Improving Performance for Streaming Video Services over CDMA-Based Wireless Networks

Transcription

1 Improving Performance for Streaming Video Services over CDMA-Based Wireless Networks Ernestina Cianca, Frank H.P. Fitzek, Mauro De Sanctis, Marco Bonanno, Ramjee Prasad, Marina Ruggieri ( )University of Rome Tor Vergata Department of Electronics Engineering Via politecnico 1, Rome, Italy ( )Department of Communications Technology, Aalborg University Neils Jernes Vej 12, 9220 Aalborg Øst, Denmark cianca@nausicaa.eln.uniroma2.it, ff@kom.auc.dk de.sanctis@ing.uniroma2.it, prasad@kom.auc.dk, ruggieri@uniroma2.it, Abstract In video streaming services, the playout begins when the queue length of the receiver buffer is above a threshold. This threshold must be large to reduce the buffer underflow probability and absorb the bit rate variations caused by the wireless channel. On the other hand, it is important to reduce this threshold in order to reduce the initial playout delay and also the size of the receiver buffer. In this paper, a video streaming service is considered, where the last link is a wireless CDMA-based link and it has been shown how a truncated power control allows to reduce the pre-roll delay without increasing the average transmission power and without degrading the video quality. The paper presents the proposed truncated power control and an analytical model to evaluate the achievable pre-roll delay reduction. Also the advantages in terms of video quality are shown. Keywords ARQ mechanism, code division multiple access, truncated power control, video quality, wireless video streaming. I. INTRODUCTION AND MOTIVATION Market research finds that mobile commerce for 3G wireless systems and beyond will be dominated by basic human communication such as messaging, voice, and video communication [1]. Because of its typically large bandwidth requirements, video communication (as opposed to the lower rate voice and the elastic ) is expected to emerge as the dominant type of service in 3G/4G wireless systems [2]. Video services are typically divided into real-time services and streaming. Our focus in this paper is on video streaming where the client may tolerate a small start up delay before the play-out commences. Video streaming schemes typically rely on the User Datagram Protocol (UDP) as the transport protocol. The use of the Transmission Control Protocol (TCP) as proposed in [3] has also some advantages, but is not addressed here. Using UDP implies that there is no reliability mechanism at transport level. The reliability of the communication over wireless channel is handled at lower layers through, for instance, the implementation of ARQ (Automatic Repeat request) schemes at the data link layer. ARQ increases the reliability of the channel but it also introduces a longer and more variable delay at link layer. In order to limit the delay, a maximum number of retransmissions, called persistency of the ARQ protocol, are allowed. Moreover, the jitter can be combated by the introduction of a receiver buffer in combination with an initial playback delay, pre-roll delay, to smooth the bit rate variations caused by the transmission channel [4]. Designers of today s commercial media streaming products find that buffering delays ranging from 5-15s typically strike a good balance between delay and playout reliability (low probability of playout interruption due to buffer underflow) [5]. In [6], the minimum initial delay for a deterministic receiver curve for a Variable Bit Rate (VBR) channel has been defined. In [7], an adaptive media playout is proposed to allow the client to buffer less data and thus reducing the pre-roll delay for a given buffer underflow probability over a wireless channel. However, those works either neglect the delay introduced by retransmissions or assume ARQ protocols with infinite persistency and hence no Fig. 1. System model of the video streaming. residual Frame Error Rate (FER) after retransmissions. In this paper, we first show that the persistency of the ARQ protocol can greatly impact the pre-roll delay, even for a Selective Repeat (SR) ARQ protocol. Therefore, in order to decrease the pre-roll delay, low persistency ARQ protocols should be used, thus resulting in less reliability. We then apply to the context of video streaming over CDMA-based wireless networks, a truncated power control that has been already shown to be efficient in reducing the average delay introduced by retransmissions at link layer for a given guaranteed reliability and power consumption [8]. Its exploitation results in two main benefits: the pre-roll delay is reduced without degrading the video quality and without increasing the required transmit power; once the pre-roll delay has been designed, better video quality can be achieved without increasing the buffer underflow probability (always at the same transmission power) In the rest of the paper, these two benefits will be shown. The remainder of the paper is organized as follows. Section II provides the system model under study. In Section III we introduce our power control policy. The performance of the proposed approach are shown in Section IV. Finally conclusions are drawn in Section V. II. SYSTEM MODEL The system model consists of a source, a server and a client. As shown in Figure 1, the video streaming passes through a wired network without losses and a lossy wireless link with CDMA-based transmission. The source can be a live program or a prestored program; in the first case the source passes a video frame to the server

