Joint Buffering and Rate Control for Video Streaming over Heterogeneous Wireless Networks
by
Lei Hua
A thesis submitted in conformity with the requirements for the degree of Master of Applied Science Graduate Department of Electrical and Computer Engineering University of Toronto
Copyright c 2010 by Lei Hua Abstract
Joint Buffering and Rate Control for Video Streaming over Heterogeneous Wireless
Networks
Lei Hua
Master of Applied Science
Graduate Department of Electrical and Computer Engineering
University of Toronto
2010
The integration of heterogeneous access networks is becoming a possible feature of 4G wireless networks. It is challenging to deliver the multimedia services over such integrated networks because of the discrepancy in the bandwidth of different networks. This thesis presents an adaptive approach that combines source rate adaptation and buffering to achieve high quality VBR video streaming with less quality variation over an integrated two-tier network. Statistical information of the residence time in each network or local- ization information are utilized to anticipate the handoff occurrence. The performance of this approach is analyzed under the CBR case using a Markov reward model. Simulation under the CBR and VBR cases is conducted for different types of network models. The results are compared with a dynamic programming algorithm as well as other naive or intuitive algorithms, and proved to be promising.
ii Acknowledgements
I would like to express my sincerest gratitude to my supervisor, Professor Ben Liang,
for this exciting opportunity to work under his supervision at this prestigious institu-
tion. During the whole process he provided me with invaluable guidance, inspiration and
support, without which I couldn’t have completed this work.
I am thankful to the members of my thesis committee, Prof. Elvino S. Sousa, Prof.
Raviraj Adve, and Prof. Jason H. Anderson for the time spent in reviewing my thesis,
and for their helpful feedback and comments on improving its content.
I thank all my current and former colleagues in my research group for their useful
inputs and suggestions on the research work itself and also the presentation of the work.
Special thanks to all of my friends at University of Toronto, Colin Jiang, Eric Yuan,
Junqi Yu, Lilin Zhang, Weiwei Li, Yuan Feng, Yunfeng Lin and others, whose company,
care and encouragement made the two years of Master’s studies much more enjoyable.
Last but never the least, I dedicate this thesis to my family, who are always there for me in my life.
iii Contents
1 Introduction 1
1.1 Overview...... 1
1.1.1 VideoStreaming ...... 1
1.1.2 Heterogeneous Wireless Networks ...... 2
1.1.3 Buffering...... 3
1.1.4 RateAdaptation ...... 4
1.1.5 ContributionoftheThesis ...... 4
1.2 ThesisOutline...... 6
2 Literature Review 7
2.1 VideoRateAdaptationTechniques ...... 7
2.1.1 Transcoding...... 8
2.1.2 Joint Source/Channel Coding ...... 8
2.1.3 ScalableVideoCoding ...... 9
2.1.4 Content-Aware Coding Techniques ...... 10
2.2 Rate Control in Heterogeneous Wireless Networks ...... 11
2.3 Buffering in Heterogeneous Wireless Networks ...... 12
3 Problem Statement 14
3.1 ApplicationScenario ...... 14
3.2 ModelsandAssumptions...... 16
iv 3.2.1 Rate Adaptation and Playback ...... 16
3.2.2 Residence Time and Rate Estimation ...... 18
3.2.3 Feedback Control Mechanism ...... 19
3.3 ProblemFormulation...... 19
4 Generic Network Model 22
4.1 ControlAlgorithms ...... 22
4.1.1 Adaptive Control Algorithm ...... 22
4.1.2 SimpleAlgorithm...... 25
4.1.3 Mean Residual Life Based Algorithm ...... 26
4.1.4 SimpleShapingAlgorithm ...... 27
4.2 Analytical Framework and Analytical Results ...... 29
4.2.1 Analytical Results for Generic Model ...... 30
5 Markov Chain Network Model 36
5.1 MarkovDecisionProcessModel ...... 36
5.2 DynamicProgrammingAlgorithm...... 38
5.3 SimulationResults ...... 39
5.4 MoreRealistic3-ZoneNetworkModel...... 43
5.5 PH-FittingofResidenceTimes ...... 45
5.6 Estimation in Adaptive Control Algorithm ...... 47
5.6.1 Utilizing Statistical Information ...... 47
5.6.2 Utilizing Localization Information ...... 48
5.6.3 Simulation Results for 3-Zone Model ...... 49
5.7 Simulating with VBR Network and VBR Video Stream ...... 52
6 Conclusion 57
Bibliography 59
v List of Tables
3.1 Notationsinsystemmodel...... 20
4.1 Analysisparameters-1...... 31
4.2 Analysisparameters-2...... 31
5.1 Simulation parameters for 2-zone Markov model ...... 40
