], power allocation to capture reuse gain shown in Fig. 2a, where P t and P d are transmit
far away from a tradi-
is studied assuming that the number of available power levels of traditional and D2D link trans- tional link and the two
resources equals the number of traditionally mitters. However, as shown in Figs. 2b and 2c,
communicating users. Furthermore, it is assumed when the two links are at close range, a higher D2D communicating
that each traditionally communicating user occu- total throughput can be achieved by allocating users are in proximity,
pies only one resource. When one D2D link orthogonal resources for the links; the total
reuses all the resources, the power allocation throughput reaches the highest when the link allocating non-orthogo-
which maximizes the D2D link throughput is with higher channel gain uses the RB.
found in [9]. When there are multiple D2D links Use of UL and DL resources for D2D links
nal resources for these
and each D2D link reuses only one resource, the affects the interference management and system two links is beneficial
optimum power allocation to maximize the total complexity differently. When DL resources are
throughput over a resource is found in [8, 10]. reused for D2D links, CCI is received by the due to limited CCI
Moreover, in [10], by evaluating all the possible users who are communicating traditionally. To between the links.
D2D and traditional link pairs for each resource, calculate power levels of D2D link transmitters
the optimum pairing to maximize the reuse gain ensuring tolerable CCI at traditionally communi-
in the network is found by using a weighted cating users, it is required to estimate the chan-
bipartite matching algorithm. However, in large nels between D2D link transmitters and
multicarrier systems (e.g., LTE-A networks), traditionally communicating users. Furthermore,
there are a large number of subcarriers or physi- CCI could be severe if a D2D pair and a tradi-
cal resource blocks (RBs) compared to the num- tionally communicating user are located at a cell
ber of users. Furthermore, the set of resources edge or at nearby cell edges [5]. On the other
allocated to one traditional link could be reused hand, when UL resources are reused for D2D
by several D2D links, where each D2D link links, CCI is received by the eNB and APs.
reuses a subset of the resources allocated to the Therefore, to manage CCI, already available CSI
traditional link, and vice versa. In this setting, of the channels between users and eNB/APs can
allocation of RBs in a cellular network based on be utilized. In addition to CCI, LTE-A network
a reverse iterative combinatorial auction-based users will suffer from ICI when DL resources
approach is investigated in [11, 12]. are utilized for D2D links, because the signals
from eNB and D2D transmitters arrive at the
CHALLENGE 3: INTERFERENCE MANAGEMENT users at different time instances. However, if UL
Intercarrier interference (ICI) and CCI caused resources are utilized for D2D links, the eNB
by D2D communication degrades the through- will suffer from ICI. As ICI is an inherent issue
put performance of the underlaying converged in conventional OFDMA-based UL systems,
D2D communication network. ICI occurs in these systems are equipped with ICI cancellation
multicarrier systems, such as an LTE-A network, schemes to combat ICI at the eNB, but not at
when the signals over different subcarriers arrive the users. Therefore, use of UL resources for
at a receiver with different delays. Therefore, to D2D links simplifies CCI and ICI management.
maximize throughput performance, it is essential Furthermore, it is beneficial to utilize UL
to manage interference. However, interference resources for D2D links as UL resources are less
management is a challenge as it complicates the utilized than DL resources due to asymmetric
resource allocation process by requiring three UL and DL traffic loads [9, 10].
additional resource allocation decisions to be Converging networks simplifies the interfer-
made: ence management in several ways. First, CCI
• Whether to allocate orthogonal or non-orthog- between a D2D and a traditional link pair varies
onal resources for D2D links with the reused resource as the channel condi-
• Whether to utilize UL or DL resources for tions over different resources vary. Moreover,
D2D links there are a large amount of resources available
• Determining which D2D and traditional link from multiple networks. Therefore, CCI in a
pair to reuse a resource and the transmit converged network can be reduced by selecting
power levels of the link pair (discussed in the resources for D2D and traditional link pairs such
previous subsection) that CCI is minimized. Second, when the D2D
In addition, the characteristics of converged net- links are set up over multiple networks, the
works, such as low transmit power levels of the transmit power of the D2D link transmitters is
multi-homing users over individual networks, divided among multiple network interfaces,
should be considered in the resource allocation reducing CCI. Third, convergence of a cellular
process to ease interference management. network and WLANs enables the use of WLAN-
To attain high system throughput, selection of based D2D links. As there are several WLAN
orthogonal or non-orthogonal resources for each frequency channels which can be utilized for

