Energy-efficient Video Streaming from High-speed Trains

Transcription

1 Energy-efficient Video Streaming from High-speed Trains Xiaoqiang Ma, Jiangchuan Liu Computing Science Schoo Simon Fraser University Hongbo Jiang Department of EIE Huazhong University of Science and Technoogy ABSTRACT The probem of streaming packetized media has been intensivey studied for a ong time. In this paper, we revisit this probem in the high-speed raiway context, where passengers encode and upoad videos through increasingy powerfu smartphones. The chaenge is highighted by the fast changing channe conditions in high-speed trains and the imited battery of ce phones. Inspired by the unique spatiatempora characteristics of wireess signas aong high-speed raiways, we propose a nove energy-efficient and rate-distortion optimized approach for video streaming. Our soution effectivey predicts the signa strength through its spatiatempora periodicity in this new appication scenario. It then smarty adjusts the GOF budget, schedues the video transmission to achieve gracefu rate-distortion performance and yet conserves the energy consumption. Performance evauation based on simuated raiway scenarios and H.264 video traces demonstrates the effectiveness of our soution and its superiority as compared to existing soutions. Categories and Subject Descriptors H.5.1 [Mutimedia Information Systems]: [Video]; C.2.1 [Network Architecture and Design]: [Wireess communication] Genera Terms Experimentation Keywords Energy-efficient Rate-Distortion optimization, High-speed trains 1. INTRODUCTION The probem of streaming packetized media has been studied for many years [1, 2, 3]. With the rapid deveopment of broadband wireess networks and powerfu smartphones, wireess video communication (e.g., Facetime [4]) is gaining Permission to make digita or hard copies of a or part of this work for persona or cassroom use is granted without fee provided that copies are not made or distributed for profit or commercia advantage and that copies bear this notice and the fu citation on the first page. To copy otherwise, to repubish, to post on servers or to redistribute to ists, requires prior specific permission and/or a fee. NOSSDAV 11, June 1 3, 2011, Vancouver, British Coumbia, Canada. Copyright 2011 ACM /11/06...$ more and more popuarity in peope s ife. In this paper, we revisit the video streaming probem in the high-speed raiway context, where passengers encode and upoad videos through their wireess mobie devices. Recent decades have witnessed the fast deveopment of high-speed raiway systems in many countries, such as Japan, France, China and so on [6]. The speed of the trains can be as high as 350 km/h and thus the trave time between distant cities has been significanty shortened. Couped with this convenience, however, is the chaenge to traditiona ceuar techniques. For exampe, the handover frequency wi be very high without redesigning the base station depoyment, eading to increased ca bocking rate and ca drop rate. The channe properties wi change very quicky, which prevent the fast power contro in WCDMA from accuratey compensating for fading [7]. The Dopper shift wi aso be significant, which degrades the system performance. A ot of work has been done towards overcoming the difficuties mentioned above. Uhirz [8] outines the concept of a GSM-based communication system for high-speed Raiway (GSM-R) and proposes to make some modifications to the standard GSM system. A inear coverage is preferred to area coverage to simpify handover. Frequency compensation is used to eiminate the Dopper shift. In order to provide broadband data services, sateite-to-wi-fi inks have been combined for Internet access on the TGV trains in France. More importanty, sateite inks have imited bandwidth and ong round trip times, which make them not idea for reatime appications, such as video streaming and video conferencing [9]. The GSM Digita Remote RF Units (GRRU) private network [10] has been depoyed aong the Beijing- Tianjin Intercity Raiway. It can provide better communication quaity and use fewer handover/reseection compared with existing techniques. Whie these soutions improve the channe properties, we sti face another key probem that band-intensive appications are very power-consuming in terms of data transmission and thus mobie devices wi run out of power in a short time. This phenomenon is especiay obvious when the signa is weak, since the transceiver needs to ampify the signa [11]. When mobie users are on high-speed trains, their ce phones wi work under fast-changing signa strength, mainy caused by the varying distance between ce phones and base stations. Ce phones need to impement power contro on the upink to compensate for the fading effect such that the received power eve at base stations wi stay fairy constant [7]. With regards to this observation, we propose a nove dynamic rate-distortion optimization approach 117