5.2 Simulation parameters for 3-zone model ...... 49
5.3 Simulation parameters for VBR network and VBR video ...... 52
vi List of Figures
3.1 Integrated two-tier network ...... 15
3.2 Relationship between the transmission sequence in time and the playback
sequenceintime ...... 18
4.1 Illustration of proportional feedback controller ...... 24
4.2 Distributionsofresidencetimes ...... 32
4.3 Analysis vs simulation results: generic model, Gamma distribution - 1 . . 34
4.4 Analysis vs simulation results: generic model, Gamma distribution - 2 . . 35
5.1 DP: variation and utilization vs. α ...... 41
5.2 Adaptive algorithm: variation and utilization vs. β ...... 41
5.3 Comparison between algorithms: utilization vs. variation ...... 42
5.4 Integrated two-tier network with 2-zone T2N ...... 44
5.5 An example of the generated user’s moving trace ...... 44
5.6 CDF’s of residence times in different zones ...... 46
5.7 PH-fittedMarkovchainnetworkmodel ...... 47
5.8 DP on 3-zone model: variation and utilization vs. α ...... 50
5.9 Adaptive algorithm on 3-zone model: variation and utilization vs. β . . . 51
5.10 Comparison DP and AA: utilization vs. variation ...... 51
5.11 VBR simulation: adaptive algorithm with statistical information . . . . . 53
5.12 VBR simulation: adaptive algorithm with localization information . . . . 53
vii 5.13 VBR simulation: simple adaptive algorithm ...... 54
5.14 Simulating VBR case - variation vs. β ...... 55
5.15 Simulating VBR case - utilization vs. β ...... 55
5.16 Simulating VBR case - utilization vs. variation ...... 56
viii Chapter 1
Introduction
1.1 Overview
1.1.1 Video Streaming
Online video has become a mainstream medium and the single most influential factor driving the need for increased mobile network capacity [8]. It would take 28 years to watch the video uploaded to YouTube in the week of April 29th, 2010 [20]; HD(high defi- nition) movies and television programs are widely available online with the help of CDNs
(Content Distribution Networks) and P2P (Peer-to-Peer) networks; video conferencing and video phones are not the exclusive rights of large companies any more, but can be enjoyed by individuals and families. It is then important and interesting to research on improving video streaming techniques.
There are two types of video streaming applications: live streaming, which captures real-time events and provides the video to users, and on-demand streaming, which offers stored video contents. Application scenarios of live streaming include video conference, video phone and live event broadcasting, which have stringent delay requirement. In this thesis we consider the transmission of pre-encoded video, which is used for delivering all kinds of published video contents and user generated contents online and is expected to
1 Chapter 1. Introduction 2
account for sixty-six percent of the world’s mobile data traffic by 2014 [21].
In comparison to other traffic flows such as Web browsing and E-mail, video streaming
has its unique characteristics and therefore may impose certain requirements on the
network. Video streaming traffic is inelastic. Unlike web browsing or file downloading, where data can be transmitted at any rate, video streaming requires certain amount of data to be delivered and decoded before the playback deadline. Hence it is sensitive to variations in both bandwidth and transmission delay. Video streaming applications are loss-tolerant. Robust coding techniques allow video to be decoded with certain loss of
data. However, this does not mean any level of loss can be tolerated. In high error-
rate networks, it is challenging to develop loss-prevention techniques for robust video
transmission.
1.1.2 Heterogeneous Wireless Networks
With the rapid growth of mobile communication technology, various wireless networking
technologies have evolved and become widely deployed all over the world, allowing people
to access the Internet with all kinds of mobile computing devices, at all times and all
places. The popular access technologies include IEEE802.11 wireless local area networks
(WLAN), WiMAX, GPRS, UMTS, and CDMA2000, etc. These technologies are hetero-
geneous in certain attributes, such as coverage area, protocol, signaling mechanism, data
rate, error rate, etc. However, it is common for the personal mobile devices (laptops,
smart phones, PDAs, digital media players) to support more than one wireless access
technologies simultaneously.
