A Low-Latency and Error-Resilient Video-on-Demand Broadcasting Protocol Using UEP-Rateless Codes Ali Talari and Nazanin Rahnavard School of Electrical and Computer Engineering Oklahoma State University Stillwater, OK 74078 Emails: {ali.talari, nazanin.rahnavard}@okstate.edu Abstract—In this paper a new reliable and low-latency Videoon-Demand (VOD) broadcasting protocol is proposed. Most existing VOD broadcasting protocols have no protection against frame loss. In order to acquire loss-resiliency, we encode each segment of a partitioned video with an unequal error protection (UEP) rateless code. Using this error correction encoding scheme beside gaining error resiliency, the inevitable VOD broadcasting startup delay reduces significantly. Furthermore, we present a modified version of our protocol with reduced bandwidth requirement on the client side.

I. I NTRODUCTION In Video-on-Demand (VOD) broadcasting, customers can choose from a list of videos at any time, and watch them after a short delay. VOD broadcasting requires a protocol that minimizes the shared video streaming costs such as required bandwidth and startup delay. Periodic broadcasting protocols ? are well-known VOD protocols which partition a video into segments, usually with increasing size, and broadcast each segment periodically in an individual logical channel. However, these protocols are not robust against channel loss. This might result in video quality degradation to a large extent in lossy channels. To overcome the channel loss, some protocols proposed to protect the segmented video with forward-error-correction (FEC) codes before transmission ?. In this paper, we propose a new VOD broadcasting protocol, which encodes a segmented video with recently proposed unequal error protection (UEP) rateless codes ??. Beside the inherent loss resiliency of the encoded video, the startup delay of the proposed protocol improves significantly compared to the case that video segments are encoded with equal error protection (EEP) rateless codes ?. The paper is organized as follows. Section II presents a brief introduction to UEP-rateless codes. In Section III, we propose a new VOD broadcasting protocol based on UEPrateless codes. Moreover, we propose a modified version of our protocol for the case in which clients have limited bandwidth. In Section IV, our proposed protocols’ performances are evaluated. Finally, Section V concludes the paper.

II. BACKGROUND ON U NEQUAL E RROR P ROTECTION R ATELESS C ODES Rateless codes ?? are a class of modern FEC codes with low computational complexity and close to capacity performance. They can potentially encode n source data packets into unlimited number of outgoing (encoded) packets. Luby proposed a group of rateless codes called LT codes ?. In LT encoding, first a packet degree, d, is chosen from a probability distribution {Ω1 , Ω2 , . . . , Ωn }, where Ωi is the probability that d = i. We may also denote thisP probability n distribution by its generator polynomial Ω(x) = i=1 Ωi xi . Next, d packets are chosen uniformly at random from source packets, and they are XORed together to form an encoded packet. During the transmission, some of the encoded packets are lost; however, regardless of which packets are lost, the client can recover the source packets when γn encoded packets are received, where γ ≥ 1 is called coding overhead. Raptor Codes ? were invented soon after LT codes. The Raptor codes are similar to LT codes including an extra precoding phase. First, the source packets are encoded with a high-rate code (such as an LDPC code), and then the precoded packets are encoded using an LT code with carefully designed Ω(x). Initially, all studied rateless codes provided equal error protection (EEP) of the entire source data. The EEP property would be sufficient for applications similar to multicasting bulk data (e.g., a software file). However, in some applications a portion of data may need more protection than the rest of data, or a portion of data may need to be recovered prior to the other parts. In ?? authors proposed a generalization of rateless codes that provides unequal error protection and unequal recovery time properties. The idea used in the generalized rateless codes is to modify the source packets selection from uniform to nonuniform. The message of n packets is partitioned into k parts. Next, pj , j ∈ {1, 2, . . . , k} is assigned to part j, where pj is the probability that a source packet from part j is chosen to build an encoded packet. In this way, more

