Delay-Sensitive Content Distribution via Peer-to-Peer Collaboration in ...

Delay-Sensitive Content Distribution via Peer-to-Peer Collaboration in Public Safety Vehicular Ad-Hoc Networks Rachad Atat1 , Elias Yaacoub2 , Mohamed-Slim Alouini1 , Fethi Filali2 , Adnan Abu-Dayya2 1

2

Department of Physical Sciences and Engineering, 4700 King Abdullah University of Science and Technology, Thuwal 23955-6900 Kingdom of Saudi Arabia. E-mail:{rachad.atat, slim.alouini}@kaust.edu.sa.

Qatar Mobility Innovations Center (QMIC), Qatar Science and Technology Park, P.O. Box 210531, Doha, Qatar. E-mail:{eliasy, filali, adnan}@qmic.com.

Abstract Delay-sensitive content distribution with peer-to-peer (P2P) cooperation in public safety vehicular networks is investigated. Two cooperative schemes are presented and analyzed. The first scheme is based on unicasting from the base station, whereas the second is based on threshold based multicasting. Long Term Evolution (LTE) is used for long range (LR) communications with the base station (BS) and IEEE 802.11p is considered for inter-vehicle collaboration on the short range (SR). The first scheme is shown to outperform non-cooperative unicasting and multicasting, while the second scheme outperforms non-cooperative unicasting beyond a specific number of cooperating vehicles, when the appropriate 802.11p power class is used. The first scheme achieves the best performance among the compared methods, and a practical approximation of that scheme is shown to be close to optimal performance. Keywords: Vehicular ad-hoc networks, content distribution, cooperative vehicular communications, LTE, 802.11p, scheduling 1. Introduction Communication plays a critical role in disaster prevention and recovery. In critical situations such as earthquake, volcano eruption, terrorist attacks, Preprint submitted to Ad Hoc Networks

June 12, 2013

and hurricanes, data transfer from a central command center, to rescuing mobile terminals such as ambulances, mobile medical treatment units, and police vehicles, needs to be done very fast and in the shortest possible time. Such critical information could include disaster related information such as electronic maps to support aid forces during their motion within a disaster area, weather conditions, nature and specifics of the disaster, safety areas, etc [1]. An example is considered in [2], where videos and information from an incident location are transmitted by a helicopter to a command center, which distributes this information to the public safety teams (police, fire, etc) heading towards, or already located at the incident location. In this paper, we propose cooperative techniques for short range (SR) collaboration to complement long range (LR) wireless transmission for public safety systems. The purpose is to ensure the transmission of the data in the shortest possible time. Clustering of moving vehicles is considered, and two different cooperative schemes, to be implemented in each cluster, are presented. The first scheme, which we will refer to as Scheme 1, consists of the base station (BS) unicasting the data at each fading realization to a single vehicle on the LR link, which in turn multicasts it to its peers on the SR links. The second scheme, which we will refer to as Scheme 2, consists of the BS multicasting the data using a pre-determined transmission rate. Vehicles with the best channel conditions and with successful reception on the LR would receive the data and multicast the data on the SR links to vehicles having achievable LR rates below the threshold rate. We consider a Long Term Evolution (LTE) network on the LR links and the Wireless Local Area Network (WLAN) 802.11p Vehicle-to-Vehicle protocol on the SR links. Both schemes are compared to non-cooperative scenarios including LR unicasting, where the BS unicasts the data to each vehicle on the LR LTE links, and LR multicasting, where the BS multicasts the data to the vehicles on the LR LTE links. LTE scheduling is considered and scenarios with different numbers of allocated LTE resource blocks (RBs) on the LR links are investigated. The paper is organized as follows. Related work is reviewed in Section 2. The system model is presented in Section 3. Cluster formation methods are described in Section 4. The delay formulation and solution for the proposed schemes are derived in Section 5. Scheduling of LTE resources is presented in Section 6. Several simulation scenarios are studied and analyzed in Section 7. Finally, conclusions are drawn in Section 8.

2

2. Related Work Several studies in the literature focus on building public safety networks to ensure that critical information is delivered in case the existing communication is damaged by an incident. Law enforcement, fire department, emergency medical services, etc, require a reliable sharing of critical information which necessitates the need of employment of public safety wireless networks [1]. Incident Area Network (IAN) is studied in [3, 4]. In case of a disaster, the IAN replaces the damaged existing communication infrastructure in order to guarantee communication of post-disaster operations such as ambulances, emergency vehicles and rescues. In [3], an anchor node plays the role of relaying the received multicast traffic from a UMTS interface towards its ad-hoc neighbors. Radio resources, battery level, adverse context conditions (signal-to-noise ratio (SNR), high mobility) are exchanged with the anchor node which in turn sends them to local radio network control that will decide which cooperative terminal will be selected. In [4], a radio resource management policy monitors periodically the conditions of the network, terminals, and rescue team priorities to deliver multicast services to the team with major need. Simulations showed that HAP (High Altitude platform), which assists incident networks in offering broadcast/multicast services, can increase the number of multicast sessions when the terrestrial network does not exist. In [5] and [6], a cooperative automatic repeat request (ARQ) is presented, where, after leaving the coverage of an access point (AP), vehicles communicate between each others to exchange the packets that were lost during the transmission from the AP to the vehicle nodes. In this way, retransmissions of packets can be avoided and packet losses decreased, thus improving the throughput and transfer delay. In [6], cars broadcast HELLO messages to know about the presence of other neighbor cars and to notify other nodes that they need to act as cooperator. A list of cooperators is contained in the HELLO message. After a vehicle finishes from downloading data from the AP, it identifies the packets lost and requests them from other vehicles. Dynamic spectrum access is suggested in [7] for vehicles, where they sense the availability of spectrum before attempting to transmit. However, due to high mobility, shadowing, and other effects, spectrum sensing by a single vehicle may not lead to accurate information about the availability of the spectrum. Thus cooperative spectrum sensing is suggested in [7]. To implement optimized collaborative content distribution, signaling in3

formation needs to be exchanged between the vehicles themselves, and between the vehicles and the BS. Hence, the signaling overhead of any collaborative method should be taken into account. Other challenges incurred during device discovery and signaling include scalability, autonomous discovery, power efficiency and privacy. LTE Direct is being investigated as a possible device-to-device (D2D) communication platform addressing these challenges [8, 9]. It is currently being standardized as a feature in 3GPP Release 12 of LTE. It allows the discovery of more than 1000 devices within ranges of several hundred meters while maintain the quality of service (QoS) requirements [8]. In LTE Direct, the network can authorize devices to communicate directly (a feature needed in public safety applications in the absence of the macrocell network). However, LTE Direct can allow an operator to authorize and control the direct connection setup between vehicles, and to determine the user traffic routing between the direct and network paths. This can provide more control in fast changing environments and allows to maintain QoS (minimize loss of information through network intervention and control). In this work, cooperative content distribution schemes are developed in order to convey the critical information in public safety networks in the shortest possible time. The novelty in the proposed schemes is in relying jointly on LTE for LR communications and on IEEE 802.11p for SR communications, while taking LTE resource allocation into account. The same concept can be easily extended to LTE-Direct using D2D communications on the SR: in terms of adapting the simulation setup, this consists of selecting suitable LTE RBs for SR transmission instead of using 802.11p channels, and updating the simulation parameters accordingly. However, since IEEE 802.11p is standardized and widely used in the literature for vehicular communications, it will be used in this paper, especially that the implementation of LTE D2D in vehicular networks is not yet standardized (at the time of writing this paper). In the non-cooperative case, LTE multicast and broadcast multimedia services (MBMS) [10, 11, 12] can be used to send the data to all the concerned vehicles. Using LTE MBMS/eMBMS for non-cooperative multicasting in this paper, we dedicate a single RB in the cell to multicast the data of interest to the set of interested vehicles. However, in the case of multicasting, transmission on a given RB is limited by the rate achieved by the vehicle having the worst channel conditions on that RB.