2 every t F seconds, while in the second case the source passes all of the video frames to the server at the beginning of the session. The server is responsible of delivering video frames from the source to the client through a heterogeneous wired/wireless network; in this paper we refer to a UDP-based transport platform. The UDP protocol does not permit to recover from data losses and this functionality is left to the link layer (see Section II-C). The server encapsulates each video frame within a UDP packet and each packet is enqueued into the UDP transmission buffer. Both at the server side and at the client side a link layer ARQ buffer and a playout/udp buffer are needed. At the client side, the playout begins when the queue length n of the playout buffer is above a specified threshold N pr. Such a phase is called pre-roll process and it is needed in order to reduce the buffer underflow probability at the expenses of an initial delay (pre-roll delay). At the client side, underflow occurs when n = 0; after the buffer underflow occurrence the receiver temporarily suspends the playout of the video and a new pre-roll process starts. Both the pre-roll delay and the buffer underflow probability depends on the pre-roll threshold N pr and on the channel reliability. Large N pr results in a small underflow probability but increase the pre-roll delay. It is common for streaming media clients to have a 5 to 15 seconds of buffering delay before playback starts. It is worth noting that the size of the receiver buffer is a crucial point for wireless terminal design. To lower the terminal costs small buffers would be desirable, but the characteristics of the wireless channel requires large buffers to overcome the instability in the serving rate of the buffer. The serving rate is variable because information can get lost and may have to be retransmitted. There are two main ways to address this problem. The first would be to live with the need of larger buffers, the second is to find solutions in the protocol design to overcome the instability of the wireless medium using error control mechanisms such as ARQ mechanisms. While video quality of a given video encoding standard depends on the images reliability and hence on the packet error rate, the customer satisfaction also depends on the initial pre-roll delay and on the buffer underflow probability. Taking into account that in our system the transport protocol is unreliable (i.e. assuming UDP), the packet error rate depends on the FER at link layer; for such a systems, a partial reliability of the transmission is gained by using the ARQ protocol at the link layer. The exploitation of ARQ protocols increases the preroll delay that is needed to get a certain upper bound on the buffer underflow probability. Packets from the playout buffer are removed at the playout rate µ(n) and after decoding they are displayed. In [9] an adaptive control of the playout speed of media is proposed in order to reduce the pre-roll delay for a given buffer underflow probability. In this paper, we assume a non-adaptive media playout for which the playout rate is constant. We then propose the use of a truncated power control to reduce the pre-roll delay for a given video quality and buffer underflow probability. A. Physical Layer In the system model we have considered a lossy wireless link with a CDMA-based data transmission. A SNIR-based (Signal to Noise plus Interference Ratio) power control is assumed where the transmission power is updated each T p seconds, in order to get a target average SNIR. The transmission is slotted in time intervals of duration T p. The transmitter uses a power P h to transmit in the time slot h; the Multiple Access Interference (MAI) seen at the receiver is modelled as a zero mean Gaussian r.v. with variance σ 2 I during the time slot h. SNIR req(p e) is the target SNIR and p e is the corresponding FER (Frame Error Rate) when a frame length of m time slots is considered. B. Channel Model A non frequency-selective multipath fading is considered. The effect of fading is then described as a multiplicative complex