important packets are assigned a higher selection probability, and therefore, they are included in more encoded packets. As a result, more important packets can be decoded with a lower error rate or equivalently with a shorter decoding latency. In a simple case, the source data is divided into two parts with higher and lower priorities. Here we recall the twopriority-level bit error rate (BER) formulas of generalized rateless codes derived in ?. Assume n1 = αn (0 < α < 1) is the number of more important bits (MIB), and n2 = (1−α)n is the number of less important bits (LIB). We set p1 = knM L . and p2 = knL for some 0 < kL < 1 and kM = 1−(1−α)k α Let yl,M and yl,L denote the BER of MIB and LIB at the lth decoding iteration, respectively. We have yl,M = e−kM µγβ(1−(1−α)kL yl−1,L −αkM yl−1,M ) ,

(1)

yl,L = e−kL µγβ(1−(1−α)kL yl−1,L −αkM yl−1,M ) ,

(2)

with β(x) = Ω′ (x)/Ω′ (1), µ = Ω′ (1), and y0,L = y0,M = 1. It can be shown that sequences {yl,M }l and {yl,L }l are convergent with respect to the number of decoding iteration l ? . Let yM and yL denote the corresponding fixed points, respectively. For a given overhead γ, we have yM < yL . We could also fix the target BER of MIB and LIB and compare γMIB and γLIB , which are the overheads needed for MIB and LIB to reach the target BER, respectively. We have γMIB < γLIB . This means that BER of MIB reaches a target BER faster (smaller overhead) than the BER of LIB. In our proposed VOD protocol, we take advantage of the faster recovery of MIB when applying UEP-rateless codes to improve the startup delay experienced by clients. Throughout this paper, we consider the following degree distribution Ω(x) derived in ? Ω(x) = 0.007969x + 0.493570x2 + 0.166220x3 + 0.072646x4 + 0.082558x5 + 0.056058x8 + 0.037229x9 + 0.055590x19 + 0.025023x64 + 0.003135x66. III. E RROR -R ESILIENT L OW-L ATENCY VOD P ROTOCOL In this section, we introduce our low-latency and lossresilient VOD protocol that employs UEP-rateless codes. We also present a variation of our scheme that is suitable when clients have low bandwidth. A. Protocol Development To Start our protocol design, we assume that a client does not perceive any degradation in the video quality at the frame loss rate of 10−3 ?. This is a very pessimistic assumption, and viewers’ satisfaction is definitely ensured. A video is first divided into several segments of increasing length (later we show how to determine the length of each segment). Let r denote the transmission rate of each

segment, and B0 denote the video playout rate. Without loss of generality, we consider the first segment of the partitioned video as the reference segment with playout duration of one unit time at playout bandwidth of B0 . By this assumption, if no coding is employed, the size of data contained in the first segment is 1 ∗ B0 , and the download time of the reference segment is Br0 unit times. However, if we encode the reference segment with a rateless code, we also need to consider the coding overhead, γ, in deriving the download time. Let us first consider the case that the reference segment is encoded with an EEP-rateless code. The download time of this segment is γEEP Br0 unit times, where γEEP is the overhead at which the loss rate of the decoded segment reduces to 10−3 frames. In our proposed protocol, we partition the reference segment into two parts with the fraction sizes α and 1 − α, and protect the first part with a higher priority, kM , and the rest of the segment with a lower priority kL < kM employing a UEP-rateless code ?. The first part (MIB) acquires the error rate of 10−3 at γMIB , and is displayed to the client, while more encoded packets from the same segment are being received. Meanwhile, the second part (LIB) is also recovered at γLIB . Other segments are similarly divided into two parts with relative sizes α and 1 − α. The basic rule in VOD broadcasting protocols is that each part and segment must be received (in our protocol, decoded) completely before its playout time so that the video is shown continuously with no interruptions. We need to choose the UEP priorities, the size of each section (equivalently α), and segment sizes accurately to meet these timing constraints. All the overheads for EEP- and UEP-rateless coding are found using iterative decoding of rateless codes from (1) and (2). The startup delay of the proposed protocol with UEPrateless coding is affected by two recovery times γMIB ∗ Br0 and γLIB ∗ Br0 . As a result, we have to carefully determine the startup delay in a way that no interruption occurs between two partitions playout. Determining the startup delay, we encounter two cases. First, when the beginning part’s recovery time plus the playout time of this portion is greater than the recovery time of the second part. This condition is shown in Figure 1(a). In this case, the startup delay is dominated and determined by the first part’s recovery time. The video starts playing upon recovery of the first portion. The second part is recovered on time, during the first part’s playout, and before its playout time. In the second case, the second part is not recovered during the first segment’s playout, thus it will not be ready on time as shown in Figure 1(b). In this case we must extend the startup delay so that the second part is recovered on time and the video can be displayed without any interruptions. This is equal to decoding and storing the first portion in a buffer, and starting the video display when we are confident about second part’s

