A Selective Downlink Scheduling Algorithm to Enhance Quality of VOD Services for WAVE Networks

2.1. Video Frame and Packet

In video compression technologies, such as MEPG and H.264 [4], encoded pictures (or frames) are arranged in groups of pictures (GOPs). An encoded video stream consists of successive GOPs. A GOP can contain the following frame types: I-frame, P-frame, and B-frame. The order of intraframes and interframes is specified in a GOP. An I-frame is a reference picture which is intracoded corresponding to a fixed image and it is independent of other pictures. A P-frame is predictive-coded frame which contains motion-compensated difference information from the preceding I- or P-frame. A B-frame is bidirectionally predictive-coded frame which contains different information from the preceding and following I- or P-frame within a GOP. I-frame and P-frame are often referred to as anchor frames. A GOP always begins with an I-frame. Afterwards, several P-frames follow. The B-frames are inserted between two consecutive anchor frames.

Since I-frames contain all the information for image reconstruction without the need of referring to other frames, the more I-frames a video stream has, the more resilient it is for transmission over wireless links. However, having more I-frames increases the video stream size. In order to save bandwidth, it is common that video streams are encoded with only one I-frame per GOP. In WAVE, the errors of video stream decoding can occur in OBUs due to packet loss or transmission error. A successfully received I-frame can correct any errors caused by any preceding frames. Therefore, I-frames are much more important than P- and B-frames. In the cases of heavy traffic load or severe packet loss, it is wise to deliver I-frames to OBUs first to enhance the quality of video playback.

Figure 2 depicts an example of video encoding pattern ("IBBPBBPBB") and video stream packetization. Encoded video frames are packetized into video packets for transportation by lower layer protocols. For instance, a video stream encoded by H.264 can be transported over real-time transport protocol (RTP) [15]. In the procedures of video packetization, the importance of a video packet can be marked as either high or low priority, so that it can be scheduled and delivered differently in lower layers. As shown in Figure 2, I-frame packets are marked as high-priority packets, whereas P- and B-frame packets are marked as low priority. The video packet priority marking will be discussed in detail in Section 3.1.

It should be noted that the sequence of video frames for transmission is different from the sequence of video playback. Taking the encoding pattern "IBBP " as an example, the actual transmission sequence is "IPBB ", that is, the P-frame is transmitted before the B-frames. This is due to the fact that decoding B-frames requires both I- and P-frames received in an ONU's buffer. In this paper, the playback deadline of a frame or packet means the latest time that a frame (or packet) has to be available in an ONU's decoding buffer.

2.2. Wave Channel Operations

Detailed description of multiple channel operations in WAVE networks can be found at [1]. In order to better understand the operations of the proposed SDS algorithm in Section 3, we summarize the typical channel operations as follows. First, upon power on, an OBU monitors the CCH until a WAVE service advertisement (WSA) sent by an RSU is received. A WSA carries the information of available SCHs and their access parameters, such as channel numbers. Based on the WSA information, the OBU then synchronizes with the RSU, and the OBU can exchange data with the RSU in SCHs.

Due to vehicles' fast movement, the topology of WAVE networks changes frequently. To maximum the time used for data exchange, WAVE devices do not use active scanning, association, and authenticating procedures as usually utilized in IEEE 802.11 [1]. WAVE defines a channel synchronization procedure which manages channel coordination. Figure 3 shows a WAVE sync interval which consists of a CCH interval, an SCH interval, and guard intervals between them. The CCH interval is used to send high-priority safety messages and WSAs, whereas the SCH interval is used for data services. The length of the CCH and SCH intervals, that is, and , can be adjusted adaptively. However, the total length of the sync interval, , is fixed to . When joining to a WAVE basic service set (WBSS), OBUs are required to synchronize with the WBSS provider, that is, an RSU, in V2I model. The synchronization ensures the OBUs to monitor CCH during CCH intervals. Video packets are delivered in SCH intervals.

The discussion above is based on single-channel WAVE devices. With a single channel, an OBU can work on either CCH or SCH at a time. If two or more channels are facilitated in a WAVE device, the operations in CCH interval and SCH interval can be conducted simultaneously. In this paper, we consider single-channel WAVE devices only as this is common for OBUs.