4

Cluster Head

multic

rds wa rea To ent a id c in

g astin

2

LR

Li nk s

r ste Clu

Base Station

Cluster Head

rds wa rea To ent a id inc

r1 ste Clu

Figure 1: General system model.

3. System Model The system model adopted in this work is depicted in Fig. 1. Vehicles of public safety teams progress in cooperative clusters. In a given cluster, K cooperating vehicles in the range of an LTE BS form a vehicular ad-hoc network and head towards an incident area. The BS should communicate a content of common interest (files, maps, live video from disaster area, etc) to the vehicles, which can communicate with each other using IEEE 802.11p on the SR. LTE communication is considered on the LR links. In LTE, the available spectrum is divided into resource blocks (RB) consisting of 12 adjacent subcarriers, allocated in a 0.5 ms time slot. The shortest assignment unit consists of two consecutive slots, i.e., for a duration of 1 ms, which is the duration of one transmission time interval (TTI) [13, 14]. A total bandwidth of Wtot = 5 MHz, subdivided into NRB = 25 RBs of 12 subcarriers each is assumed. Each subcarrier has a bandwidth of Wsub = 15 kHz, such that the bandwidth of an RB is WRB = 180 kHz [13]. On the LR, we will consider 5

several cases where allocation is performed on the basis of one or more RBs for each cluster of vehicles. The BS downlink power, considered to be 5 Watts, is assumed to be subdivided equally among the RBs. The SR links use the WLAN 802.11p protocol which is different from 802.11 standard in that it added support for WLAN in a vehicular environment. It supports three different bandwidths: 5 MHz, 10 MHz and 20 MHz [15, 16]. The 802.11p usually operates in the 5.8 GHz and 5.9 GHz frequency. Four different power classes are defined in the 802.11p standard. They are presented in Table 1. Table 1: The Power Classes of 802.11p

Power Class A Class B Class C Class D

Max Output Power(dBm) 0 10 20 28.8

3.1. Channel Model The channels are modeled by pathloss, shadowing and fading. We use the term “node” to refer to either a vehicle or to the BS. The channel gain between nodes k and j over channel x (x could be any LTE subcarrier or the 802.11p channel used for SR communication) is given by: (x)

(x)

Hkj,dB = (−κ − υ log10 dkj ) − ξkj + 10 log10 Fkj ·

(1)

In (1), the first factor captures propagation loss, with dkj the distance between nodes k and j (where k = 0 corresponds to the BS in case of LR communication, and k, j ≥ 1 correspond to the vehicles), κ the pathloss constant and υ the path loss exponent. The second factor, ξkj , captures log-normal shadowing with a standard deviation σξ , whereas the last factor, (x) Fkj , corresponds to Rayleigh fading (generally considered with a Rayleigh parameter a such that E[a2 ] = 1). In the channel model, spatial shadowing correlation is taken into account since shadow fading values depend on the fixed location of obstacles [17]. Spatial correlation can be described as a measure of of how fast the local mean power evolves as the vehicle moves along a certain route [18]. We will 6

apply the correlated shadowing model of [18] and [19] where the shadowing correlation is expressed as: Λξ (∆dT , ∆dR ) = exp (−

∆dR + ∆dT ln 2), dcor

(2)

where ∆dR and ∆dT represent the movements of receiver and transmitter, respectively, and dcor is the decorrelation distance taken to be 20 meters for urban environment and 5 meters for the indoor environment. As the vehicles move, the model of (2) can be applied to determine the correlation between shadowing values at the different vehicle positions, in addition to the correlation between shadowing values of the same moving vehicle on the link with the BS (by setting ∆dT = 0 since the BS is fixed). In addition, for fast Rayleigh fading, a block fading model is considered, where the fast fading remains constant for a fixed time Tdec which is the channel de-correlation time. Then the channel conditions change and remain constant for another Tdec , and so on. 3.2. Data Rates Continuous rates can take any non-negative real value according to the Shannon capacity formula log(1 + γk,x ), where γk,x is the signal to noise ratio (SNR) of k over channel x. Conversely to continuous rates, discrete rates represent the quantized bit rates rk,x achievable in a practical system as follows:  r0 , η0 ≤ γk,x < η1      η1 ≤ γk,x < η2  r1 , η2 ≤ γk,x < η3 (3) rk,x (γk,x ) = r2 ,  . .   .. ..     rL−1 , ηL−1 ≤ γk,x < ηL , where ηl represents the SNR target in order to achieve the rate rl with a predefined BER. Note that in the limit, we have r0 = 0, η0 = 0, and ηL = ∞. In this paper, we consider discrete sets of Modulation and Coding Schemes (MCS) for the rate calculations according to (3), as is the case in practical standards. Thus, we adopt the discrete rates used in LTE and 802.11p. The 14 MCS used in LTE obtained from [20] are shown in Table 2, and the eight MCS used in 802.11p in addition to their corresponding data rates can be found in [15, 16] and are shown in Table 3. 7

Table 2: Discrete rates and SNR thresholds with 14 modulation and coding schemes

MCS

rl (bits/symb) ηl (dB)

No Transmission QPSK, R = 1/8 QPSK, R = 1/5 QPSK, R = 1/4 QPSK, R = 1/3 QPSK, R = 1/2 QPSK, R = 2/3 QPSK, R = 4/5 16-QAM, R = 1/2 16-QAM, R = 2/3 16-QAM, R = 4/5 64-QAM, R = 2/3 64-QAM, R = 3/4 64-QAM, R = 4/5 64-QAM, R = 1 (uncoded)

0 0.25 0.4 0.5 0.6667 1.0 1.333 1.6 2.0 2.6667 3.2 4.0 4.5 4.8 6.0

-∞ -5.5 -3.5 -2.2 -1.0 1.3 3.4 5.2 7.0 10.5 11.5 14.0 16.0 17.0 26.8

Table 3: The 802.11p discrete rates and SNR thresholds for different modulation and coding rates

Modulation BPSK BPSK QPSK QPSK 16QAM 16QAM 64QAM 64QAM

Coding Rate 1/2 3/4 1/2 3/4 1/2 3/4 2/3 3/4

Rate(Mbps) 3 4.5 6 9 12 18 24 27

η (dB) −2.2 −0.3 1.3 4.6 7 11.15 14 16

The long range discrete rate RL,k of a vehicle k, receiving from the BS with a rate rk,x = rl bits/symbol over the subcarriers of the RBs allocated

8

to it, can be calculated as follows: RL,k

RB SC TTI rl · NRB,k · NSC · NSymb · NSlot = , TTTI

(4)

RB where NRB,k is the number of RBs allocated to k, NSC is the number of SC subcarriers per RB (equal to 12 in LTE), NSymb is the number of symbols per subcarrier during one time slot (set to six or seven in LTE, depending TTI is the number of whether an extended cyclic prefix is used or not), NSlot time slots per TTI (two 0.5ms time slots per TTI in LTE), and TTTI is the duration of one TTI (1ms in LTE) [13].