function whose envelope is assumed to have a Rayleigh distribution. Therefore, by denoting with g(h) the squared envelope of the complex function describing the fast fading effect, with A 0(h) its mean at the beginning of the time slot h, g(h) is assumed to have the following p.d.f.: p L(g) = 1 g(h) A 0(h) exp A 0 (h). (1) C. Radio link layer ARQ protocol A radio link protocol with negative acknowledgments such as the (NACK)-based selective repeat ARQ error recovery mechanism [10] is considered. The persistency of the protocol is denoted by n max, which represents the maximum number of allowed retransmissions. The UDP packet is divided into N Radio Link Layer (RLL) frames, each m time slots long. The receiver requests the retransmission of data frames that are not received correctly. When the receiver finds a frame in error it sends back a NACK requesting for retransmission and a timer is set. After a retransmission timeout t RT O, the timer expires and if the lost frame has not been received yet, the receiver sends back another NACK and another timer starts. After a number n max of timer expirations the radio link layer aborts the attempt to pass a correct frame to its upper layer and passes whatever it has got. Both the residual FER of the ARQ scheme and the average link layer transmission delay depends from the FER at physical layer (we mean, the FER without retransmissions), here denoted with p e, and from n max. Assuming no errors in the reception of the acknowledgments, the probability of correct reception P c can be written as: P c = (1 p e) + p e(1 p e) n i=1 III. TRUNCATED POWER CONTROL p (i 1) e. (2) The transmission power at the time slot h is denoted by P h. The following transmission policy is adopted [8]: to transmit if P h is lower than a power threshold P ; to idle otherwise. We expect it to idle during deep fade or high interference level, waiting to transmit when the channel attenuation and/or the interference level is lower. Therefore, this truncated power control introduces a delay in the transmission of a frame. In order to evaluate its average value, the following model for the transmit/idle state process has been used. A. Transmit/Idle state process According to the channel model (Section II-B), the idle state is only associated to deep fades due to multipath. Therefore, the power threshold introduced by the truncated power control represents an upper bound on the extra power needed to compensate the deep fades due to multipath, assuming that the shadowing and path loss can always be compensated. Furthermore, the power is updated each T p seconds. During a T p long time slot the transmit power is constant. The binary process that describes transmit/idle

3 state is obtained by comparing P h with P in each time slot and hence, by comparing g(h)/a 0 with the threshold 1/F = SNIR 2 0 (I + N)/P. Let us denote with d the average duration of an idle-state (in number of time slots). According to the hypotheses, d depends on the correlation of the channel fading process. Therefore, the statistics of the idle state process can be estimated by the statistics of the fade duration [11]. In [12] it was shown that a Rayleigh fading envelope can be well approximated by a first-order Markov process with continuous amplitude. Therefore, the transmit/idle state process can be modelled by a two-state Markov model with transition matrix: ( ) p 1 p M c = (3) 1 q q where p and 1 q are the probabilities that in the i-th time slot the transmitter is in a transmit state, given that the (i 1)-th time slot was in a transmit and in a idle state, respectively. F denotes the so called fade margin [11]. The steady-state probability that the transmitter be in an idle state P I = P rob(p n P ) is [13] - [14]: P I = [ 1 e 1/F ]. (4) Note that (1 q) 1 represents the average length of an idle state (in number of time slot T p) and, hence will be denoted with d. For a Rayleigh multipath fading process, the correlation in time is well characterized by the Doppler frequency, here denoted with f D. Therefore, the Markov parameter q can be derived as in [13] - [14]: with: q = 1 Q(θ, ρθ) Q(ρθ, θ) e 1/F 1 θ = (5) 2/F 1 ρ 2, (6) where ρ is the Gaussian correlation coefficient of two successive samples of the channel gain. In a fading channel with Doppler frequency f D, this coefficient can be written as: ρ = J 0(2πf DT p) (7) where J 0( ) is the Bessel function of the first type and zero order. In equation (5), Q(, ) is the Marcum Q function given by: Q(x, y) = y e (x2 +w 2 ) 2 I 0(xw)wdw. (8) where I 0 is the modified Bessel function of the first type and of zero -th order. Note that the correlation properties of the fading process depend only on f D. Let us denote with d the average value of the delay for transmitting a frame of m time slots, in presence of truncated power control. Following