IEEE Wireless Communications • December 2014 101

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 102

these links, multiple D2D links can be setup and

simultaneously operated among the users in
x 105 vicinity without causing CCI. For example, IEEE
802.11n supports three non-overlapping channels
10 in 2.4GHz frequency band [7].
Total throughput (b/s)

9
PROPOSED THREE-TIMESCALE
8
RESOURCE ALLOCATION SCHEME
7
In this section, we propose a novel resource allo-
6 cation scheme for underlaying cellular/WLAN
converged D2D communication network, shown
5 in Fig. 1, overcoming the various challenges
10
mentioned earlier. The proposed scheme is
10
8 designed with two objectives:
5 6
4
• Maximize the total system throughput subject
2 to user QoS and total power constraints
Pd (W) 0 0
Pt (W) • Minimize the signaling overhead and the com-
(a)
putational complexity such that this scheme
can be employed in practical systems
The total system throughput is the sum of all the
D2D and traditional link throughputs achieved
over both networks. The QoS constraints ensure
x 105
that the total throughput achieved by a user via
4.7 the networks satisfies the user’s minimum through-
put requirement. The total power constraints
Total throughput (b/s)

4.6
ensure that the sum of transmit power allocated to
4.5
the two network interfaces of a UE does not
4.4 exceed the total power available at the UE.
4.3 As shown in Fig. 3, the proposed resource
4.2
allocation scheme operates on three different
timescales. The first timescale is the slowest,
4.1 while the third timescale is the fastest (i.e., a time
4 slot in the first timescale is the longest, while that
10 in the third timescale is the shortest). Mode selec-
10 tion is performed in the first timescale. Resources
8
5 6 of the cellular network and the WLANs are joint-
4
2 ly allocated in the second timescale. As the cellu-
Pd (W) 0 0
Pt (W) lar network has a short resource allocation
interval, resources of the cellular network are
(b) reallocated in the third timescale.
The proposed scheme addresses the various
challenges stated earlier as follows.
• To address Challenge 1: A low-complexity joint
x 105
resource allocation for the cellular network
4.5 and WLANs, which operate on the third and
second timescales, respectively, is performed
Total throughput (b/s)

4
based on the average channel gains; and effi-
3.5
cient and feasible resource allocation deci-
3 sions are made by considering the PHY and
2.5 MAC features of the networks.
2
• To address Challenge 2: Complexity and sig-
naling overhead are reduced by performing
1.5 mode selection in the first (i.e., a slow)
1 timescale; hop gain is captured in the mode
10 selection by utilizing two resources, which can
10 be allocated to a D2D link or a traditional
8
5 6 link with a UL and a DL, to calculate the
4
2 throughput of each mode; and allocation of
Pd (W) 0 0
Pt (W) non-orthogonal resources is simplified by allo-
cating resources in two steps.
(c) • To address Challenge 3: Non-interfering
WLAN-based D2D links are utilized; CCI and
Figure 2. Throughputs achieved reusing a RB for a D2D and a traditional ICI mitigation is simplified by using UL
link: a) D2D and traditional links are far away from each other; b) two resources for the D2D communication within
links are in proximity, and the traditional link has a higher channel gain; c) the cellular network; severe CCI is avoided by
two links are in proximity, and the D2D link has a higher channel gain. preventing the allocation of non-orthogonal
resources for the links in proximity; and CCI
is further reduced by enabling UE multi-hom-