3.3. Parameters and Variables The parameters that affect the delay in both schemes are the following: • ST : the size of the content to be sent from the BS. • RL,k : transmission rate on the LR links from the BS to vehicle k when the BS uses unicasting on the LR. • RL,T h : transmission threshold rate on the LR links from the BS to vehicles, when the BS uses threshold-based multicasting. • RS,kj : achievable transmission rate on the SR links from vehicle k to vehicle j. • SR (n): Number of remaining data bits at fading realization n. • nT : Number of channel variations until the whole content of size ST is distributed. In other words, we have SR (nT ) > 0 and SR (nT + 1) = 0. • k ∗ (n): Vehicle selected for transmission at channel realization n in order to minimize the content distribution delay at that channel realization. 4. Cluster Formation Methods In this section, after reviewing the most relevant cluster formation methods in vehicular networks, a simple approach that can be implemented with the schemes proposed in Section 5 is described.

9

4.1. Literature Review for Cluster Formation Several papers discuss cluster formation in vehicular networks. In [21], an energy efficient hierarchical clustering algorithm in wireless vehicular sensor networks is presented. It is based on a multi-tier system that reduces the amount of information that cluster nodes need to process. The first layer chooses the smallest transmission radius and controls it for one hop and the upper layer chooses a larger transmission radius by increasing the hop distance until the network radius reaches its maximum value. Nodes calculate the participation standards of candidate Cluster Head (CH), then nodes choose the best CH according to ratio between the node’s residual energy and the distance to the sink node. A hierarchical clustering algorithm for Vehicular ad-hoc networks (VANET) which does not require knowledge of the nodes locations is presented in [22]. The algorithm proves its robustness to topology changes caused by vehicular movement. It selects a ClusterRelay node to forward control messages from CH to slaves or vice versa. A SYNC message is defined as a message used to assign slots for cluster member nodes. A slave node that hears a SYNC message from two Cluster-Relays will announce itself as the nominated CH. After a random period of time, some nominees drop their nominations and remain slaves. In [23], clusters are not formed based on the location but on the direction of the vehicles to help in traffic management at intersections. Vehicles have digital maps that are split into different regions. Vehicles traveling in the same direction will be in one cluster. The first vehicle to enter the region in a particular direction will be referred to as CH. A direction based propagation function is used to eliminate unnecessary information exchange. A new clustering scheme called robust mobility adaptive clustering (RMAC) for high dynamic VANET is presented in [24], where vehicular mobility metrics, speed, location, and direction of travel are considered for clustering. RMAC forms one-hop clusters where a node assigns precedence to each of its neighbors. The one with highest precedence is selected as the CH. Due to high mobility, a node will need to store information on all nodes, which is a costly process. Instead, each node maintains neighbor information within its Zone of Interest (ZOI) and discards other information. Results show 35 seconds of cluster stability in high density networks with more than 50 nodes. A clustering scheme that supports multimedia and data applications using dedicated short-range communications (DSRC) supporting seven channels is presented in [25]. A vehicle is equipped with two transceivers: one used for communications with the CH vehicle over CRC (cluster range control) and 10

the other is used to transmit non-real time traffic over ICC (inter cluster control channel). A quasi-CH that is neither a CH nor a cluster member, receives a valid advertisement message from nearby CHs to join their clusters. A CH broadcasts ITJ (invite-to-join) messages every t units of time. It checks the signal strength and if it is greater than the predefined threshold, the ITJ message is considered valid. If a quasi-CH cannot receive a valid ITJ message within t units, it will elect itself as CH. An algorithm called Location Routing Algorithm with Directional ClusterBased Flooding (LORA-DCBF) is presented in [26]. Each CH maintains a Cluster Table which contains the addresses, directions and geographic locations of member and gateway nodes. Gateway nodes reduce duplicate retransmissions within the same region. Member nodes become gateway nodes when receiving messages from more than one CH. The source node updates its direction and location information before sending a packet and when the destination node receives the packet, its direction and location are updated which makes the algorithm suitable for fast mobile nodes. A mobility-based clustering scheme for VANET called Affinity PROpagation for VEhiclar networks, (APROVE) is presented in [27]. The algorithm elects a CH using affinity propagation. It finds clusters that minimize the distance from the CH to cluster members and relative mobility. Results show that clusters created are stable and have long average CH duration and long average Cluster member duration and low average rate of CH changes when compared to other known clustering algorithms. 4.2. Proposed Method for Cluster Formation Cluster formation can be done according to the methods described in [21, 22, 23, 24, 25, 26, 6, 27]. In the following, we present a cluster formation approach customized to the schemes proposed in Section 5 and having some intersections with [6, 27]: • Vehicles take turn in transmitting a certain message. We denote it as a Hello message as in [6]. The message contains information about the transmitting vehicle (its ID, IP address, etc), its channel state information (CSI) on the LR, i.e. its achievable rate on the RBs it is using to communicate with the BS, and a pilot sequence on the SR so that the other vehicles estimate their SR CSI with the transmitting vehicle.

11

• A vehicle that hears the message can estimate its SR CSI with the transmitting vehicle using the pilot sequence. When, in its turn, the receiving vehicle transmits its own message, the other vehicles will be able to estimate their SR CSI with it. • When no more Hello messages are transmitted on the SR, each vehicle would have a list of neighbors along with its CSI with them on the SR links, in addition to the LR CSI with the BS of all these neighbor vehicles. • Since the vehicles are moving together and correspond to the same team (e.g., police, or fire dept, or medical, etc) over the same public safety network, they can form a cooperative cluster. • The vehicle having the best LR CSI sends the information about the cluster to the BS: the cluster members, its SR CSI with them, and its CSI feedback about LR CSI with the BS. The BS would have a list of the cluster members after the above approach is implemented. The BS sends the data on the LR to the vehicle selected as CH by the proposed methods of Section 5, so that it relays it to the cluster members via SR multicasting. Although in the proposed schemes of Section 5 the CH may vary at each fading realization, the clusters themselves could be considered stable (the cluster members are the same regardless of the CH variations). In fact, since the vehicles of each public safety team move together to the disaster area, from the same starting point, the above approach can be considered as an initialization phase before any of the proposed schemes of Section 5 is implemented. This initialization approach can be repeated when needed, with large intervals between the repetitions, due to the high cluster stability expected. In vehicular networks other than public safety networks, the previous cluster stability assumption might not hold due to the dynamic uncoordinated movements of the vehicles to different destinations. In such a scenario, the proposed clustering approach might lead to non-negligible overhead that could hinder the fast content distribution benefits. 5. Proposed Collaborative Methods In this section, the proposed collaborative schemes for public safety vehicular ad-hoc networks are presented. Sections 5.1 and 5.2 present the schemes 12

based on LR unicasting and multicasting, respectively. These methods assume that vehicle clusters are already formed and hence they are implemented inside each cluster independently. 5.1. Scheme 1: SR Collaboration with LR Unicasting In this section, a vehicle is selected to multicast the data on the SR for the purpose of delay minimization. We consider that the selected vehicle receives the data on the LTE LR link before transmitting to the other vehicles on the 802.11p SR link. Thus, as the channel changes during each Tdec , the BS would select the optimal vehicle k ∗ that would minimize the content distribution delay in the cluster of vehicles. The fading fluctuations will lead to variations at each iteration of the vehicles receiving on LTE and transmitting on the SR 802.11p links. The remaining data bits at each fading realization are expressed as: SR (n) = ST −