the procedure described in [8], the delay d can be written as: d = P Id T p. (9) This delay can be seen as a longer bit interval T = T s(1 + d/(mt p)) and, hence, a lower effective data rate R = 1/T. However, in spite of reduced effective data rate, the truncated power control allows to get a higher channel reliability without increasing the transmission power and, hence, the overall throughput at Radio Link Layer could be lower. IV. PERFORMANCE EVALUATION A. Pre-roll delay vs Persistency The aim of this Section is to show the impact of ARQ persistency on the dimensioning of the pre-roll delay with and without truncated power control. The video stream is characterized by a duration τ and a sampling curve p(t) which is defined as the overall amount of data (e.g., measured in bits) produced by the video encoder up to the time t. In case of Constant Bit Rate (CBR) video stream p(t)=r * t, where R denotes the transmission bit rate of the transport channel. To avoid buffer underflow, which occurs in case that not enough data is present at the decoder at the time a certain video frame has to be decoded and displayed, the pre-roll delay must be properly designed. Even if CBR is considered, the error-prone wireless channel with ARQ protocol results in a VBR reception. Let us denote with r(t) the receive curve, which specifies the total amount of data received error-free up to the time t at the receiver. Obviously, r(t) is monotically increasing. Moreover, note that we are not considering an infinite-persistent ARQ protocol, therefore, the received video is characterized by r(t) and also a residual FER, or probability of correct reception P c that is given by (2). To avoid buffer underflow at the receiver buffer, the preroll delay has to be chosen such that for any instant t, at least p(t ) bits are available at the decoder, i.e.: R : r(t) p(t ), t. (10) Let us define the pseudo-inverse function of the monotonically increasing sampling curve p(t) as p ( 1) (x) := min{t : x p(t)}. As it is shown in [6], the minimum pre-roll delay to avoid buffer underflow and the decoder buffer size is chosen as: = max{t p ( 1) (r(t))} (11) t In order to evaluate from Eq. (11), we need to model r(t), which depends on the channel and the ARQ protocol. Explicit modelling of a retransmission protocol in the context of Markov chain based analysis is challenging. Moreover, the aim of this Section is not to provide accurate dimensioning of but rather to estimate the impact of ARQ persistency. With this aim we can assume that retransmissions of lost packets are modelled as a reduction of the throughput. Therefore, if R denotes the original bit rate, the bit rate at the receiver is R/A, where A is the mean number of frame transmissions and can be computed as: n max+1 A(n max, p e) = (1 p e) i p e (i 1) +p (nmax+1) e (n max +1) i=1 (12) In (12), no errors in the reception of the acknowledgments are assumed and the channel error process is a memoryless process. The latter assumption hold if a perfect power control in the CDMA system is assumed. With the truncated power control the new data rate is R = R /A, where R has been defined in Section III-B and A is the mean number of frame transmissions of the SR scheme with persistency n max and with the same average transmission power as the case without truncated power control. If A < A, the lower effective data rate with truncated power control could be compensated, thus giving a better overall throughput. Therefore, for our purposes is computed as if r(t) is deterministic and achieved by p(t) with a lower data rate. In the following results a video stream of duration τ = 60s is considered. In Table 1, the pre-roll delay has been evaluated for a channel BER=0.001, with and without truncated power control. Moreover, different value of the persistency have been considered. An acceptable value for the residual packet error rate is usually

4 Tab. 1. Parameters of the transmission for a channel BER= n max = 2 n max = 1 without TrPC = 24s = 19s P c = 0.89 P c = 0.77 with TrPC = 11s = 13s P c = 0.98 P c = 0.98 assumed to be 0.1. This means that the probability of correct reception should be P c = 0.9. Without truncated power control, to get this acceptable value of P c, a persistency of n max = 2 is needed. This results in a pre-roll delay = 24s that is far higher than the range of acceptable values (5-15s). The truncated power control allows to reduce the pre-roll delay to = 11s, that is now acceptable. Moreover, a higher P c is achieved, thus also the video quality has been improved. In the same Table, the case n max = 1 is also considered. Again, the truncated power control allows to reduce the delay and get a higher quality. B. Video Quality We now assume a pre-roll delay of 11s. Figures 2 and 3 compares the quality of two video with the same average