102 IEEE Wireless Communications • December 2014

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 103

ing for both D2D and traditional modes, and

Repeat mode
properly calculating the UE transmit power. selection A time slot of the first timescale
To make efficient and feasible resource allo-
A time slot of
cation decisions, user throughputs over each net- the second timescale
work are accurately calculated based on the . . .
PHY and MAC technologies of the network.
Throughputs over RBs of the cellular network
and time slots of contention-free polling-based (a)
channel access of WLANs are calculated using Allocate resources Allocate resources
the Shannon capacity formula, since the alloca- of the WLAN A time slot of the second timescale of the WLAN
tions of these resources are centrally controlled
by eNB and APs, respectively. Average through- Contention-free Contention-based
channel access period channel access period
put achieved by the ith user via contention-based
channel access of WLAN is calculated taking the .... RTS CTS Data ACK ....
collisions in the channel into account. For four- Beacon Contention-free
way handshaking, it is given by [3, 4, 7] transmission time slots Request- Clear-to-send Acknowledgment
to-send message
Allocate resources of message
the cellular network
βD A time slot of the third timescale
RiCB = N
t0 + N β ∑
D . . .
(1)
2 + H j Pj )
w
j =1 B log (1
(b)

where P wj is the transmit power level of the jth Figure 3. Proposed three-timescale resource allocation scheme: a) opera-
user over the WLAN, N is the number of users tions of the first and second timescales; b) detailed view of a second
in the WLAN, H j is the square of the channel timescale time slot.
gain divided by the noise power, D is the packet
size, B is the WLAN bandwidth, b is the proba-
bility of a successful transmission, and t0 is a sys- one RB for UL and the other for DL; in D2D
tem-specific constant that varies with N [3]. mode, throughput is calculated allocating both
Throughput function RCB i is a concave increasing RBs for the D2D link to capture the hop gain.
function of Pwj, j. Furthermore, throughputs are calculated using
Since cellular networks and WLANs are average channel gains and unit transmit power
designed to support high- and low-mobility levels for two reasons:
users, respectively, cellular network resources • Instantaneous channel gains and user transmit
are allocated for both high- and low-mobility power levels vary over the time slots of the
users, while WLAN resources are allocated for third timescale.
low-mobility users only. Furthermore, in the cel- • Within each time slot of the first timescale,
lular network, D2D links are allocated UL there are multiple time slots of the third
resources as UL resources are underutilized, and timescale.
in order to manage CCI and ICI without signifi- In WLANs, mode selection is performed in a
cantly increasing the system complexity. similar manner, but using two time slots. Users
use the same mode for contention-based and
FIRST TIMESCALE: MODE SELECTION contention-free channel access mechanisms.
Mode selection is performed in the first When a D2D communicating user needs to
timescale in order to reduce the involved com- access non-D2D services, such as the Internet,
putational complexity and signaling overhead by email, and voice mail, the user is also allocated
less frequently (i.e., in the first timescale) mak- direct (i.e., traditional) links to the eNB or an
ing the mode selection decisions and estimating AP. However, if traditional mode has been
the channels required for mode selection. In the selected for the user for D2D communication
mode selection process, users are allowed to use over a particular network, allocation of such a
different modes for different networks as the direct link is not required since both D2D and
wireless channel gains over one network differ non-D2D services can share the link between the
from those over another network. user and the eNB or AP. This is performed via
The first step of the mode selection process is steps 3,4, and 5 of the mode selection algorithm.
to allocate WLAN-based D2D links for users Furthermore, step 4 also ensures the allocation
who are outside AP coverage areas, as these of a direct link between a user and the eNB
links provide high capacity without causing CCI. when the user uses only a WLAN-based D2D
Due to the high capacity of these links, when a link for D2D communication and needs to access
D2D user pair is allocated one of these links, non-D2D services.
they are not allocated cellular network resources
for D2D communication. On the other hand, SECOND TIMESCALE: JOINT RESOURCE ALLOCATION
when a user is within AP coverage, the user can
access both networks using the multi-homing FOR THE CELLULAR NETWORK AND WLANS
capability. This is performed via steps 1 and 2 of The second timescale resource allocation jointly
the mode selection algorithm shown in Fig. 4a. allocates cellular network and WLAN resources,
In the cellular network, the mode for each distributes power available at multi-homing UEs
user is selected based on the achievable through- between their two network interfaces, and
put using each mode, utilizing two RBs. In tradi- ensures QoS satisfaction. The resource alloca-
tional mode, throughput is calculated allocating tion is executed in two steps in order to simplify