n−1 X

RL,k∗ (y) (y) · Tdec ·

(5)

y=1

¡ ¢ Hence, during each fading realization, k ∗ receives B(k ∗ (n), n) = min RL,k∗ (n) (n) · Tdec , SR (n) bits on the LR, since the remaining bits SR (n) might be less than the total bits that can be transmitted on the LTE link between the BS and k ∗ , RL,k∗ (n) (n) · Tdec . We consider that the B(k ∗ (n), n) bits received on the LR can all be distributed on the SR at fading realization n since SR rates significantly higher than LR rates can be achieved due to the reduced distance between the members of the vehicle cluster. It should be noted that although high rates are achievable by LTE/LTEA compared to IEEE 802.11p, they are reached when a bandwidth of 20 MHz (100 RBs) is allocated to a unique user in the absence of intercell interference. In a loaded network with several users and a limited number of RBs allocated per user or cluster, as is the case in the paper, the “realistic” rates would be less. Using an SR technology does not automatically lead to higher rates: The high rates on the SR are due to the relatively short distances between vehicles. Thus, selecting the CH as the vehicle having favorable LR channel conditions reduces the delay on the LR link. The rate on SR links, in case the vehicles are moving together in a platoon, would be generally high since the vehicles are in relative proximity. In fact, if it is more beneficial for a vehicle to receive from the BS using LTE then receiving from k ∗ , then obviously this vehicle should be in a different cluster, or should be directly connected to the 13

BS (thus forming a cluster of its own), than being in the same cluster with k∗. The method followed in Scheme 1 to minimize the content distribution delay can be summarized as follows: • Step 1: Select the vehicle k ∗ (n) that can minimize the delay at fading realization n, i.e., that can deliver one bit of data in the shortest possible time. • Step 2: After determining k ∗ (n), send as many bits as possible to k ∗ (n) on the LR link during fading realization n. In order to minimize the delay for distributing the data to the members of the cooperative cluster, the BS should select vehicle k ∗ (n) at the nth channel realization, during Step 1 of the algorithm of Scheme 1, with k ∗ (n) given by:   1 1 1bit · k ∗ (n) = arg min Dcoop,k (n) = arg min  (6) + k k RL,k (n) min RS,ki (n) i6=k

Then the BS sends B(k ∗ (n), n) bits to k ∗ (n) during Step 2 of the algorithm of Scheme 1. The result of (6) is interesting. It indicates that the best vehicle should have the best combination of LR (represented by RL,k (n)) and SR (represented by mini6=k RS,ki (n)) characteristics, and it is not necessarily the vehicle having the best LTE rate on the LR. However, to implement the solution of (6), the BS needs information about the SR rates RS,ki . Feeding back this information in real-time in a fast fading scenario might not be practical. Therefore, we consider an LR approximation to (6), where the BS selects the vehicle having the highest LTE rate on the LR, in order to multicast the data using 802.11p on the SR: µ ¶ 1 ∗ k(LRapp) (n) = arg min = arg max RL,k (n)· (7) k k RL,k (n) With (7), no need to feed-back SR information to the BS. LR channel state information (CSI) can still be fed-back, as this is common in LTE and other wireless communication systems. In fact, LTE can maintain link performance at speeds up to 350 km/hour [28]. Thus, the selection in (7) can easily be implemented in practice by the BS. In the results of Section 7, we compare 14

the performance of the LR approximation to the optimal solution of (6) and show that it can achieve a good performance slightly worse than the optimal solution. This performance is due to the smart grouping of vehicles into cooperative clusters: the relatively short SR distances between all vehicles in the same cluster lead to relatively high SR data rates on most SR links. Consequently, the selection of the vehicle having best LR conditions as CH would not lead to a drastic performance degradation compared to selecting the best CH according to (6). The delay needed to distribute B(k ∗ (n), n) bits when selecting k ∗ (n) to relay the data in the nth channel realization is given by: Dcoop,k∗ (n) (n) = B(k ∗ (n), n)·   1 1  · + ∗ RL,k∗ (n) (n) min R (n) S,k (n)i ∗

(8)

i6=k (n)

The total delay to distribute the content is given by summing (8) over n (from n = 1 to nT ). 5.2. Scheme 2: SR Collaboration with LR Multicasting In Scheme 2, the BS multicasts the data on the LR using a rate RL,T h achievable by an SNR threshold γT h , rather than limiting the transmission rate to the worst user as in the non-cooperative multicasting scenario. A vehicle receives successfully from the BS if it has an SNR γk ≥ γT h . We consider the indicator variable δk (n): ½ 1 γk (n) ≥ γT h (n) δk (n) = (9) 0 γk (n) < γT h (n), where δk (n) = 1 means that vehicle k received successfully on the LR at the nth channel realization. The set of vehicles that received successfully on the LR at the nth channel realization is expressed as: U ∗ (n) = {k | γk (n) ≥ γT h (n)}·

(10)

We consider that the vehicles in U ∗ (n) receive the data on the LR before selecting one of them to transmit to the vehicles within the same cluster on

15

the SR. Hence, the delay incurred by having k ∈ U ∗ (n) multicast one bit of data on the SR after receiving it on the LR, is given by: 1bit Dcoop,k (n) =

1 RL,T h (n)

+

1 , min R (n) S,ki ∗

(11)

i∈U / (n)

where the first term corresponds to the time needed by vehicle k to receive the data on the LR, and the second term corresponds to the time needed for the other MTs to receive the data transmitted by k on the SR. Since multicasting is used for SR transmission, vehicle k transmits using the minimum achievable rate within the vehicles in the same cooperative cluster that failed to receive the data on the LR. Hence, all vehicles ∈ / U ∗ (n) will need the same time to receive the data. Thus, it is clear from (11) that in order to minimize the delay for distributing the data to the members of the cooperative cluster, k ∗ should be selected as follows: k ∗ (n) = arg min Dcoop,k (n) ∗ k∈U (n)

= arg max { min RS,ki (n)}· ∗ ∗

(12)

k∈U (n) i∈U / (n)

The maximum number of bits that can ³be transmitted by the ´BS during fading realization n using Scheme 2 is min RL,T h (n) · Tdec , SR (n) . As a result, the time Dk (n) needed by each individual vehicle to receive the content is expressed as: Dk (n) = min (RL,T h (n) · Tdec , SR (n))· µ ¶ 1 (1 − δk (n)) + , RL,T h (n) minj ∈U / ∗ (n) RS,k∗ (n)j (n)

(13)

where the first term corresponds to the transmission time on the LR and the second term indicates that vehicle k waits for the content on the SR only if it fails to receive on the LR, hence the presence of the term (1 − δk (n)). 5.3. Transmission Time versus Processing Time In the proposed schemes, the content distribution time is considered as the sum of the time needed to receive the content on the LR and the time needed to multicast it on the SR. Although this may seem as an approach that accommodates only transmission time, it is actually a worst case scenario that 16

3FNW

3FNW

3FNW/

7UDQVPLVVLRQ7LPHRQWKH/5

3FNW

3FNW

3FNW/

3URFHVVLQJ7LPHDIWHU5HFHSWLRQRQWKH/5

3FNW 3FNW

3FNW/

7UDQVPLVVLRQ7LPHRQWKH65

Figure 2: Content distribution time.