reduction of the throughput due to the retransmission and truncated power control (TrPC). The original video is encoded with the encoding standard H.263. Note that the video sequences have been intentionally encoded at low bit rate (102kbps) in order to better highlight the effect of transmission errors. Also the encoded video sequences are shown. The Figures clearly shows an improvement of the quality that is achieved without increasing the power consumption. The Peak Signal to Noise Ratio (PSNR) has been used as a measure of the video quality. The average PSNR with truncated power control is around 20dB (Fig. 2) and without truncated power control is 17dB (Fig. 3). However, the pre-roll delay is very high due to the high channel BER. When a lower value of the pre-roll delay is considered ( = 13s), as in Figures 4 and 5, the improvement of the quality achieved with the truncated power control is more evident. The difference on the average PSNR is now around 7dB. Moreover, by comparing Figures 2 and 4, we can conclude that the system with the truncated power control is less sensitive to the persistency of the ARQ protocol. Therefore, a reduction of about 10s in the pre-roll delay can be achieved at expense of only a slight degradation of the quality. In fact, the difference in PSNR between Figures 2 and 4 is around 1.5 db. V. CONCLUSION This paper proposes a truncated power control to improve the performance of video streaming services over CDMA-based wireless links in terms of reduction of the pre-roll delay and improved video quality. First of all, the impact of the persistency of the ARQ error control protocol on the value of the minimum preroll delay that is needed to get a given buffer underflow provability has been evaluated. Moreover, it has been shown that on a channel with a BER=0.001, a reduction of the pre-roll delay from 24s to 11s can be achieved by using the proposed truncated power control. Moreover, once the pre-roll delay has been designed, the truncated power control allows to increase the video quality. An improvement of the PSNR of 7dB has been found. REFERENCES [1] D. Bose. M-commerce for 3G. Presentation at 11th Time to Market Symposium, Sky Garden, Sony Center Berlin, Germany, September [2] L. Roberts and M. Tarsala. Inktomi goes wireless; forms alliances. In CBS MarketWatch, March 14th [3] F. H. P. Fitzek,R. Supatrio, A. Wolisz, M. Krishnam, M. Reisslen, Improving QoS for QoS Streaming Video Applications over UDP in CDMA-Based Network. Fig. 2. Comparison of the original image with the encoded one and the received one with P c = 0.992, achieved with n max = 2 and with truncated power control. Pre-roll delay of = 24s. Fig. 3. Comparison of the original image with the encoded one and the received one with P c = 0.9 with n max = 2, without truncated power control. Pre-roll delay of = 24s. [4] V. Varsa, I. Curcio, Transparent end-to-end packet switched streaming service (PSS); RTP usage model (Release 5), 3GPP Tr V1.4.0, [5] Window Media Player [online]. [6] T. Stockhammer, H. Jenkac, Streaming Video over Variable Bit-Rate wireless channels, IEEE Trans. on Multimedia, vol.6, no. 2, Apr [7] M. Kalamn, E. Steinbach, B. Girod, Adaptive media playout for low-delay video streaming over error-prone channels, IEEE Trans. on Circuits and Systems for Video Tech., Vol. 14, No. 6, June [8] E. Cianca, M. De Sanctis, M. Ruggieri, R. Prasad, Truncated Power Control for Improving TCP/IP performance over CDMA wireless links, to be published on IEEE Trans. on Wireless Communications. [9] M. Kalman, E. Steinback, B. Girod, Adaptive Media Playout for Low- Delay Video Streaming Over Error-Prone Channels, IEEE Transsactions on Circuits and Systems for Video Technology, vol. 14, no. 6, June [10] S. Lin, D. J. Costello, M. J. Miller, Automatic-Repeat-Request Error Control Schemes, IEEE Communication Magazine, vol. 22, no. 12, pp. 5-17, 1994

5 Fig. 4. Comparison of the decoded image with the original transmitted image with P c = 0.99 with n max = 1 and with truncated power control. Pre-roll delay of = 13s. [11] Li Fu Chang, Throughput Estimation of ARQ Protocols for a Rayleigh Fading Channel Using Fade- and Interfade-Duration Statistics, IEEE Transactions Vehicular Technology, vol. 40, no. 1, pp , February [12] H. S. Hang, On verifying the first-order Markovian assumptions for a Rayleigh fading channel model, Proceedings of IEEE ICUPC94, San Diego, pp , September [13] M. Zorzi, R. R. Rao, L. B. Milstein, On the accuracy of a first order Markov model for data block transmission on fadin channels, Proceedings of IEEE ICUPC 95, pp , November [14] K.S. Miller, Multidimensional Gaussian Distribution, New York: Wiley, Fig. 5. Comparison of the decoded image with the original transmitted image P c = 0.77 with n max = 1, without truncated power control. Preroll delay of = 13s.