IEEE Wireless Communications • December 2014 103

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 104

Resources are allocated the allocation of non-orthogonal resources, and the same reason, average channel gains are used
the calculation of transmission power levels for for mode selection (as discussed earlier). Using
using instantaneous D2D and traditional links that share these average channel gains in the joint resource allo-
resources. In the first step, resources are allocat- cation problem, the amounts of power that
channel gains over the ed for the traditional and D2D links that use should be allocated for UE cellular network
WLANs while using aver- orthogonal resources. In the second step, interfaces and the throughputs that should be
remaining D2D links are allocated non-orthogo- achieved through the cellular network are deter-
age channel gains over nal resources. mined. In the third timescale, resources of the
the cellular network, as Resources are allocated using instantaneous cellular network are allocated based on these
channel gains over the WLANs while using aver- power and throughput levels, and utilizing
there are multiple third age channel gains over the cellular network, as instantaneous channel gains.
there are multiple third-timescale time slots As reusing resources for D2D and traditional
timescale time slots within one time slot of the second timescale. For links that are in proximity is inefficient, D2D
within one time slot of
the second timescale.
Start Executed at CCS
For the same reason,
average channel gains No
Step 3
are used for mode End
Yes Determined mode
for each D2D pair User i (or j)
needs to access
selection. (users i and j)
non-D2D services
No
Yes
Step 1 Step 4
Yes Determine No Allocate ith user
modes over Traditional mode
Users i and j are in an is selected for cellular a direct link to
AP coverage cellular network cellular network
and over WLAN network

No Yes
Allocate a Step 5
Step 2 WLAN based
Yes D2D link
Users i and j are in ith user is in an AP Yes Allocate ith user
UE WLAN radio coverage and D2D mode a direct link to
communication is selected for WLAN WLAN
range Determine
mode over
cellular network No
No

(a)

Executed at CCS Executed at eNB

Step 8
Start
Wait until CCS receive Yes Allocated resources for
calculated Rci , Rwi , pci each user-i, except D2D
and pwi for all i mode users at least L
distance from eNB

Step 6
No
Yes
Allocated resources End
optimally Step 9 For ith user, allocate RBs and
determine pci , based on dual
No Allocated resources variables
Yes
for each WLAN based
D2D link user-i
Update and distribute For ith user, allocate RBs
dual variables
and determine pci and pwi ,
No based on dual variables
Step 10
ith user access Yes
Executed at APs cellular network via a Allocate total power of ith
direct link user for the WLAN based
Step 7 D2D link
Yes No
Allocated resources
for each user-i in an For ith user, allocate RBs
AP coverage and determine pci such that
Step 11
CCI at eNB is less than Ic ,
No Allocated resources based on dual variables
for each D2D mode user-i No
For ith user, allocate contention- at least L distance from
free time slots and determine eNB Determine average CCI
pwi , based on dual variables received from traditional
mode users over RBs
Yes

(b)

Figure 4. a) Mode selection algorithm; b) second timescale joint resource allocation algorithm.

