Performance evaluation for the proposed routing algorithm

In this subsection, this study evaluates our forwarding scheme over a simple bi-directional highway scenario consisting of 6 lanes in total, (i.e., 3 lanes per direction) as depicted in Figure 19. The vehicular traffic traces are generated by the freeway mobility model of IMPORTANT [27]. When vehicles leave the road, there is a probability to make them re-enter the road from the same/different direction lanes.

The simulation scenarios here can be also mapped to transmit video packets in the specific road segments. Furthermore, both the position of the video source and the destination in our setting are fixed for simplicity. The testing map contains a 4 km straight highway, and the video source is always 500 meters away from the one end of the road. Actually, in highway scenarios, the position of the video destination is not

necessary to be static (e.g., vehicle). The source vehicle can predict the current position of the destination by the driving direction and velocity information. The distance between the source and the destination referred by the MDP-based forwarding scheme can also be updated depended on the requirement of accuracy.

Figure 19: The used highway scenario.

TABLE II: Parameters used for algorithm evaluation.

Parameter value

Rd. seg. Length (m) 500, 1000, 1500, 2000, 2500, 3000

Lane width (m) 3.75

Road density (#/km) 90 (sparse), 120 (medium), 180 (dense) Vehicle speed (m/s) 22.22 ~ 33.33 (= 80 ~ 120 km/hr)

Acceleration (m²/s) 3.0

Video deadline (s) 0.1, 0.2, 0.3

HELLO interval (s) 2

Cell length (m) 50

4.2.1. RSNRs under different road lengths and road densities

To observe the impacts of the road segment lengths and the road densities for our forwarding scheme, we design simulation scenarios which compose different combinations of such parameters. By referencing the simulation settings of [16], we classify the road density into 3 categories: sparse (90 vehicles/km), medium (120 vehicles/km) and dense (180 vehicles/km). The length of road segments are classified into 6 cases: 500m, 1000m, 1500m, 2000m, 2500m and 3000m. As we can see from Figure 20, the proposed forwarding algorithm can yield excellent video quality for the

end user even the road length (or the distance between the source vehicle and the destination vehicle) is far as 3 km and the shadowing propagation model is applied.

According to our simulation results, we find that the average video qualities of the medium case are slightly greater than the sparse case and the dense case. The medium case can potentially offer more vehicles to be the forwarder candidates than the sparse case, and thus the packet senders have more opportunities to pick good vehicles. For the aspect of the dense case, the periodic HELLO messages broadcast mechanism permit the extra provided vehicles to bring more chances of packet collision, wireless channel congestion, so the distortions are severer than the medium case.

Figure 20: PSNR results of different road density and road segment length under 0.3 second video deadline.

4.2.2. PSNRs under different video deadlines

The video decoding deadline is especially an important consideration for such

time-sensitive applications. An expired video packet cannot be decoded by the destination vehicle’s decoder; therefore, the contained video frame information will lose, and the video quality will be distorted. Three different video deadlines are set, 0.3 second, 0.2 second and 0.1 second, to observe the impacts from deadline expiration. As shown in Figure 21, the shorter video deadline leads to worse PSNR, because more packets violate the deadline constraint and then become helpless in the decoding processes.

Figure 21: PSNR results of different video deadline under the medium density.

4.2.3. Comparison of the forwarding schemes

This simulation compares the achieved performance of the three forwarding schemes: the proposed MDP-based forwarding, the greedy forwarding and the random forwarding. The proposed forwarding scheme applies the best hop distance according to the distortion information to make the packet forwarding decisions. The

greedy forwarding is widely-used in many literatures [6][16][28]: the sender node tries to pick the neighbor node which is closest to the destination node as the next hop.

This approach can find the path with minimum hop count. Finally, the random forwarding is designed to pick the next hop arbitrarily from all the closer neighboring vehicles to the destination vehicle within the transmission range. From Figure 22, we realize that the MDP-based forwarding scheme can achieve excellent video quality than greedy and random forwarding when the wireless radio is not ideal. All the facts related to video distortion, such as radio fading and the video decoding deadline, have been considered by the proposed distortion model, so the MDP-based forwarding can know how to pick the next hop accordingly. We notice that even the random forwarding is also outperform than the greedy one, because the hop distances picked by the greedy forwarding are very probably larger than the distances picked by the random forwarding, such hop count based approach is inappropriately under the fading channel environment.

Figure 22: PSNR results of different forwarding schemes.

4.2.4. Comparison of the all action consideration and the single action considerations

In this subsection, the simulation is to delve how action considerations influence the performance of the proposed MDP-based forwarding scheme. The applied vehicular densities in the above simulations are actually not sparse enough even the

“sparse” density. Under such density settings, a sender vehicle can easily find its next hop in the best cell in most cases, so we cannot clearly figure out the benefit of considering different actions. For this reason, we design two additional scenarios with sparser vehicular densities: 25 vehicles/km and 50 vehicles/km. Moreover, we also wonder the performance under different cell lengths, because the smaller cell lengths make the target cell selections more diversely, sender vehicles have more chances to pick different cells to forward packets, and the larger cell lengths are in the opposite

way. The cell length settings are 30 m and 50 m.

(a)

(b)

(c)

(d)

(e)

(f)

Figure 23: All action consideration vs. single action considerations

The simulation results are depicted as Figure 23. We can observe that the all action consideration approach yields better PSNR performance than the single action consideration approaches, and the video quality improvement is relatively large when the vehicular density is lower. In the sparser density scenarios, the destination nodes may be unable to decode the received video packets successfully because the needed I-frames were lost during the packet delivery procedures. The data counted by Figure 23 (b) and (e) contain such problems, so some of the lines are in oscillation form. We filter the data to ignore the ones with zero PSNR value, and replot the figures as Figure 24 (c) and (f) to let the PSNR trends make sense. There is another point we can observe, the smaller cell length consideration does not affect the PSNR performance too much, and this shows we do not really need small cell length in the MDP-based forwarding scheme, thus the extra time complexity can be further reduced.