Į

w

1-Į

ȖMIB*B0/r ȖLIB*B0/r

Į

0.4 0.39 0.38 0.37 0.36 0.35 0.34 0

(a) First part (MIB) recovery-time dominates.

w

0.41 Normalized startup delay, w

on-time recovery. In this case the startup delay is controlled by the second part’s recovery time. Note that γMIB and γLIB are determined by plugging the values of kM , kL , and α into (1) and (2), and finding the values of γMIB and γLIB for which BERs yM and yL are equal to 10−3 .

1-Į

0.2 α

0.4

1.5

2

M

ȖMIB*B0/r ȖLIB*B0/r

Fig. 2.

(b) Second part (LIB) recovery-time dominates. Fig. 1. First segment and startup delay, w, which is a function of γM IB , γLIB , B0 , r, and α.

According to Figure 1 and the discussion provided, we can derive a formula for the startup delay of each segment in terms of the recovery overheads of its partitions and the transmitting bandwidth of the segment. The normalized startup delay formula is clearly given by w = max(γMIB

B0 B0 ∗ , γLIB ∗ − α). r r

(3)

For the EEP-rateless coding case, α would be equal to zero and (3) reduces to w = γEEP ∗

B0 . r

(4)

Our goal is to choose kM and α in (1) and (2) such that the startup delay of the VOD protocol is minimized. Figure 2 depicts the startup delay, w, for different values of kM and α when r = 3B0 . As is shown, w has a global minimum value at kM = 1.56 and α = 0.10. Table I summarizes the optimal values of kM , α, and the corresponding γMIB , γLIB , and minimized startup delays for r = 2B0 to r = 5B0 . We use these optimum values in our simulations and protocol design. TABLE I O PTIMUM VALUES OF kM , α, AND THE CORRESPONDING γM IB , γLIB , AND MINIMIZED STARTUP DELAYS , w, WITH UEP- RATELESS CODING . r α kM γM IB γLIB w

2B0 0.14 1.43 1.032 1.296 0.516

3B0 0.10 1.56 1.031 1.287 0.343

4B0 0.06 1.75 1.030 1.270 0.257

5B0 0.05 1.83 1.030 1.264 0.206

1 k

Normalize startup delay, w, vs. kM and α when r = 3B0 .

Equivalently, the normalized startup delays for r = 2B0 to r = 5B0 , with EEP-rateless coding are given in Table II. Note that for all the cases we have γEEP = 1.211. TABLE II N ORMALIZED STARTUP DELAY, w, WITH EEP- RATELESS CODING . r w

2B0 0.605

3B0 0.403

4B0 0.302

5B0 0.242

Now, we can derive the formula to find the sizes of the segments in our proposed VOD protocol. We determine the second segment size with the criteria that summation of the startup delay, w, and the duration of the first segment playout time, is the startup delay for the second segment. The division of this time duration to w, gives the second segment’s length. This rule applies to other segments as well. The general segmentation formula is given by

Si =

(

w+

P1i−1

j=1

w

i = 1, Sj

i ≥ 2,

(5)

where Si is the normalized ith segment size, and the startup delay, w, for UEP-rateless and EEP-rateless coding is given by (3) and (4), respectively. The same two-priority-level UEP-rateless code is applied on all segments, therefore, all segments include a portion which is recovered in advance, and is displayed to the client before the second part of the segment is recovered. The proposed broadcasting protocol segmentation is depicted in Figure 3, when the number of segments, S, is equal to 4. S is chosen according to the system requirements and the available system resources. Although a larger S results in a shorter startup delay, it requires a higher total video broadcast bandwidth, rS, and also causes heavier computational complexity.