104 IEEE Wireless Communications • December 2014

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 105

links in WLANs are allocated orthogonal • Assisted global navigation satellite systems (A- As cellular networks
resources. Similarly, in the cellular network, GNSS) positioning
D2D links that lie within distance of L from the • Observed time difference of arrival (OTDOA) have a shorter resource
eNB are allocated orthogonal resources. Fur- positioning
thermore, from Eq. 1, it can be seen that the Position information can be exchanged between
allocation interval than
transmit power levels of all the users in a WLAN UEs and an eNB via LTE positioning protocol WLANs, resources of the
are correlated, because a user’s throughput via (LPP).
contention-based channel access depends on the cellular network are real-
transmit power levels of all the users in the THIRD TIMESCALE: located in the third
WLAN. Also, the users in a WLAN may access
the cellular network using a part of the power CELLULAR NETWORK RESOURCE ALLOCATION timescale utilizing
available in the UEs. Therefore, all the multi- As cellular networks have a shorter resource
homing users who access both networks should allocation interval than WLANs, resources of
instantaneous channel
be jointly allocated resources during the first the cellular network are reallocated in the third gains. Furthermore, by
step. To facilitate that, D2D links in the cellular timescale utilizing instantaneous channel gains.
network, which are among these multi-homing Furthermore, by using a fast timescale, multiuser using a fast timescale,
users, are allocated orthogonal resources. diversity over fast fading wireless channels is multiuser diversity over
Remaining D2D links in the cellular network are exploited. In this timescale, resources are allo-
allocated non-orthogonal resources, realizing the cated following the same two-step process as in fast fading wireless
reuse gain in the system. the second timescale. In the first step, the ith
In the first step, resources of the cellular net- multi-homing user has total power of Pci to com-
channels is exploited.
work and the WLANs are jointly allocated sub- municate over the cellular network, and requires
ject to two main constraints: a minimum rate of Rmin – Rwi via the cellular net-
• P ci + P wi  P T , where P ci is the transmit power work. P ci and R wi are calculated in the second
level of the ith user over the cellular network timescale. The second step remains unchanged.
and PT is the total available power
• R ci + R wi  R min , where R ci and R wi are the
throughputs achieved by the ith user via the IMPLEMENTATION AND
cellular network and the two channel access PERFORMANCE EVALUATION
mechanisms of a WLAN, respectively, and
Rmin is the required throughput In this section, we discuss the semi-distributed
These are the total power and QoS constraints implementation of the proposed resource alloca-
for multi-homing users. If a D2D communicating tion scheme and evaluate its performance. The
user accesses non-D2D services, respective QoS system consists of an LTE-A network and IEEE
constraints are considered for D2D and non- 802.11n WLANs operating in 2.1 and 2.4 GHz
D2D services. In this first step, the eNB is frequency bands, respectively.
assumed to receive the worst CCI of I c from The semi-distributed implementation shown
D2D links, and the power available at multi- in Fig. 5 reduces the signaling overhead and sig-
homing users is allocated to the two network naling delay, distributes the computational bur-
interfaces. In the second timescale the joint den over the networks, and prevents a single
resource allocation algorithm shown in Fig. 4b, point of failure. Different functions of the
resource allocation for multi-homing users dur- resource allocation scheme are performed at
ing the first step is performed via steps 7 and 8, APs, an eNB, and a centralized control server
while that for WLAN-based D2D link users is (CCS) that is connected to the LTE-A EPC
performed via steps 9 and 10. It should be noted through a packet data network gateway (PDN-
that these WLAN-based D2D link users are GW). APs and the CCS communicate through a
coordinated by the eNB as they are not within WLAN access gateway (WAG), evolved packet
an AP’s coverage. Calculation of the dual vari- data gateway (ePDG), and the PDN-GW. The
ables is discussed later. eNB and CCS communicate through a serving
In the second step, cellular network resources gateway (S-GW) and the PDN-GW.
are allocated (reused) for the D2D links that use Mode selection is performed at the CCS. To
non-orthogonal resources subject to the total determine the user modes, average channel gains
power and QoS constraints. CCI received by the of the traditional and potential D2D links over
D2D links is taken into account, and transmit both networks are sent to the CCS. Once the user
power levels of the D2D link transmitters are modes are determined, the selected modes are
calculated such that they do not exceed Ic at the informed to the APs and eNB to set up the links.
eNB. CCI received by a D2D link receiver can The first step of the second timescale
be calculated as power and RB allocations for resource allocation is to jointly allocate cellular
the traditional links are already completed in the network and WLAN resources. Resource alloca-
first step. The second step is performed via step tion for each network is performed at its base
11 of the second timescale joint resource alloca- station (i.e., eNB or AP) and controlled by the
tion algorithm. CCS such that resource allocation for the entire
In order to reduce the required number of system can iteratively converge to the global
channel estimations for the second step, average optimum. Specifically, the CCS broadcasts the
channel gains, which can be estimated based on dual variables (i.e., Lagrange multipliers) that
the distances, are used for the calculation of correspond to the total power and QoS con-
received/caused CCI. To determine the dis- straints. Then the APs and eNB allocate
tances, positions of the UEs can be calculated resources based on the received dual variables,
using two techniques that are supported by LTE and feedback P ci, P wi, R ci and R wi, i to the CCS.
networks: Finally, CCS updates the dual variables and