With the coexistence of heterogeneous wireless networks and the devices supporting
multiple access technologies, the integration of heterogeneous wireless networks is be-
coming a trend and is part of the 4G network design [30]. This feature allows user to
seamlessly switch among different wireless network interfaces and enjoy greatly enlarged
coverage and more reliable wireless access on a single device. Chapter 1. Introduction 3
However, there are many challenges in deploying such an inter-technology roaming environment. Active research topics on heterogeneous wireless networks involve admis- sion control, hand-off mechanism, mobility management, traffic flow assignment, etc.
The heterogeneity of wireless access technologies also imposes great challenges on video streaming applications running on a mobile device in the integrated network.
In heterogeneous wireless networks, handoffs inside one technology and between tech- nologies, can cause extra delays, which exaggerates the challenge on the delay require- ments of video streaming applications. A more substantial problem in the heterogeneous wireless networks is that, different access networks offer different ranges of bandwidth, which greatly exacerbates the variations in streamed video quality if we simply match the video source rate to the available transmission rate. Hence, in this thesis we mainly focus on reducing the variation in streamed video quality while maintain high average quality.
1.1.3 Buffering
Two types of video streaming techniques are commonly applied in both wired and wireless networks to combat the varying network bandwidth and delays: buffering and video rate adaptation.
Buffering sustains the video playback when available bit rate (ABR) is low, by prefetching and storing a certain amount of data ahead of (playback) time. With a
finite buffer size, two types of event will happen and may cause detrimental effects to the streaming process: buffer underflow and buffer overflow. Underflow may happen when the playback rate (data consumption rate) is higher than the transmission rate, which leads to playback jitters (stops). Overflow may happen when the transmission rate is higher than the playback rate, while at the same time the buffer size is small. Buffer overflow may lead to loss of data and then playback jitters.
Another factor to consider in buffering is the initial buffering delay, i.e. the waiting Chapter 1. Introduction 4
time between starting the buffering and starting the playback. There is a trade-off
between the initial buffering delay and the buffer size when we aim to provide satisfiable
video streaming service[19].
In this thesis, we consider the longer-term variations in the transmission rate in heterogeneous wireless networks, hence we assume an infinite buffer size. Also we set the initial buffering delay to be minimal. We are primarily interested in how much data to buffer for the future in every time slot and at what quality should we buffer it.
1.1.4 Rate Adaptation
Rate adaptation techniques match the video source rate to the network transmission rate, when the transmission rate is low, at the cost of lowering the perceived quality of decoded video. Various video rate adaptation techniques have been proposed over time, such as transcoding, joint source/channel coding, multiple file/rate switching, scalable video coding, and content-aware coding techniques [2]. We present some of them in
Chapter 2.
While theoretically rate adaptation can ensure continuous playback as long as the
ABR is higher than the minimum required rate of the specific adaptation technique, it introduces fluctuations in the perceived quality of the video, which can be annoying to users. This problem is exaggerated in heterogeneous wireless networks since the variation of ABR there is much higher than in homogeneous wireless networks.
1.1.5 Contribution of the Thesis
In this thesis, we consider the problem of streaming pre-encoded video on a moving mobile terminal (MT) in heterogeneous wireless networks. The video source is stored in a remote server and transmitted through the backbone network to the local access points
(AP) or base stations (BS) and then to the MT through different wireless networks. The bottleneck of the connections always lies in the last hop (i.e. the wireless hop.) There is Chapter 1. Introduction 5
a receiving buffer on the device, which is used for storing prefetched video contents.
We focus on coping with the variable ABR in heterogeneous wireless networks. The effects of other network characteristics, such as varying end-to-end delay and high error rate, are assumed to be resolved using any available technique. Our objectives are con- tinuous playback, high image quality, and low variation in the perceived image quality
(or constant-quality playback).
To achieve these objectives, we propose to combine buffering and rate adaptation
techniques with prediction of certain attributes of the network. Our scheme predicts
the residence times at each individual network, then dynamically allocates the ABR to
each unit of video sequences being transmitted, hence controlling the buffer and rate
adaptation at the same time, under the constraint of fully utilized network resources.
In order to determine the optimal way to allocate the ABR, we propose an adaptive
video rate control scheme using a linear feedback control technique on a generic network
model for a two-zone network. In designing the scheme, we divide the whole streaming
process within the heterogeneous wireless networks into cycles and try to achieve local
optimality within each cycle.