2 that smarty contros the video rate, together with a transmission scheduing according to the received signa strength. Given the periodicity of channes, this strategy remarkaby saves the power on data transmission and yet achieve gracefu degradation of video quaity. The remainder of this paper is organized as foows, In Section 2, we introduce some reated work on energy efficient approaches in wireess and ceuar networks. In Section 3, we propose the system mode, formuate the optimization probem and describe our soution. Simuation resuts on video sequences is iustrated in Section 4. Section 5 gives some discussion and we concude this paper in Section RELATED WORK Existing techniques can aready guarantee high-quaity voice services for users moving at high speed. Yet providing high speed data transmission for peope in fast-moving vehices is sti an open issue. The FAMOUS [5] architecture is proposed to offer mutimedia services to fast moving users. For the train scenario, the authors propose to combine a Radio-over-Fiber (ROF) network with movabe ces. Fokum and Frost [9] present a comprehensive survey of approaches for providing broadband Internet access to trains. They present a taxonomy of architecture according to access network technoogies incuding ROF, IEEE , sateite and so on. They then compare various impementations in both Europe and North America and summarize the essons we can earn from. In the research fied of wireess networks, the cross-ayer design approach is proposed to achieve better system performance. In [12], the authors summarize the chaenges and principes of cross-ayer wireess mutimedia transmission and propose a new paradigm to improve mutimedia streaming quaity and reduce power consumption. Eric et a. [13] further expore a cross-ayer design framework for rea-time video streaming in Ad hoc networks. Signa strength can be viewed as the refector of current channe properties. If we can predict the signa strength, adjustment can be made in advance. In [14], the authors propose a ong range onine prediction method, which is ocation-independent. However, they ony verify their method at waking speeds, where the signa strength does not change very fast. Power contro is cosey couped with signa strength. It is especiay important for upink to keep the received signa strength eve at base stations within sight variation. Without power contro, we wi see the so-caed near-far probem in CDMA systems [7]. A ot of previous works have been done to impement efficient and accurate power contro [15]. However, power contro wi significanty increase the transmission energy consumption when mobie users are far from base stations. The existing work [11] proposes to transmit ess data when the channe is bad. There have been significant studies on rate-distortion optimization for video streaming. Chou and Miao [1] formuate how and when to transmit a group of interdependent data packets with deivery deadines in a rate-distortion way. He et a. [16] further add a power dimension to the rate-distortion mode to perform onine resource aocation and energy optimization. Eckehard et a.[17] present an adaptive media payout technique that adjusts the payout speed according to channe conditions to improve the tradeoff between buffer underfow probabiity and atency [17]. Our study has been inspired by these works; yet we focus on streaming from high-speed trains that creates new chaenges. 3. SYSTEM MODEL In this section, we first estabish the energy mode in the high-speed raiway scenario. We then formuate the powerrate-distortion optimization probem and describe how to achieve energy efficiency dynamic rate contro and transmission scheduing. 