IEEE Wireless Communications • December 2014 105

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 106

tions per user reduces with D, as more users are

Data traffic flow allocated WLAN-based D2D links and more
Authentication D2D links are allocated non-orthogonal cellular
signaling
network resources during the second step.
Algorithm-related PDN-GW
signaling flow Next, we compare the throughput and SI per-
Centralized control formance of the proposed resource allocation
server (CCS)
scheme with that of a cellular/WLAN converged
Evolved network and a conventional system. In the cellu-
packet core
Internet ePDG
(EPC) lar/WLAN converged network, resources are
HSS allocated based on the second and third
WLAN direct
3GPP AAA timescale operations. In the conventional system,
server resource allocation for each network is per-
IP access S-GW
WAG
formed individually.
Internet service
WLAN 3GPP
According to the throughput and SI perfor-
provider (ISP) mance shown in Figs. 6a and 6b, the
IP access
cellular/WLAN converged network provides high-
er performance than the conventional system. The
proposed scheme provides further performance
link enhancements, and its performance increases with
D2D UE
WLAN AP UE D. When D increases, more D2D links can be
eNB UE established, because the number of potential D2D
WLAN AP
UE users in the system increases with D. As a result,
the performance of the proposed scheme increas-
es with D. When D = 40 percent, the proposed
Figure 5. Semi-distributed implementation of the proposed resource alloca- scheme improves throughput by 3.4 and 10 times
tion scheme. compared to the throughputs achieved in the cel-
lular/WLAN converged network and the conven-
tional system, respectively. The reasons for such
broadcast them back. As shown in step 6 of Fig. enhanced performance are that joint allocation of
4b, this process continues until the resource allo- resources in multiple networks, exploitation of
cation reaches global optimality. better wireless channels available between the
The second step of the second timescale users in proximity, realization of hop and reuse
resource allocation and the third timescale gains, utilization of WLAN-based D2D links, and
resource allocation are performed at the eNB as efficient use of orthogonal and non-orthogonal
both of them allocate cellular network resources resources to manage interference. This perfor-
only. Furthermore, such implementation pro- mance comparison demonstrates the throughput
vides a low signaling delay, which is essential for and QoS improvements that can be achieved by
third timescale operations due to very short time converging multiple networks and enabling D2D
slot duration. communication within a converged network.
For the performance evaluation, we consider
25 high-mobility and 25 low-mobility users in the CONCLUSIONS
system. All the users are capable of multi-hom-
ing, and D percent of them can communicate In this article, we have studied resource alloca-
using D2D mode. The total power available at tion for an underlaying converged D2D commu-
each user is 27 dBm. Durations of a time slot in nications network that consists of an LTE-A
the first, second, and third timescales are 640, network and IEEE 802.11n WLANs. A resource
64, and 1 ms, respectively. Rayleigh fading wire- allocation scheme has been proposed to maxi-
less channels with path loss exponent of 4 are mize the throughput of the system subject to
used. We set L = 200 m and I c = –62 dBm in QoS satisfaction. The proposed scheme has been
the LTE-A network. QoS satisfaction is quanti- designed based on the diverse PHY and MAC
fied by using the satisfaction index (SI), which is technologies of different networks, and to man-
defined as SI = E {1RRmin + 1Rmin>R ◊ R/Rmin}, age interference and reduce the high complexity
where R is the achieved user throughput, and and signaling overhead caused by the mode
1ab = 1 if a  b or 1ab = 0 otherwise. selection process. To further reduce the signal-
Since the proposed scheme performs mode ing overhead and delay while preventing a single
selection on a slower timescale, the average point of failure, the proposed scheme has been
number of channel estimations and the signaling implemented in a semi-distributed manner. Sim-
overhead are reduced by 8.3, 15.9, and 29.1 per- ulation results have demonstrated that the pro-
cent for D = 10, D = 20, and D = 40 percent, posed scheme significantly improves the system
respectively. In addition, by executing second throughput and QoS satisfaction.
and third timescale resource allocations at APs This work can be extended by further consid-
and the eNB, the signaling overhead is reduced ering the UE energy efficiency and the backhaul
by another 58.4 percent as a large volume of CSI capacities. Maintaining a high UE energy effi-
is not sent to the EPC network. ciency is important to provide UEs with a longer
The first step of second timescale resource operating time. In a converged network consist-
allocation has the highest complexity as it jointly ing of heterogeneous wireless networks, consid-
allocates cellular network and WLAN resources. eration of the backhaul capacities is crucial as
It converges within 7.93, 7.62, and 6.91 iterations the capacities of some networks could be bottle-
per user for D = 10, D = 20, and D = 40 per- necked by the backhaul capacities (e.g., WLANs
cent, respectively. The required number of itera- and femtocells).