includes transmission and processing delays. In fact, the delay expressions in (8) and (13) are based on the assumption that transmission on the SR starts after the reception of the B(k ∗ (n), n) bits on the LR is complete. Assuming the bits transmitted at fading realization n consist of L packets, Fig. 2 describes the situation. The transmission of packet 1 on the SR could start just after its processing has ended (second time bar in Fig. 2). However, this transmission is delayed until packet L is completely received on the LR. The assumption in this paper is that the processing of packet L is completed before its transmission on the SR starts, which is generally the case, as Fig. 2 clarifies the concept. The reason for selecting this worst case scenario, that constitutes an upper bound (UB) on the actual delay, is that processing time depends on the devices used, and is also dependent on the application and operating system used (e.g. see [29], Sections V-C to V-E). The “best-case” scenario would consist of having a negligible processing time, and where the transmission on the SR starts as soon as the first bit is received on the LR. This would lead to a lower bound (LB) on the actual delay. The results of the proposed cooperative methods in Section 7.2 will be based on the worst case delay (UB). However, a comparison of the UB and LB will be performed in Fig. 10, in order to capture the range of actual performance between these two bounds. It should be noted that even if the transmission of a packet on the SR starts immediately after it is processed, the optimal solutions derived in Sections 5.1 and 5.2 for Schemes 1 and 2, respectively, would still be the

17

same, although the actual results of Section 7 would be better. 6. LTE Resource Allocation In this section, we present the resource allocation techniques adopted on the LTE LR communications for each of the investigated scenarios. We (c) consider that the BS allocates a fixed number of RBs NRB to be used on the LR transmissions to a given vehicle cluster. 6.1. LTE Resource Allocation for Scheme 1 Denoting by rk,x (n) the rate achievable by vehicle k over RB x at the nth channel realization, and by IRB,k the set of RBs allocated to k, then resource allocation is performed according to Algorithm 1. Algorithm 1 LTE Resource Allocation for Scheme 1 1: for k = 1 → K do 2: Set IRB,k ← ∅ Set RL,k (n) = 0 3: (c) 4: for i = 1 → NRB do 5: x∗k (n, i) ← arg max rk,x (n) x∈I / RB,k

Set IRB,k ← IRB,k ∪ {x∗k (n, i)} 6: RL,k (n) ← RL,k (n) + rk,x∗k (n,i) (n) 7: 8: end for 9: end for 10: Use RL,k (n) obtained above in (6) or (7) to determine k ∗ (n) for Scheme 1 or its LR approximation, respectively. 11: Allocate the RBs in the set IRB,k∗ to k ∗ and free the other RBs to be used by the BS for other clusters. (c)

Algorithm 1 finds the best NRB RBs for each vehicle. Then, the vehicle achieving the best performance according to the metrics of (6) or (7), when (c) its best NRB RBs are allocated to it, is selected as cluster head. 6.2. LTE Resource Allocation for Scheme 2 With Scheme 2, we define Ux∗ (n) = {k | γk,x (n) ≥ γT h (n)}, i.e., the set of vehicles that can receive the data successfully at the rate RL,T h (n) when RB x is used for multicasting by the BS. The notation | · | is used to denote 18

set cardinality. In addition, we denote by IRB,Scheme2 the set of RBs to be used with Scheme 2. The resource allocation for Scheme 2 can be defined as expressed in Algorithm 2. Algorithm 2 LTE Resource Allocation for Scheme 2 1: IRB,Scheme2 ← ∅ 2: for x = 1 → NRB do 3: Nserved,x ← |Ux∗ (n)| 4: end for (c) 5: for i = 1 → NRB do 6: x∗ (n, i) ← arg max Nserved,x x∈I / RB,Scheme2

7: IRB,Scheme2 ← IRB,Scheme2 ∪ {x∗ (n, i)} 8: end for 9: U ∗ (n) = ∪x∈IRB,Scheme2 Ux∗ (n)

Algorithm 2 finds, for each RB, the number of vehicles Nserved,x that can successfully receive the data when RB x is used for threshold-based multicasting according to Scheme 2 at a rate RL,T h (n). The algorithm finds (c) the NRB RBs such that, when the same data is multicast on each of these RBs, the maximum number of vehicles can successfully receive this data on the LR. 6.3. LTE Resource Allocation for Non-Cooperative Unicasting In the non-cooperative scenario with LR unicasting, for a fair comparison (c) with Scheme 1, we consider that NRB are allocated to each vehicle, as long as there are available RBs. We define Iavail−RB as the set of available RBs and Iavail−V as the set of available vehicles during the scheduling sequence. The resource allocation approach is described in Algorithm 3. This approach allocates RBs to vehicles in a way to maximize LR perfor(c) mance by selecting the best NRB RBs for each vehicle, as long as there are RBs available. 6.4. LTE Resource Allocation for Non-Cooperative Multicasting In the case of non-cooperative multicasting, transmission on a given RB is limited by the rate achieved by the vehicle having the worst channel conditions on that RB. Denoting by IRB,M the set of RBs used for non-cooperative 19

Algorithm 3 LTE Resource Allocation for Non-Cooperative Unicasting 1: Set Iavail−RB ← {1, 2, · · ·, NRB } 2: Set Iavail−V ← {1, 2, · · ·, K} 3: Set IRB,k ← ∅∀k 4: while (Iavail−RB 6= ∅) AND (Iavail−V 6= ∅) do 5: Find the pair (Vehicle k ∗ ,RB x∗ ) such that: (k ∗ (n), x∗ (n)) = {arg max rk,x (n)} k,x

6:

(14)

Mark RB x∗ (n) as occupied: Iavail−RB ← Iavail−RB − {x∗ (n)}

7:

Add RB x∗ (n) to the set of RBs allocated to k ∗ (n): IRB,k∗ (n) ← IRB,k∗ (n) ∪ {x∗ (n)}

8: 9:

(c)

if |IRB,k∗ (n) | = NRB then Mark vehicle k ∗ (n) as served: Iavail−V ← Iavail−V − {k ∗ (n)}

10: end if 11: end while

Algorithm 4 LTE Resource Allocation for Non-Cooperative Multicasting 1: Set IRB,M ← ∅ 2: Set RL,M (n) = 0 (c) 3: for i = 1 → NRB do ³ ´ 4: x∗ (n, i) ← arg max min rk,x (n) x∈I / RB,k

k ∗

Set IRB,M ← IRB,M ∪ {x (n, i)} 5: 6: RL,M (n) ← RL,M (n) + mink rk,x∗ (n,i) (n) 7: end for

20

LR multicasting, and by RL,M (n) the multicasting rate, then resource allo(c) cation is performed according to Algorithm 4. Algorithm 4 selects NRB RBs such that, when the data to be transmitted is subdivided over these RBs, the highest multicasting rate is achieved. 7. Results and Analysis 7.1. Simulation Model Three clusters of vehicles heading towards an incident area are considered. These could correspond, for example, to police, fire, and medical emergency teams heading towards an incident area. Since each team moves from its own headquarters (located in different locations in general), the clusters are formed according to the method of Section 4.2 and remain the same during the simulation time. In the simulations, we assume that the vehicles of each cluster move along a highway section of 400 meters long and 30 meters wide in one direction without sharp turns and whose origin is at a distance dLR = 700 m from the BS. Velocities for each vehicle are generated in the range of 70 km/hour to 100 km/hour. Furthermore, we consider a file of size ST = 1 Mbits to be transmitted to all requesting vehicles. In addition, we consider an LTE bandwidth on the LR WLR = 5 MHz, subdivided into NRBs = 25 RBs of 12 subcarriers each [13, 14]. We consider a 5 Watts BS transmit power, subdivided equally among all RBs. Channel parameters are obtained from [30]: κ = −128.1 dB, υ = 3.76, and σξ = 8 dB. The results are obtained with a 95% confidence interval. The time where fading is considered constant is taken to be Tdec = 10 ms. This is in line with the guidelines discussed in [29] (Section V-F) based on channel coherence time. In fact, on the SR, the speeds vary between 70100 km/hour. With IEEE 802.11p at 5.9 GHz carrier frequency, and an average speed difference between neighboring vehicles of 15km/h, this would correspond to a channel coherence time of around 12 ms. On the LR, with an average speed of 85 km/h (with respect to the fixed LTE BSs), and a carrier frequency of 1.8 GHz, this would correspond to a channel coherence time of around 6.7 ms. Speeds of 70 km/h and 100 km/h would correspond to coherence times of 8.1 ms and 5.7 ms, respectively. Thus, setting Tdec = 10 ms would be a reasonable tradeoff that is close enough to the LR and SR coherence times.