3.1 Energy mode Path oss mode, which is important for ink budget, has attracted the interest of many researchers. The basic mode [18] is as foows: P L(d) = P L(d 0) + 10n og ( d d 0 ) + X σ (1) where P L(d) is the average path oss vaue at distance d from a measured ocation to the transmitter; n is the path oss exponent, which depends on the environment; d 0 is the cose-in reference distance, and P L(d 0) is based on either practica measurements or on a free space path oss mode at distance d 0 from the transmitter; X σ is a zero-mean Gaussian distributed random variabe (in db) with standard deviation σ (aso in db), and is computed from measured data. The above parameters statisticay describe the path oss mode. The path oss can be transformed from the measured received signa strength as [19]: P r(d) = P t + G P L(d) P L other (2) where P r(d) and P t (both in dbm) are the power eve of received signa and transmitted signa, respectivey; G is the antenna gain and P L other is the attenuation caused by other factors. G and P L other are constants for the same type of base stations. Equation (1) impies that the arge-scae variation of path oss ony depends on the distance from the transmitter. In rea ife, especiay in metropoitan areas with dense buidings, the sma-scae fading and mutipath effect wi cause rapid and significant changes in signa strength over a sma trave distance or time interva. However, since most highspeed raiways are constructed in suburban and rura areas that can be approximatey regarded as open space environment, mutipath effect can be argey negected. As mentioned in Section 2, power contro needs to be impemented on ce phones to compensate for the variation of path oss. The power needed to transmit a bit at distance d is [21]: u(d) = ad n + c 0 (3) where a and c 0 are constants; n is the path oss exponent. For the sake of simpicity, in the remaining part of this paper, we ony consider the former part on the right side of the above equation, ad n, as the transmission consumption and use c(d) to denote it. In practice, however, the distance can not be directy measured, so ce phones can adjust transmission power based on received signa strength. The ratio of transmission power at different paces A and B is then given by: p(d A) p(d B) = 10 Pr (db ) Pr (da) 10 (4) 118

3 where d A and d B are the distances from the transmitter at A and B, respectivey. To better iustrate the fast-changing received signa strength and transmission power as we as the reationship between them, we now show a series of simuated resuts. Our simuation is based on the path oss mode in open area in [19], which is measured aong the Zhengzhou-Xi an high speed raiway environment at the 930 MHz band: P r(d) = og(d) (5) The standard deviation of shadowing is chosen as In order to reduce the construction cost and keep trains running at reativey constant speed, the high speed raiway is amost a straight ine in a consideraby ong segment. This construction strategy simpifies the evauation of path oss aong the raiway. We give the Beijing-Tianjing Intercity Raiway as an exampe in Figure 1. Received signa strength (dbm) Location (m) x 10 4 Figure 2: Received signa strength aong the raiway and each edge directed from data unit to data unit means that data unit must be decoded in order to decode data unit. We use to denote this reationship. A sampe dependence graph of a group of (IBPBP) frames is shown in Fig. 3. Figure 3: Dependency between IBPBP video frames Figure 1: Beijing-Tianjin Intercity Raiway from Googe Map [20] This construction strategy simpifies the evauation of path oss aong the raiway. In our simuation, we use a straight ine to represent a segment of 35 km raiway, aong which 18 base stations every 2 km have been depoyed. The straight distance between each base station and the rai is 150 m. We cacuate the received signa strength every 10 m, and the transmission energy at each ocation is represented by the ratio of it to the minimum energy consumption aong the this segment of rai. We use the ratio rather than an absoute vaue of power consumption because the ratio is independent of the types of base stations and ce phones that are used, and thus our mode can work in genera cases. From Fig. 2 we can find out that the difference of received signa strength can be as high as 20 db and thus the maximum transmission energy is over 100 times more than the minimum. The peaks in Fig. 2 correspond to the ocations with the minimum distance of 150 m to base stations. 3.2 Probem formuation A video sequence can be encoded into frames at different rates with corresponding distortion. The encoded data are packetized into a finite set of data units. The interdependency of data units can be represented by a directed acycic graph. Each node of the graph corresponds to a data unit, We use D 0 to denote the distortion if no data unit is decoded, D the reduction of distortion if data unit is decoded on time, and B the size in bits of data unit. For each data unit, et t 1, t 2,..., t N be N discrete transmission opportunities and et t DT S be the deivery deadine. A transmission poicy π = (π(1), π(2),...