106 IEEE Wireless Communications • December 2014

GAMAGE_LAYOUT_Author Layout 12/18/14 4:17 PM Page 107

1000
Proposed scheme (D = 40%) 1
900 Proposed scheme (D = 20%)
Proposed scheme (D = 10%)
800 Cellular/WLAN converged network 0.9
Average user throughput (Mb/s)

Conventional system
700
0.8

Satisfaction index
600

500 0.7

400 0.6
300
0.5
200
Proposed scheme (D = 40%)
100 0.4 Proposed scheme (D = 20%)
Proposed scheme (D = 10%)
50 Cellular/WLAN converged network
0.3 Conventional system
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
Required user throughput (Mb/s) Required user throughput (Mb/s)
(a) (b)

Figure 6. Performance comparison: a) throughput performance; b) QoS satisfaction.

REFERENCES BIOGRAPHIES
[1] X. Pei et al., ”Radio-Resource Management and Access- AMILA THARAPERIYA GAMAGE [S’07] (amila.gamage@uwater-
Control Mechanism Based on a Novel Economic Model loo.ca) received his B.E. degree in electronics and telecom-
in Heterogeneous Wireless Networks,” IEEE Trans. munications engineering from Multimedia University,
Vehic. Tech., vol. 59, no. 6, July 2010, pp. 3047–56. Malaysia, in 2008 and his M.E. degree in telecommunica-
[2] R. Ferrus, O. Sallent, and R. Agusti, ”Interworking in tions engineering from the Asian Institute of Technology,
Heterogeneous Wireless Networks: Comprehensive Thailand, in 2011. He is currently working toward his Ph.D.
Framework and Future Trends,” IEEE Wireless Com- degree in the Department of Electrical and Computer Engi-
mun., vol. 17, no. 2, Apr. 2010, pp. 22–31. neering, University of Waterloo, Canada. From 2008 to
[3] A. T. Gamage, H. Liang, and X. Shen, “Two Time-Scale 2009, he was a solutions architect with Dialog Telekom
Cross-Layer Scheduling For Cellular/WLAN Interwork- PLC, Sri Lanka. His current research interests include
ing,” IEEE Trans. Commun., vol. 62, no. 8, Aug. 2014, resource management for interworking heterogeneous net-
pp. 2773–89. works, cooperative communication, and cloud computing.
[4] A. T. Gamage and X. Shen, ”Uplink Resource Allocation He is a corecipient of Best Paper Awards at IEEE ICC 2014.
for Interworking of WLAN and OFDMA-Based Femtocell
Systems,” Proc. IEEE ICC ’13, Budapest, Hungary, June HAO LIANG [S’09, M’14] (hao2@ualberta.ca) received his Ph.D.
2013, pp. 4664–68. degree in electrical and computer engineering from the Uni-
[5] G. Fodor et al., ”Design Aspects of Network Assisted versity of Waterloo in 2013. From 2013 to 2014, he was a
Device-to-Device Communications,” IEEE Commun. postdoctoral research fellow at the Broadband Communica-
Mag., vol. 50, Mar. 2012, pp. 170–77. tions Research and Electricity Market Simulation and Opti-
[6] 3GPP TS 36.300, “LTE; Evolved Universal Terrestrial mization Labs at the University of Waterloo. Since 2014, he
Radio Access (E-UTRA) and Evolved Universal Terrestrial has been an assistant professor in the Department of Electri-
Radio Access Network (E-UTRAN); Stage 2,” Rep. cal and Computer Engineering at the University of Alberta,
V11.6.0, 2013. Canada. His research interests are in the areas of smart grid,