3.1. System Architecture

The architecture of WAVE downlink scheduling for video streaming is shown in Figure 4(a). There are two types of transport protocols used in WAVE MAC: WAVE short message protocol (WSMP) and Internet protocol (IP). WSMP is used to transport WAVE system information, such as management, control, and safety information. WAVE short message (WSM) and WSA are transported by WSMP. IP is used to convey video packets and other application data. WSMP packets can be exchanged on both CCH and SCH. However, IP packets are only allowed on the SCH. These two protocol types are identified by the EtherType field of IEEE 802.2 header. The channel router routes packets to CCH or SCH according to this field.

Inside the CCH, there is a CCH classifier which categorizes packets into four queues identified by access category index (ACI). These four queues have different traffic specifications, such as access priorities, which are defined in IEEE 802.11e EDCA. Each access category has its own channel access function and uses specified EDCA parameters. These parameters are (1) arbitration interframe space (AIFS) which defines the minimum time interval between the wireless medium becoming idle and the start of another transmission; (2) contention windows (CW) which defines a random number generation window for random collision backoff mechanism; (3) transmit opportunity (TXOP) limit which defines the maximum time duration for which a WAVE device can transmit after obtaining a TXOP. The sum of TXOPs of these four queues is a CCH interval, that is, , where is the TXOP of the queue in the CCH channel. The SCH scheduler schedules these four access category queues based on their access parameters. There is no specific scheduling algorithm defined in the WAVE standard, and the scheduling operations may differ in different implementations.

In the proposed scheduling architecture, SCH's packet queuing procedure is similar to that of CCH's. The SCH classifier is in charge of putting packets into these four access category queues. It is the same as CCH that . SCH differs from CCH in the following manners: (1) the queue (Figure 4(c)) is dedicated for video streaming packets, and it is further separated into subqueues for each OBU to store high- and low-priority video packets; (2) the SCH scheduler implements the proposed SDS algorithm (as discussed in Section 4).

Before sending video packets to a channel router, the priority of each video packet is marked by the priority marking module. Some reserved bits in video bitstream header can be used to identify the importance of a packet. Taking H.264 over RTP [15] as an example, the with values 30 and 31 are not specified in the standard [15]. As shown in Figure 4(b), we can mark high-priority video packets as type 30, and low priority video packets as type 31.

3.2. System Operations

Figure 5 illustrates the operations in an RSU. In a CCH interval, as shown in Figure 5(a), the RSU announces the video streaming services provisioned via WSA messages. The video streams are identified by provider service identifiers (PSIDs). Upon receiving the WSAs, an OBU which is interested to playback the video stream replies to the RSU to join in a WBSS. The OBU also sends its current location (coordinates from global positioning system, or GPS), direction, and speed (estimated by the OBU itself based on the GPS coordinates) to the RSU. Based on the OBU's information, the RSU estimates the time that the OBU leaves its radio coverage (or cell), that is, the out-of-cell time , where is the index of the OBU. The RSU also monitors the signal strength of each OBU and estimates a transmission rate, denoted as , for each OBU. The RSU maintains a WBSS user list which keeps each OBU's PSID, , and (as shown in the middle of Figure 5).

Before delivering packets in an SCH interval, the RSU has to decide how to use the SCH interval for all OBUs. This is where scheduling algorithms come to play. In the proposed architecture, the SDS algorithm schedules downlink packets based on a service list to specify the sequence of the OBUs to be served and which packets in the queues are granted to deliver to OBUs. As shown in Figure 5, OBU 3 will be served first and then OBU 1. There are 21 packets ( ) in the high-priority queue and 31 packet ( ) in the low-priority queue granted to be delivered to OBU 3. For OBU 1, only high-priority packets are granted to be delivered in this SCH interval. The high-priority packets of OBU 1 are not granted, and they are probably expired and thus discarded. The SDS algorithm generates or updates the service list in the RSU just before each SCH interval. The service list is indexed by OBU IDs and it is ordered by a weight factor (as discussed in Section 4 below).

After the service list is generated or updated, as shown in Figure 5(b), the RSU serves the OBUs according to this list one by one by following the order. The granted video packets are delivered within the current SCH interval. The above operational procedure repeated in each WAVE sync interval.

5.1. Simulation Setup

We have developed a WAVE network simulator to simulate WAVE networks. It simulates a two-lane straight highway of 30 km with each lane for each direction of traffic flow. In each lane, there are more than one sublane all carrying traffic to same direction. These sublanes enable vehicles overtaking each other. There are 50 RSUs uniformly distributed in the highway, each has 300 m transmission coverage radius. This makes a full coverage of the simulated highway. Simulated vehicles with mounted single-channel OBUs randomly enter the highway with exponentially distributed interarrival time at a rate of 100 vehicles per minute. Vehicle speed follows a Gaussian distribution with a mean of 100 km/h and a standard deviation of 10 km/h. The capacity of the RSU is 54 Mbps. Channel data rate varies between 1 to 54 Mbps and is a function of distance in our simulation for simplicity. The EDCA parameters used in the WAVE MAC are listed in Table 1. The contention window size is milliseconds and milliseconds.