π(n) {0, 1} N is used to describe the transmission of a data unit, where π(i) = 0 means that the data unit shoud not be sent at opportunity i whie π(i) = 1 means that the data unit shoud be sent at opportunity i if no acknowedgement packet was received from the feedback channe before t i. We define an error ɛ(π) for poicy π as the probabiity that the data unit cannot be successfu received before its deadine t DT S, ɛ(π) = P {F T T > t DT S t i} (6) i:π(i)=1 where F T T is the forward trip time. The cost ρ(π) for poicy π is defined as the expected number of data unit transmissions, that is [22], ρ(π) = P {RT T > t i t j} (7) i:π(i)=1 j<i:π(i)=1 where RT T is the round trip time. Here we do not make any particuar assumption for the probabiity distribution of F T T and RT T. A shifted Gamma distribution, however, is widey used [1]. 119

4 The expected power consumption for transmitting data unit is therefore, ρ(π) C() = B c(d(t j)) (8) j=1 where d(j) is the distance from the train to the base station at opportunity t j when data unit is sent. We then consider the probem of rate-distortion optimized transmission of a group of interdependent data units. For the sake of simpicity, we consider a group of frames with the size of L and each data unit contains exact one frame. This GOF can be encoded in M different rates and each rate corresponds to a distortion. The transmission poicy for the GOF is described by a poicy vector π = (π 1,..., π L), where π, 1, 2,..., L is the transmission poicy for the th data unit of the GOF. The expected power consumption for transmitting this GOF is, C( π ) = ρ(π) L B (k)c(d(t j)) (9) =1 j=1 where B (k), k {1,..., M} is the size of data unit when this GOF is encoded in rate R(k). The expected distortion for the group is, D( π ) = D 0(k) L =1 D (k) (1 ɛ(π )) (10) Our objective is to jointy optimize the poicy vector to minimize the expected distortion subject to a constraint on the expected power consumption. We can sove the probem by finding the poicy vector π and the GOF rate aocation R that minimizes the expected Lagrangian, as in [1] J( π ) = D( π ) + λc( π ) = D 0k + ( L D (k) ) ɛ(π )) (1 =1 ρ(π) + λ j=1 B (k)c(d(t j)) (11) where (B (k), D (k)) {(B (1), D (1),..., (B (M), D (M)} To minimize J( π ) for a given λ and an arbitrary pair of (B (k), D (k)), k 1,..., M, we extends the Iterative Sensitivity Adjustment (ISA) agorithm [1]. The main idea is minimizing one variabe at a time, whie keeping other variabe constant unti convergence. For exampe, we can start from any initia poicy vector π (0) and determine the minimum J( π ) as foows. Seect one data unit n 1,..., L to find its optima poicy vector at step n. Then for n, set π (n) where π (n) = π (n), whie for = n, et = arg min J(π (n) 1,..., π (n) π 1, π, π (n) +1 = arg min π S (n) ρ(π) ɛ(π ) + λ j=1 S (n) (k) = D (k),..., π(n) L ) B (k)c(d(t j)) (12) (1 ɛ(π (n) )) (13) We can find the optima poicy vector for a the M ratedistortion pairs. The poicy vector and the rate-distortion pair with the minimum vaue of J( π ) are seected to be the transmission scheduing and rate aocation for this GOF, respectivey. 3.3 Signa strength prediction The power consumption for transmitting bits at each transmission opportunity is cosey reated to the signa strength at that time as we mentioned before. We describe how to predict the signa strength prediction as foows. We assume that the base stations are of the same type, and the environmenta conditions have no significant change for a reativey ong segment of raiway, for exampe, tens of kiometres. Then the arge-scae variation of the signa strength depends ony on the distance from the ce phone to its connected base station. The ce phone first coects traces of received signa strength vaues at a constant frequency (In our experiment,we coected three periods of signa strengths to obtain the statistics such as the mean vaue and standard deviation). Since the train runs at a reativey constant speed, the period of variation T can be easiy computed. Then we can have a coarse evauation of the signa strength for the future. Suppose the current time is t 1 and the received signa strength is P (t 1), and the average received signa strength at the corresponding time in the coected traces (t T, t 2T...) is P (t 1). We can use weighted sum (1 α) P (t 1)+α P (t 1) as the prediction of the signa strength at time instance t 1+T. We have vaidated the accuracy of this prediction method in Section 4. We aso pan to expoit new technoogies, such as eectronic compasses and acceerometers, to further improve the prediction accuracy and we wi further discuss it in Section PERFORMANCE EVALUATION In this part, we perform simuations to evauate our agorithms and aso give some anaysis based on the