[7] IEEE P802.11-REVmb/D12, Part 11, “Wireless LAN Medi- wireless communications, and wireless networking.
um Access Control (MAC) and Physical Layer (PHY)
Specifications, Mar. 2012. RAN ZHANG [S’13] (r62zhang@uwaterloo.ca) received his B.E.
[8] C.-H. Yu et al., ”Resource Sharing Optimization for degree in electronics engineering from Tsinghua University,
Device-to-Device Communication Underlaying Cellular China, in 2010. He is currently pursuing his Ph.D. degree
Networks,” IEEE Trans. Wireless Commun., vol. 10, no. with the Broadband Communication Research Group, Uni-
8, Aug. 2011, pp. 2752–63. versity of Waterloo. His current research interests include
[9] J. Wang et al., ”Resource Sharing of Underlaying Device-to- resource management in heterogeneous networks, carrier
Device and Uplink Cellular Communications,” IEEE Com- aggregation in LTE-A systems, wireless green networks, and
mun. Lett., vol. 17, no. 6, June 2013, pp. 1148–51. electrical vehicle charging control in smart grids.
[10] D. Feng et al., ”Device-to-Device Communications
Underlaying Cellular Networks,” IEEE Trans. Commun., XUEMIN (SHERMAN) SHEN [M’97, SM’02, F’09] (sshen@uwa-
vol. 61, no. 8, Aug. 2013, pp. 3541–51. terloo.ca) is a professor and University Research Chair,
[11] C. Xu et al., ”Efficiency Resource Allocation for Device- Department of Electrical and Computer Engineering, Uni-
to-Device Underlay Communication Systems: A Reverse versity of Waterloo. He was Associate Chair for Graduate
Iterative Combinatorial Auction Based Approach,” IEEE Studies from 2004 to 2008. His research focuses on
JSAC, vol. 31, no. 9, pp. 348-358, Sep. 2013. resource management in interconnected wireless/wired net-
[12] C. Xu et al., Resource Management for Device-to- works, wireless network security, social networks, smart
Device Underlay Communication, Springer, 2014. grid, and vehicular ad hoc and sensor networks. He served
[13] M. J. Neely, E. Modiano, and C.-P. Li, ”Fairness and as Technical Program Committee Chair/Co-Chair for IEEE
Optimal Stochastic Control for Heterogeneous Net- INFOCOM ’14, IEEE VTC ’10 Fall, Symposia Chair for IEEE
works,” IEEE/ACM Trans. Net., vol. 16, no. 2 , Apr. ICC ’10, Tutorial Chair for IEEE VTC ’11-Spring and IEEE ICC
2008, pp. 396–409. ’08, and Technical Program Committee Chair for IEEE
[14] H. S. Wang and N. Moayeri, ”Finite-State Markov GLOBECOM ’07. He also serves/served as Editor-in-Chief for
Channel — A Useful Model for Radio Communication IEEE Network, Peer-to-Peer Networking and Application,
Channels,” IEEE Trans. Vehic. Tech., vol. 44, no. 1, Feb. and IET Communications. He is a registered Professional
1995, pp. 163–71. Engineer of Ontario, Canada, an Engineering Institute of
[15] H. S. Chang et al., ”Multitime Scale Markov Decision Canada Fellow, a Canadian Academy of Engineering Fel-
Processes,” IEEE Trans. Autom. Control, vol. 48, no. 6, low, and a Distinguished Lecturer of the IEEE Vehicular
June 2003, pp. 976–87. Technology and Communications Societies.

IEEE Wireless Communications • December 2014 107