21

2 Scheme 1 LR App Scheme 2 No Coop Unicast No Coop Multicast

1.9

Maximum Delay in sec

1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 1

0

2

4 6 Number of cooperating vehicles

8

10

Figure 3: Maximum distribution time for γT h = 30 dB and using power class D and one RB.

7.2. Simulation Results This section presents the simulation results in terms of maximum distribution time per cluster, i.e., the time needed for all the vehicles in the cluster to receive the whole content. The results are presented versus the number of vehicles per cluster. With γT h = 30 dB for Scheme 2, and transmitting data on one RB on the LR links and using power class D for the transmit power on 802.11p, Fig. 3 shows that Scheme 1 achieves the lowest distribution time, followed by the no cooperation with unicasting case where the BS transmits in parallel to all users on their highest RB rate. In addition, Scheme 2 achieves a shorter distribution time than no cooperation with multicasting, which achieves the worst distribution time since the transmission is limited to the minimum rate among vehicles. As we lower the transmit power on the SR to power class C, Scheme 1 still achieves the lowest distribution time when the number of vehicles is below nine as shown in Fig. 4. However, beyond nine vehicles, ’no cooperationunicasting’ achieves the lowest distribution time. The reason is that by decreasing the transmit power on the SR links, the SR rates also decrease and vehicles will be getting the content at slower rate on SR links, which increases the content distribution time. Decreasing the transmit power on the SR further to power class B, the results of Fig. 5 show that cooperation is no longer efficient in terms of distribution time and that non-cooperative unicasting and multicasting become better, regardless of the number of cooperating vehicles. The reason is that 22

2 LR App Scheme 1 Scheme 2 No Coop Multicast No Coop Unicast

1.9

Maximum Delay in sec

1.8 1.7 1.6 1.5 1.4 1.3 1.2

0

2

4 6 Number of cooperating vehicles

8

10

Figure 4: Maximum distribution time for γT h = 30 dB and using power class C and one RB.

4.5

Maximum Delay in sec

4

3.5 Scheme1 LR App Scheme2 No Coop Unicast No Coop Multicast

3

2.5

2

1.5 0

2

4 6 Number of cooperating vehicles

8

10

Figure 5: Maximum distribution time for γT h = 30 dB and using power class B and one RB.

at a very low SR transmit power, SR rates start to become very low, compared to a higher transmit power available on the LR at the LTE BS. The results of power class A are expected to be worse than those of power class B. A remarkable conclusion from Figs. 3 to 5 is that the LR approximation of Scheme 1 is very close to the optimal solution in terms of performance. 23

This result is interesting, since it indicates that the LTE BS does not need to have information about the SR CSI. Standard LR CSI feedback is sufficient to select an appropriate vehicle in order to multicast data to its peers using 802.11p. This performance is due to the smart grouping of vehicles into cooperative clusters: the relatively short SR distances between all vehicles in the same cluster lead to relatively high SR data rates on most SR links. Consequently, the selection of the vehicle having best LR conditions as CH would not lead to a drastic performance degradation compared to selecting the best CH according to Scheme 1. In fact, when vehicles are close enough to each other, the first term in (6) becomes dominant in the delay calculation, and thus the contribution of LR transmission becomes more evident. This is naturally not the case when the distances between vehicles increase, and thus the SR data rates decrease. For the following results, we will be using Power class D. From Fig. 6, we notice that for Scheme 2, when the number of vehicles is low, the maximum delay is too high. The reason is that, when γT h = 30 dB, the probability of having one vehicle exceed γT h is low. Hence, when the number of vehicles is low, it might occur that none of them is receiving any data on the LR and in that case, they have to wait for Tdec to elapse and check if they can receive data at the next fading realization. Comparing Fig. 3 that uses one RB on the LR and Fig. 6 that uses two RBs on LR, we notice that ‘no cooperation with multicasting’ performs better when using two RBs. In fact, when unicasting on LR and as the number of vehicles increases, some users may not be allocated an RB to transmit data and thus the delay Tdec is added to the overall delay. On the other hand, multicasting on LR transmission is limited to the minimum rate among vehicles and thus it is guaranteed that the transmission on the LR will occur. Increasing the number of allocated RBs leads to higher achievable multicasting rates. Fig. 7 shows that Scheme 2 becomes better than ‘no cooperation with unicasting’ when the number of vehicles exceeds 7 when using 6 RBs on the LR. In fact, the scheme ‘no cooperation with unicasting’ starts increasing linearly as the number of vehicles increases. The reason behind this linear increase is that beyond 4 vehicles, we do not have enough RBs for all vehicles (the total number of RBs is 25 and six RBs are allocated per vehicle) and some of them do not get any data so they wait for Tdec each time. Furthermore, with two RBs, Fig. 8 shows that maximum delay for Scheme 1 is about 0.8 sec and for ‘no cooperation with multicasting’ about 1.7 sec. When using 6 RBs, Fig. 8 shows that maximum delay has decreased for Scheme 1 to 0.4 24

12

Maximum Delay in sec

10

8

6

4

Scheme1 LR App Scheme2 No Coop Unicast No Coop Multicast

2

0 0

2

4 6 8 Number of cooperating vehicles

10

Figure 6: Maximum distribution time for γT h = 30 dB and using two RBs.

12 Scheme1 LR App Scheme2 No Coop Unicast No Coop Multicast

Maximum Delay in sec

10

8

6

4

2

0 0

2

4 6 8 Number of cooperating vehicles

10

Figure 7: Maximum distribution time for γT h = 30 dB and using six RBs.

sec and for ‘no cooperation with multicasting’ to 0.78 sec. The main conclusions from the simulation results can be drawn as follows. Non cooperative unicasting requires a certain number of RBs to be allocated to each vehicle. As the number of vehicles increases, the number of available RBs becomes insufficient to accommodate all vehicles, which leads to performance degradation. Scheme 2 achieves good performance when one RB is used for multicasting. When the number of RBs dedicated by the BS to a single cluster increases, the performance of Scheme 2 is enhanced. However, the performance of the other schemes gets enhanced faster which makes them outperform Scheme 2, except for non-cooperative unicasting when the number of vehicles increases. In fact, with Scheme 2, the BS is not using CSI information in order to perform multicasting. Reducing RL,Th increases the content distribution time, whereas increasing RL,Th requires a higher SNR γTh to be available on the LR links for the vehicles to receive the data. When 25