resuts. We use the experimenta settings mentioned in Section 3. We use Matab R2010b as our simuation too. We ran each experiment 10 times, and cacuated the mean vaue. 4.1 Signa strength prediction We first show the accuracy of our signa strength prediction agorithm. By cacuating the reative errors with different choices of α, we can choose the α with smaest error in our foowing simuation. We use the first three periods of vaues as the coected traces and cacuate the average vaue at each point. The fourth period is regarded as the current vaue and is used to predict the corresponding vaue in the fifth period. α is chosen from 0 to 1 with an interva of The simuation resuts with 95% confidence eve are shown in Fig. 4. It is ceary that, with α = 0.25, the prediction agorithm produces the smaest reative error. Hence, we choose this vaue in the simuation. 4.2 Streaming quaity and power consumption We next evauate Streaming quaity and power consumption compared to a baseine strategy that encodes the video sequence into the highest quaity and upoads the data units at the first possibe transmission opportunity for different deivery deadines. We first show the received video quaity degradation of our approach which is measured by the 120

5 Reative error (%) Power consumption ratio α ch Deivery deadine (s) Figure 4: Reative errors of prediction with different α PSNR decrease. We then compare the two approaches in terms of power consumption on transmission. As mentioned in Section 3.1, we cacuate the received signa strength and energy radio of each ocation every 10 meters. We assume that both the signa strength and energy ratio wi not dramaticay change during the 10 meters. The time interva between two transmission opportunities is 100ms. We empoy the QCIF size video sequence of Foreman and use JM 17.2 software of the H.264/AVC standard [23] to encode and decode it. The video sequence is encoded into 50, 100, and 150 kbps, and the corresponding PSNR are 34.8 db, db, and db, respectivey. The frame rate is 10 fps and we repicate the video sequence to 3 minutes. The structure of GOF is set of 10 frames with one I frame foowed by nine P frames. We sove the optimization with the basic unit of GOF. The maximum transmission rate is 128 kbps and the mean RTT is 150ms. The transmission oss rate is set to be 10%. In the simuation, we randomy use the starting point to transmit and vary the deivery deadine from 1 to 5 seconds. With the fixed time interva of 100ms, the number of transmission opportunities varies from 10 to 50. With increased deadine, the number of transmission opportunities wi aso be increased. The resuts are shown in Fig. 5 and Fig. 6. ch PSNR (db) Deivery deadine (s) Figure 5: Received video quaity degradation compared with baseine approach We divide the transmission power consumption of our Figure 6: Power saving ratio compared with baseine approach method by that of the baseine strategy to get the power saving ratio. We can see that with increased deivery deadine, both the distortion and transmission power consumption get reduced. The reason is that with more transmission opportunities, we can better expoit the signa strength prediction. Considering that the worst case compexity to sove (13) increases exponentiay with the number of transmission opportunities, we suggest that deivery deadine around 3 seconds can achieve a good tradeoff between the distortionpower and the time needed to find the optima transmission poicy. To get better resuts, we can encode the video sequence into more rate-distortion pairs. This method, however, woud increase the computation compexity. Fortunatey, we can get a approximate resut by considering ony a subset of ratedistortion pairs according to the signa strength prediction. To be precise, we can choose the pairs with higher rate when the signa strength is strong in the near future, and choose the pairs with ower rate when the signa strength is weak in the near future. We eave this as part of our future work. 5. DISCUSSION In this section, we wi discuss some practica issues worthy of noting. First, the speed of trains is not absoutey constant. Instead, it can be increased and decreased, and can aso have some fuctuations. The change of speed wi affect the accuracy of signa strength prediction. GPS can be used to get the ocation information, but it consumes a ot of power to track sateites [11]. Due to the ong propagation deay of sateite communication inks and the high speed of trains, its error of ocaization can be very significant. Recenty, nove ocaization approaches, e.g. [24], thus have to be incorporated. Eectronic compasses and acceerometers, which are readiy avaiabe in modern ce phones, can aso be utiized to figure the ocation information, which is more energy-efficient and accurate than GPS. We are currenty examining to use acceerometers to track the speed change of trains, which can be used to obtain more accurate channe conditions. It coud be aso combined with the adaptive media payout technique to adapt the video payout speed to the speed change of trains. Second, the distance between base stations may not be 121