ACI	AC			AIFSN	TXOP
1	Background			9	0.5 ms
0	Best effort			6	1.5 ms
2	Video			2	5.0 ms
3	Voice			2	5.5 ms

Since the focus of the proposed algorithm is to schedule the packets in the RSU queues, handover is not considered directly in this paper. In the simulations, we assumed that the video packets associated to an OBU are always available in its currently connected RSU. If the OBU is not associated with the RSU due to mobility, the proposed algorithm will discard all the packets related to the OBU in the RSU queues.

In the simulations, we assume that all OBUs request VOD services after they enter the simulated highway. The requested video streams are randomly selected from the aforementioned four benchmark video streams. Except for the video data ( ), other types of downlink traffics are generated by following rules: (1) voice traffic is generated randomly in the range of (32 kbps, 128 kbps) (2) self-similar BE traffic is generated randomly in the range of (0 kbps, 128 kbps) (3) background traffic is generated randomly in the range of (0 kbps, 256 kbps).

In the simulations, the proposed SDS algorithm is compared with the conventional FCFS [6], EDF [7], and MFL [5] algorithms in terms of video packet delivery ratio and video playback delay. We run the simulation 50 rounds and collect the results for analysis.

5.2. Delivery Ratio of Video Packets

More high-priority packets delivered means better video playback quality at OBUs. We measure the number of high- and low-priority packets delivered in all simulation runs with different traffic loads (from 1 to 160%). The average delivery ratio of high- and low-priority packets is shown in Figures 7 and 8, respectively.

Figure 7 shows the average delivery ratio of high-priority packets. When the total traffic load is under 80%, there is no high-priority packet loss. After the traffic load reaches a saturation point (100% of traffic load), the delivery ratio drops. The FCFS algorithm drops the fastest. This is because it does not distinguish the difference in packets. By distinguishing playback deadlines, the EDF and MFL perform better than FCFS in all video streams. Amongst EDF and MFL, MFL performs better due to the fact that it considered some specific features of vehicular networks, such as remaining dwelling time and tolerable maximum transmission delay. As EDF and MFL do not take the importance of different types of video packet into account, they treat high- and low-priority packets in the same way. The SDS algorithm outperforms the others due to the fact that the high-priority packets are preferentially treated. In all four video streams, the SDS algorithm always achieves the highest delivery ratio. Even though the overall traffic load is 160%, the SDS algorithm's average delivery ratio is still higher than 80% in all video streams. This reflects the specific feature of the SDS algorithm, that is, to deliver high-priority packets first.

Apart from the delivery ratio of high-priority packets, we are also interested to analyze the delivery ratio of low-priority packets. As shown in Figure 8, when the traffic load becomes higher than the saturation point, the delivery ratios of all algorithms drop quickly for all video streams. As can be observed, the delivery ratio of the SDS algorithm is lower than EDF and MFL in most of the cases. It is even lower than FCFS in video stream Tokyo Olympics after the traffic load is larger than 140%. This reflects the fact that the high-delivery ratio of high-priority packets is at the cost of low-delivery ratio of low-priority packets when the traffic load is very heavy.

5.3. Video Playback Delay

Apart from higher delivery ratio of high-priority packets, another main objective of the proposed SDS algorithm is to reduce video playback delay. In the situations of heavy traffic load and severe channel condition, video playback delay is unavoidable. In this simulation, video playback delay is defined as the accumulated time of the video discontinuity during the overall video stream playback. Video discontinuity is caused because the required high-priority packets (which constitute I-frames) are undelivered within their playback deadlines.

Figure 9 depicts the average video playback delay in all simulation runs. As can be observed, FCFS has the longest delay in all simulations. EDF and MFL perform better than FCFS, and MFL is slightly better than EDF. It can be also observed that video streams with a smaller packet size, such as the video stream silence of the lambs, suffer less playback delay. Since high-priority packets are always delivered first, our SDS algorithm achieves the smallest delay in all video streams. The playback delay of the SDS algorithm is nearly half of the MFL for all cases.