3 Scheme1−1 RB Scheme1−2 RBs Scheme1−4 RBs Scheme1−6 RBs LR App−1 RB LR App−2 RBs LR App−4 RBs LR App−6 RBs No Coop Multicast−1 RB No Coop Multicast−2 RBs No Coop Multicast−4 RBs No Coop Multicast−6 RBs

2.5

Maximum delay in sec

2

1.5

1

0.5

0

0

1

2

3

4 5 6 Number of cooperating vehicles

7

8

9

10

Figure 8: Maximum Distribution Time for Scheme 1, the LR approximation, and non cooperative multicasting.

the high RL,Th cannot be achieved, the vehicles have to wait for other fading realizations in order for at least one of them to have an SNR greater than γTh , which increases the delay. On the other hand, non-cooperative multicasting and Scheme 1 along with its LR approximation benefit well from the increase of the number of RBs allocated to the cluster. The enhancements as the number of RBs increases are shown in Fig. 8 for these three schemes. In fact, with multicasting, the data is transmitted over all the RBs allocated by the BS to the cluster. This leads to higher transmission rates on the LR and hence lower time to distribute the content, especially that with non-cooperative multicasting, the worst-case CSI is taken into account conversely to Scheme 2. With Scheme 1 and its LR approximation, all the RBs allocated to the cluster are used to unicast the data to the selected cluster head. Since this cluster head is selected based on its high achievable rates, and since increasing the RBs allows it to achieve higher rates, then this justifies the superior performance of Scheme 1 over the other schemes, followed by its LR approximation. Hence, using a limited number of RBs on the LR jointly with SR collaboration can achieve a fast content distribution time. The remaining RBs can be spared to communicate with other vehicle clusters. Thus, the approach presented in this paper complements the use of LTE, for example in incident 26

1 0.9 0.8

Matching Probability

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0

2

4 6 Number of cooperating vehicles

8

10

Figure 9: Probability of selecting the same CH by both the optimal solution and LR approximation.

1.6 1.4

Maximum Delay in sec

1.2 Equal Parts: Sequential (1 RB) Equal Parts: Parallel (K RBs) Scheme 1 − UB: 1 RB Scheme 1 − LB: 1 RB Scheme 1 − UB: K RBs Scheme 1 − LB: K RBs

1 0.8 0.6 0.4 0.2 0

0

2

4 6 Number of cooperating vehicles

8

10

Figure 10: Comparison of the optimal solution of Scheme 1, using the upper and lower bounds, and previous related work.

scenarios as in [2], by allowing the use of less LTE resources (communication with only a single vehicle in each cluster in Scheme 1 and its LR approximation) while ensuring faster distribution of the data through the use of SR collaboration over 802.11p. It should be noted that, although the performance of LR approximation ∗ is close to optimal, k(LRapp) is different than k ∗ most of the time. In fact, ∗ Fig. 9 shows that the probability of having k(LRapp) = k ∗ decreases as the number of vehicles increases. It should also be noted that the previous results of the proposed schemes correspond to the scenario of Fig. 2, representing an upper bound on the actual delay. Fig. 10 shows a comparison of the delays in both the upper and lower bound scenarios described in Section 5.3. The actual performance would be between these two bounds. The delay gap between the UB and LB 27

scenarios is relatively small. Thus, these two bounds can be considered tight with the simulation framework of this paper. Furthermore, the UB and LB are compared in Fig. 10 to a cooperative content sharing approach widely used in the literature [31, 32, 33, 34]. This approach consists of dividing the content into a number of parts equal to the number of cooperating vehicles, and sending one part to each vehicle on the LR in order to share it with the other vehicles on the SR. Sending the data can be done sequentially on the LR (1 RB for all vehicles, and the parts are sent in sequence), or in parallel (for K vehicles, transmission occurs simultaneously on K RBs, one for each vehicle), depending on the bandwidth available. For a fair comparison with the sequential and parallel scenarios, the UB and LB of Scheme 1 were implemented with 1 RB and K RBs dedicated for LR transmission per CH, respectively. The performance of Scheme 1 was superior in both scenarios. 7.3. Remarks, Limitations and Extensions It should be noted that all vehicles in the simulation model are assumed to be able to communicate successfully with the LTE BS(s). SR IEEE 802.11p communication is used to complement the LTE network in order to ensure faster content distribution to the vehicles, when the rates achievable on the LR are lower than the SR rates. Thus, 802.11p is not used to replace any loss of LTE connectivity, although this could be an interesting topic for future research. In the results of Section 7.2, multicasting was used in each cluster on the IEEE 802.11p SR links. Hence, only one channel per cluster is needed for SR content distribution, with only the CH transmitting at a given instant. Therefore, there is no risk of collisions or interference inside each cluster. Furthermore, interference between clusters is not an issue, since the clusters are assumed to correspond to different public safety teams moving from different origins along different routes. Even in the case of neighboring clusters, three orthogonal 802.11p channels can be used, one in each cluster. Thus, interference can be avoided between each cluster and its two nearest neighboring clusters. However, it should be noted that sub-clustering inside the group of vehicles of each public safety team is not considered in this paper. In other words, we adopt the model of Fig. 1 and not that of Fig. 11. In fact, it might be more efficient to subdivide the cluster of each team into different sub-clusters, with

28

Cluster Head

multic

rds wa rea To ent a id c in

g astin

Base Station

r3 ste Clu

ter 4

Li nk s

Clus

LR

Cluster Head

Cluster Head

rds wa rea To ent a id inc

r1 ste Clu

r2 ste Clu

Cluster Head

Figure 11: System model with advanced clustering.

each sub-cluster having its own CH, as shown in Fig. 11. However, this approach requires dynamic cluster modifications at fast rates and a significant amount of overhead. Implementing such an approach to reduce LR delays while taking the needed overhead into account is a challenging task worthy of future investigation. 8. Conclusions Cooperative delay-sensitive content distribution in public safety vehicular ad-hoc networks was investigated, and two novel schemes were proposed. LTE was used for long range communications and 802.11p was used for cooperation on the short range. A high mobility environment with correlated shadowing was considered. The performance was studied in the presence of LTE resource allocation on the long range, with a variable number of LTE resource blocks allocated for the investigated methods. The proposed schemes were compared to each other in addition to the non-cooperative 29

content distribution techniques, in terms of the maximum time needed for content distribution. Simulation results showed that the best performance was achieved when the base station unicasts the content to a selected vehicle on the long range, and that vehicle multicasts the data to other vehicles within its cooperating cluster. A practical approximation to the optimal solution was also presented. Furthermore, it was shown that within ranges of practical interest, the approach is beneficial with power classes C and D of 802.11p. Acknowledgment The authors would like to thank the anonymous Reviewers and the Editor for their comments that helped in significantly enhancing the clarity and quality of the paper. This work was made possible by NPRP grant # 09-180-2-078 from the Qatar National Research Fund (a member of The Qatar Foundation). The statements made herein are solely the responsibility of the authors. References [1] J. Song, L. Trajkovic, Modeling and performance analysis of public safety wireless networks, in Proc. 24th IEEE International Performace, Computing and Communications Conference (IPCCC 2005) (Phoenix, Arizona, April 2005) 567 – 572. [2] Motorola, Barricaded suspect incident analysis: enhancing critical incident response with public safety LTE (2011). [3] F. Araniti, M. D. Sanctis, S. Spinella, M. Monti, E. Cianca, A. Molinaro, A. Iera, M. Ruggieri, Cooperative terminals for incident area networks, in Proc. 1st International Conference on Wireless Communication Vehicular Technology, Information Theory and Aerospace and Electronic Systems Technology 2009 (Wireless VITAE’09) (Aalborg, Denmark, May 2009) 549–553. [4] F. Araniti, M. D. Sanctis, S. Spinella, M. Monti, E. Cianca, A. Molinaro, A. Iera, M. Ruggieri, Hybrid system HAP-WiFi for incident area network, in Proc. of 2nd International ICST Conference on Personal Satellite Services (PSATS 2010) Conference (Rome, Italy, February 2010) 436–450. 30