S1

B0

S2

S3

S4

B0

w

r r r r

S1

S2

S3

S4

w

S1

r

S2

r S3

r S4

Fig. 3.

r

Proposed protocol segmentation.

Since UEP/EEP-rateless encoded packets are broadcast continuously, there are not any segmentation boundaries on the broadcasting channels in Figure 3. Whenever a client tunes into the video stream, he can start buffering the video segments. This is an advantage of VOD protocols with UEP/EEP-rateless encoding since in conventional periodic broadcasting protocols, if a client tunes into the video stream in the middle of the first segment broadcast, he has to wait until the first segment is rebroadcast to start buffering. As we discussed earlier, our protocol is also loss resilient. We can easily modify this protocol so that, no matter what the value of the channel loss rate is, we can have a continuous playout. If the aggregate channel loss rate equals p, by increasing the server transmission’s rate with the factor a, where a = 1/(1 − p), we can adequately compensate the loss incurred by the channel. The client still receives enough required number of packets for successful decoding. Video server should monitor the channel to choose the appropriate a, or it can always transmit the video stream with the highest predefined a, which is a characteristic of the channel. Moreover, VOD broadcasting protocols with rateless encoding, maintain the error rate of the video in an acceptable range with a reasonable overhead. This will provide quality of service to users. When the video is transmitted with conventional VOD protocols with no forward-error-correction, the quality of the video is affected by channel loss, and as a result, no minimal quality can be guaranteed for the clients. Especially, when the error rate of the channel is high, uncoded VOD broadcasting protocols are unusable. B. Modified Low Bandwidth Protocol So far, in our proposed protocol, we have assumed that server and clients have the same bandwidth equal to rS. However, in some cases clients might have limited amount of bandwidth due to the technical limitations or the bandwidth cost. Our proposed protocol can be easily modified such that each client has to receive only two segments in parallel. This reduces clients’ required bandwidth from rS to 2r. Decreasing the client’s bandwidth leads to an increase in the startup delay due to new segmentation size requirement; however, it does not affect the UEP-rateless codes’ performance. We still encode the segments with the same UEPrateless code, and compare the resulting startup delay with EEP-rateless encoding. In Figure 4 the segmentation of the modified protocol for lower client bandwidth is illustrated. The startup delay is still given by (3) and (4), and the

S1 S2 S3 S4

Fig. 4. Segmentation in our modified protocol. Users download only from two channels in parallel at each time.

segmentation size of the modified version of the protocol, which satisfies a continuous playout, is given by  1 i = 1,  w+1 i = 2, (6) Si =  Si−1 w +Si−2 i ≥ 3. w

The segmentation sizes for both proposed protocol schemes, using optimum values from Table I and II, with EEP- and UEP-rateless encoding, is depicted in Figure 5. IV. P ERFORMANCE E VALUATION