6 equa. To sove this probem, the mobie phone users go opposite directions can exchange the ocation information of base stations and the statistica information of signa strength. This can hep to improve the prediction accuracy. These kinds of information can be embedded into the GPS map or saved in base stations and fetched by the passing by ce phones. 6. CONCLUSION The rapid deveopment of high-speed trains raises new chaenges to ce phone vendors with respect to QoS management. The imited battery capacity and the increasingy popuar bandwidth-intensive and power-hungry appications, such as fies sharing and video streaming, attract a ot of researchers to address the power contro probem to proong the battery ife. In this paper, by utiizing the construction feature of high-speed raiways and the base stations as we as the motion feature of trains, we revisited the probem of rate-distortion optimized video streaming. In future work, we pan to further expoit the periodicity of the signa strength to design approximate agorithm to sove the optimization probem. Acknowedgments This research is supported by a Canada NSERC Discovery Grant, an NSERC DAS grant, an NSERC Strategic Project Grant, and an MITACS NCE Project Grant. 7. REFERENCES [1] P. A. Chou, Z. Miao. Rate-distortion optimized streaming of packetized media. IEEE Transactions on Mutimedia, 8(2): , [2] J. Chakareski, P. A. chou. Appication ayer error-correction coding for rate-distortion optimized streaming to wireess cients. IEEE Transactions on Communications, 52(10): , [3] H. Seferogu, O. Gurbuz, O. Ercetin, Y. Atunbasak. Rate-distortion based rea-time wireess video streaming. Journa Image Communication, 22(6): , [4] [5] F. D. Greve et a. Famous: a network architecture for deivering mutimedia services to fast moving users. Wireess Persona Communications,: 33(3-4): , [6] rai [7] H. Homa, A. Toskaa. WCDMA for umts: hspa evoution and te. Wiey, 2007 [8] M. Uhirz. Concept of a gsm-based communication system for high-speed trains. In Proceeding of IEEE VTC 1994, pages , [9] D. T. Fokum, V. S. Frost. A survey on methods for broadband internet access on trains. IEEE Communications Surveys & Tutorias, 12(2): , [10] X. Cheng, Y. Li, X. Cao. The discussion on gsm coverage scheme of high-speed raiway, In Proceedings of IEEE ICCTA 2009, pages , [11] A. Schuman, V. Navda, R. Ramjee, N. Spring, P. Deshpande, C. Grunewad, K. Jain, V. Padmanabhan. Bartendr: a practica approach to energy-aware ceuar data scheduing. In Proceedings of MobiCom 2010, pages 85-96, [12] M. van Der Schaar, N. Sai Shankar. Cross-ayer wireess mutimedia transmission: chaenges, principes, and new paradigms. IEEE Wireess Communications, 12(4):50-58, [13] E. Setton, T. Yoo, X. Zhu, A. Godsmith, B. Girod. Cross-ayer design of ad hoc networks for rea-time video streaming. IEEE Wireess Communications Magazine, 12(4):59-65, [14] X. Long, B. Sikdar. A rea-time agorithm for ong range signa strength prediction in wireess networks, In Proceedings of IEEE WCNC 2008, pages , [15] R. D. Yates. A framework for upink power contro in ceuar radio systems. IEEE Journa on Seected Areas in Communications, 13(7): , [16] Z. He, W. Cheng, X. Chen. Energy Minimization of Portabe Video Communication Devices Based on Power-Rate-Distortion Optimization. IEEE Transactions on Circuits and Systems for Video Technoogy, 18(5): , [17] M. K. Eckehard, E. Steinbach, B. Girod. Adaptive Payout For Rea-Time Media Streaming. ICIP 2002, pages 45-48, [18] T. S. Rappaport. Wireess communications. Prentice Ha, [19] R. He, Z. Zhong, B. Ai. Path oss measurements and anaysis for high-speed raiway viaduct scene, In Proceedings of IWCMC 2010, pages , 2010 [20] [21] I. Stojmenovic, X. Lin. Power-aware ocaized routing in wireess networks. IEEE Transactions on Parae and Distributed Systems, 12(11): , [22] M. Röder, J. Cardina, R, Hamzaoui. On the compexity of rate-distortion optima streaming of packetized media. In Proceedings of DCC 2004, pages , [23] [24] I. Constandache, R. R.Choudhury, I. Rhee. Towards mobie phone ocaization without war-driving. In Proceedings of IEEE INFOCOM 2010, pages 1-9,