[5] O. Trullols-Cruces, J. Morillo-Pozo, J. Barcelo, J. Garcia-Vidal, A cooperative vehicular network framework, in Proc. IEEE International Conference on Communications (ICC’09). [6] J. Morillo-Pozo, O. Trullols, J. Barcelo, J. Garcia-Vidal, A cooperative ARQ for delay-tolerant vehicular networks, in Proc. 28th International Conference on Distributed Computing Systems Workshops (ICDCS’08) (Beijing, China, June 2008) 192–197. [7] D. Borota, G. Ivkovic, R. Vuyyuru, O. Altintas, I. Seskar, P. Spasojevic, On the delay to reliably detect channel availability in cooperative vehicular environments, in Proc. 73rd IEEE Vehicular Technology Conference (VTC Spring). [8] Qualcomm, LTE Direct [last accessed: June 2013]. URL http://www.qualcomm.com/research/projects/lte-direct [9] Qualcomm, LTE Direct: The Case for Device-to-Device Proximate Discovery, Qualcomm White Paper. [10] 3rd Generation Partnership Project (3GPP), 3GPP TS 23.246 Technical Specification Group Services and System Aspects; Multimedia Broadcast/Multicast Service (MBMS); Architecture and functional description, version 11.1.0, Release 11 (2012). [11] 3rd Generation Partnership Project (3GPP), 3GPP TS 26.346 Technical Specification Group Services and System Aspects; Multimedia Broadcast/Multicast Service (MBMS); Protocols and codecs, version 11.4.0, Release 11 (2013). [12] S. Lu, Y. Cai, L. Zhang, J. Li, P. Skov, C. Wang, Z. He, Channel-Aware Frequency Domain Packet Scheduling for MBMS in LTE, in Proc. IEEE VTC-Spring 2009 (Barcelona, Spain, April 2009) 1–5. [13] 3rd Generation Partnership Project (3GPP), 3GPP TS 36.211 3GPP TSG RAN evolved universal terrestrial radio access (E-UTRA) physical channels and modulation, version 11.2.0, Release 11 (2013). [14] 3rd Generation Partnership Project (3GPP), 3GPP TS 36.213 3GPP TSG RAN evolved universal terrestrial radio access (E-UTRA) physical layer procedures, version 11.2.0, Release 11 (2013). 31

[15] Rohde and Schwarz, WLAN 802.11p measurements for vehicle to vehicle (V2V) DSRC, Rohde and Schwarz Application Note. [16] IEEE 802.11, Wireless LAN medium access control (MAC) and physical (PHY) layer specifications, amendment 6: wireless access in vehicular environments (2010). [17] I. Forkel, M. Schinnenburg, M. Ang, Generation of two-dimensional correlated shadowing for mobile radio network simulation, In Proc. of the 7th International Symposium on Wireless Personal Multimedia Communications (WPMC 2004). [18] Z. Wang, E. Tameh, A. Nix, Joint shadowing process in urban peer-topeer radio channels, IEEE Transactions on Vehicular Technology 57 (1) (January 2008) 52–64. [19] K. Yamamoto, A. Kusuda, S. Yoshida, Impact of shadowing correlation on coverage of multihop cellular systems, in Proc. IEEE International Conference on Communications (ICC 2006) (Istanbul, Turkey, June 2006) 4538–4542. [20] 3rd Generation Partnership Project (3GPP), 3GPP TR 36.942 3GPP TSG RAN evolved universal terrestrial radio access (E-UTRA) radio frequency (RF) system scenarios, version 8.1.0, Release 8 (2008). [21] Y. Liu, N. Xiong, Y. Zhao, A. Vasilakos, J. Gao, Y. Jia, Multi-layer clustering routing algorithm for wireless vehicular sensor networks, Institution of Engineering and Technology (IET) Communications 4 (7) (April 2010) 810–816. [22] E. Dror, C. Avin, Z. Lotker, Fast randomized algorithm for hierarchical clustering in vehicular ad-hoc networks, in Proc. 10th IFIP Annual Mediterranean Ad Hoc Networking Workshop (Med-Hoc-Net’2011) (Sicily, Italy , June 2011) 1–8. [23] N. Maslekar, M. Boussedjra, J. Mouzna, L. Houda, Direction based clustering algorithm for data dissemination in vehicular networks, in Proc. IEEE Vehicular Networking Conference (VNC’09). [24] R. Goonewardene, F. Ali, E. Stipidis, Robust mobility adaptive clustering scheme with support for geographic routing for vehicular ad hoc 32

networks, Institution of Engineering and Technology (IET) Intelligent Transport Systems 3 (2) (June 2009) 148–158. [25] H. Su, X. Zhang, Clustering-based multichannel MAC protocols for QoS provisionings over vehicular ad hoc networks, IEEE Transactions on Vehicular Technology 56 (6) (November 2007) 3309–3323. [26] S. Momeni, M. Fathy, N. Mozayani, Vehicle Ad-Hoc Networks Clustering, in Proc. 4th Advanced Satellite Mobile Systems Conference (ASMS’08). [27] C. Shea, B. Hassanabadi, S. Valaee, Mobility-based clustering in VANETs using affinity propagation, in Proc. IEEE Global Telecommunications Conference (GLOBECOM’09). [28] J. Zyren, Overview of the 3GPP long term evolution physical layer, Freescale Semiconductor White Paper, July 2007. [29] G. Marfia, M. Roccetti, A. Amoroso, G. Pau, Safe Driving in LA: Report from the Greatest Intervehicular Accident Detection Test Ever, IEEE Transactions on Vehicular Technology 62 (2) (February 2013) 522–535. [30] 3rd Generation Partnership Project (3GPP), 3GPP TR 25.814 3GPP TSG RAN physical layer aspects for evolved UTRA, v7.1.0 (2006). [31] F. Albiero, M. Katz, and F. Fitzek, Energy-Efficient Cooperative Techniques for Multimedia Services over Future Wireless Networks, IEEE ICC’08 (2008) 2006–2011. [32] F. Albiero, J. Vehkaper¨a, M. Katz, and F. Fitzek, Overall Performance Assessment of Energy-Aware Cooperative Techniques Exploiting Multiple Description and Scalable Video Coding Schemes, Communication Networks and Services Research Conference (CNSR’08) (2008) 18–24. [33] M. Ramadan, L. Zein, Z. Dawy, Implementation and Evaluation of Cooperative Video Streaming for Mobile Devices, In Proceedings of IEEE PIMRC 2008, Cannes, France, September 2008. [34] E. Yaacoub, L. Al-Kanj, Z. Dawy, S. Sharafeddine, A. Abu-Dayya, Energy Efficient Fair Content Distribution over LTE Networks using Bluetooth Piconets, 26th Wireless World Research Forum (WWRF), Doha, Qatar, April 2011. 33