To show the applicability of our scheme to any arbitrary distribution, the performance
of our proposed scheme in a simplified constant-bit-rate (CBR) network scenario with
CBR video source is evaluated in an analytical framework based on Markov chains, where
the state space is dimensioned by the normalized quality levels and the buffered lengths
of video at the end of each cycle. We associate with each state a cost being the quality
variations within the cycle and calculate the average cost per cycle. Other naive and
intuitive algorithms are also studied within the analytical framework in order to show
the advantage of our adaptive scheme.
Then, for the special case of exponential network residence times, we formulate the
streaming process into a finite-horizon controlled Markov Decision Process (MDP), and
solve the optimization problem using a dynamic programming based optimal control Chapter 1. Introduction 6 algorithm. Although this method is assumed to provide the theoretical optimality, it involves a large amount of computation, cannot deal with increasing dimensionality, and is not applicable to more generalized residence time distributions. On the other hand, the aforementioned adaptive algorithm is much simpler than the dynamic programming algorithm in terms of the amount of computation involved, works with more generalized distributions, and requires less knowledge of the network. Through simulations we show that this scheme provides near optimal performance.
Furthermore, we develop a more realistic network model by modeling the movement of the MT and fitting the actual residence times using Phase-Type distributions. We also increase the number of network zones to three. We show through simulation that our scheme also gives near-optimal performance under the new model. Furthermore, we simulated our scheme with variable-bit-rate (VBR) networks and VBR video sources, and explored the effects of utilizing different estimations of residence times, i.e. the statistical information extracted from history mobility traces, and the geographical information provided by localization service. Our algorithm proves to provide significantly improved performance with VBR networks and VBR video source than the naive algorithms.
1.2 Thesis Outline
The thesis is organized as follows. The next chapter reviews work in the related areas of video rate control and buffering techniques in both homogeneous and heterogeneous wireless networks. The system setup and the problem statement are presented in Chapter
3. Chapter 4 is focused on a generic network model and the design of our adaptive control algorithm. We also present the analytical performance evaluation framework. In
Chapter 5, we introduce the Markov Chain network model and the dynamic programming algorithm. We also extend the system models to a 3-zone model with PH-fitted residence times and simulate the VBR case in Chapter 5. Finally, Chapter 6 concludes the thesis. Chapter 2
Literature Review
This chapter briefly reviews the existing research progress on video streaming technologies on both homogeneous and heterogeneous wireless networks and the current challenges, which motivated our research work. We first discuss some research works on video rate adaptation in general variable-bit-rate (VBR) networks. Then we present some related research on buffering and rate control techniques in video streaming over heterogeneous wireless networks.
2.1 Video Rate Adaptation Techniques
Various video rate control or adaptation techniques have been proposed to combat short- term variations in homogeneous VBR wireless channels when performing video streaming.
Specifically for streaming pre-encoded video, research focus has been put on transcod- ing [22, 3, 11], joint source/channel coding [12, 14, 5], scalable video coding [9], and other techniques such as content-aware or motion-aware coding [18, 7, 25, 27]. While in commercial systems, the multiple file/rate switching techniques are widely implemented
[2].
These rate adaptation proposals work well in homogeneous wireless network where average ABR doesn’t vary over time. However, they could not provide satisfactory per-
7 Chapter 2. Literature Review 8 formance in terms of quality variation in heterogeneous wireless networks, since purely adapting the source bit rate to the channel bit rate will lead to a large variation of video quality over different sub-networks.
2.1.1 Transcoding
Transcoding is a technique to adapt the video source rate through recompression. [22,
3, 11] are three examples of research studies on video rate adaptation with transcoding.
These techniques dynamically choose the quantizer used in encoding each frame or block, and try to minimize the total distortion while matching the video source rate with the network rate. Their focus is mainly on analyzing the specific encoding technique and extracting the rate-distortion models. The heavy computation involved in transcoding is its main disadvantage.
2.1.2 Joint Source/Channel Coding
The authors of [12] show that the perceptual source distortion decreases exponentially with the increasing MPEG-2 source rate, and the perceptual distortion due to data loss is directly proportional to the number of lost macro blocks. Hence they propose to use Joint Source/Channel Coding (JSCC) technique, specifically adding FEC (forward error correction) bits, to protect the data from loss. The optimal channel coding FEC parameters can be selected according to the aforementioned relationships and the total rate of received video stream can be controlled to minimize the total distortion.