In order to evaluate our proposed VOD protocols, we compare the startup delay of our scheme with UEP-rateless coding to startup delay of the case where video is encoded with an EEP-rateless code. We compare the two protocols for different bandwidths allocated per individual channel, for r = 2B0 to r = 5B0 , and when number of segments varies from 2 to 8 segments. In Figure 6 the percentage of the reduction made in the startup delay of our proposed protocol is illustrated. From Figure 6, we can see a significant improvement in the systems performance for the same bandwidth. For example, if UEP-rateless codes are used instead of EEP-rateless codes, when the video is partitioned into six segments and r = 5B0 , the startup delay decreases about 55%. Similarly, we compare the modified VOD protocol with the segmentation given by (6) in two cases of EEP-rateless and UEP-rateless coding. Figure 7 depicts the percentage by which the startup delay declines for different bandwidths and different number of segments, when a UEP-rateless code is employed instead of an EEP-rateless code. From Figures 6 and 7, it can be seen that our proposed protocol and its modified version that are based on UEPrateless coding outperform the case where EEP-rateless coding is employed. We also note that as the bandwidth allocated to each stream increases, the efficiency of the proposed protocols also increases. An increase in the performance can also be observed when the number of segments increases. V. C ONCLUSION In this paper we proposed a new VOD broadcasting protocol with unique features of error resiliency and low startup delay. These features were acquired by dividing the video segments into two partitions, and encoding each segment with a UEP-rateless code. Rateless codes get very close to

70 UEP−rateless coding EEP−rateless coding

200

Startup delay reduction percentage (%)

Normalized segment size (Si)

250

150

100

50

0

Bandwidth = 2*B

65

0

Bandwidth = 3*B

0

60

Bandwidth = 4*B

0

55

Bandwidth = 5*B

0

50 45 40 35 30 25 20

1

2

3 4 Segment number

5

6

15

2

3

100

50

2

3 4 Segment number

6

7

8

70

5

6

(b) Modified VOD protocol segment sizes, when the bandwidth of server is equal to rS and the bandwidth of clients is equal to 2r.

Startup delay reduction percentage (%)

Normalized segment size (Si)

UEP−rateless coding EEP−rateless coding

1

5

Fig. 6. Percentage of reduction made in the startup delay of our original proposed VOD protocol with UEP-rateless coding compared to the case where EEP-rateless coding is employed.

150

0

4

Number of segments (S)

(a) Proposed VOD protocol segment sizes, when the bandwidth of server and the bandwidth of clients are equal to rS.

Bandwidth = 2*B

65

0

Bandwidth = 3*B

0

60

Bandwidth = 4*B

0

55

Bandwidth = 5*B

0

50 45 40 35 30 25 20

Fig. 5.

Segmentation sizes for the proposed protocols.

15

2

3

4

5

6

7

8

Number of segments (S)

channel capacity, and have low encoding and decoding computational complexity, which makes the proposed protocol feasible in practice. We also showed that our proposed VOD scheme can easily be modified for the case that clients have a lower bandwidth than the server. Simulation results show that our VOD broadcasting protocols with UEP-rateless coding outperforms the case where EEP-rateless coding is employed. R EFERENCES A. Hu, “Video-on-demand broadcasting protocols: a comprehensive study,” in Proc. of IEEE INFOCOM, Apr. 2001., vol. 1, pp. 508–517. A. Mahanti, D. Eager, M. Vernon, and D. Sundaram-Stukel, “Scalable on-demand media streaming with packet loss recovery,” in Proc. of SIGCOMM’2001, p. 12, 2001. N. Rahnavard, B. Vellambi, and F. Fekri, “Rateless codes with unequal

Fig. 7. Percentage of reduction made in the startup delay of our modified proposed VOD protocol with UEP-rateless coding compared to the case where EEP-rateless coding is employed.

error protection property,” IEEE Transactions on Information Theory, vol. 53, pp. 1521–1532, April 2007. N. Rahnavard and F. Fekri, “Generalization of rateless codes for unequal error protection and recovery time: Asymptotic analysis,” IEEE International Symposium on Information Theory, pp. 523–527, July 2006. M. Luby, “LT codes,” in The 43rd Annual IEEE Symposium on Foundations of Computer Science, pp. 271–282, 2002. A. Shokrollahi, “Raptor codes,” IEEE Transactions on Information Theory, vol. 52, pp. 2551–2567, June 2006. D. Wijesekera, J. Srivastava, A. Nerode, and M. Foresti, “Experimental evaluation of loss perception in continuous media,” Multimedia Systems, vol. 7, no. 6, pp. 486–499, 1999.