Similarly, [14] and [5] both consider the Joint Source/Channel Coding problem with
FEC channel coding and focus on how to choose the channel coding parameters. In [14], the authors translate the Quality of Service (QoS) requirements of the video streaming applications into a threshold of occupancy of playback buffer. By adapting the JSCC parameters their scheme tries to maintain a certain level of buffer occupancy to sustain continuous playback. Chapter 2. Literature Review 9
In [5], a probabilistic QoS requirement, i.e. the buffer starvation probability has been proposed. The authors use cycle-based rate control with cycles being successively alternating between good(non-fading) and bad(fading) period, while guaranteeing an upper bound on the probability of starvation at the playback buffer. The cycle-based idea inspired us to divide the streaming process in heterogeneous networks into cycles, but our “cycle” have a completely different definition from theirs in that our cycle contains intervals of the MT residing in different sub-networks.
While the Joint Source /Channel Coding technique can adapt the source rate within
certain range, it usually involves cross-layer design with information flows across PHY
/MAC /Network layers, which might be applicable in homogeneous networks but could
become extremely complex in terms of implementation and computation in heterogeneous
networks.
2.1.3 Scalable Video Coding
The layered or scalable video coding techniques are said to be suitable for adapting
to longer-term bandwidth fluctuations [2]. There has been a substantial body of re-
search works developing efficient scalable compression techniques. A scalable extension
of H.264/AVC [31], Scalable Video Coding (SVC) [24] has been standardized, which pro-
vides scalability of temporal, spatial, quality resolution, or a combination of scalability on
these three dimensions, of a decoded video signal through adaptation of the bit stream.
Fine Granularity Scalability (FGS) coding [17] and Fine Granularity Scalability Tempo-
ral (FGST) coding [29] have also been adopted as amendments to the MPEG-4 standard.
Multiple Description Coding (MDC) [13] is another type of scalable video coding where
each description (substream) of the video stream is of equal weight and independent of
each other in contrast to the Base Layer/Enhancement Layer structure of SVC.
The authors of [10] developed a heuristic rate control algorithm for 2-layer FGS coded
video over TCP-friendly “connection”, which can achieve the same level of smoothness Chapter 2. Literature Review 10 over both TCP and TCP-friendly protocols. Their algorithm works with CBR coded video and the loss model is simple. The authors of [16] proposed a stochastic dynamic programming algorithm for VBR scalable coded video with a more realistic loss model.
The authors of [9] studied the problem of minimizing the average distortion of FGS
video under a limited transmission rate. The authors provided a framework which jointly
considers the effects of packet scheduling at the sender and the error concealment at the
receiver.
The authors of [33] explicitly considered the effect of fading in wireless channel and
develops cross-layer rate adaptation algorithm for layered video in fading channels. The
complexity in cross-layer design makes it difficult to implement even in homogeneous
wireless network.
The authors of [23] introduced a novel streaming strategy to improve the probability
of successfully stream a scalable coded video sequence by adaptively selecting the number
of layers according to mobility information in Ad-Hoc wireless networks. This is relevant
to our work in that our proposed algorithm can also utilize the mobility and location
information to predict the MT’s movement and channel status, as described in Chapter
5.
In our control scheme, we can use either the transcoding or scalable coding techniques
to perform rate adaptation. However, we assume a generic rate-quality relationship in
our model and that one cannot change the quality of a video sequence that is already
transmitted.
2.1.4 Content-Aware Coding Techniques
There exists many other video rate adaptation algorithms which try to achieve different
QoS objectives. Content-aware encoding/playout has been an interesting and contro-
versial topic for video rate adaptation, as there exists no generally accepted standard
for perceived quality of motion pictures when we consider the presentation of the actual Chapter 2. Literature Review 11 content, instead of quantifiable metrics such as resolution, frame rate and PSNR (Peak
Signal-to-Noise Ratio). Representative works of content-aware rate adaptation/playout control include [18, 7, 25, 27], etc. These works try to analyze the amount of motion or interested objects in each frame, and allocate the available bit rates unfairly among different objects / frames to achieve best perceived quality when the network rate is not high enough to present the full pictures.
2.2 Rate Control in Heterogeneous Wireless Net-
works
Video streaming in heterogeneous wireless networks has been a relatively new topic. Most of the available works address the issues in architecture design and hand-off handling.
The authors of [26] analyzed the effects of handoffs on rate control and proposed a cross- layer solution to anticipate the handoff occurrence and to adjust the data rate. They use transport-layer dummy packets to probe the channel in their solution, while in this thesis we propose to utilize the statistical information of the residence times and ABR in each sub-network.
Some researchers consider video streaming in a multiple stream environment with
heterogeneous access technologies and focus on fairness or priority among all users/flows.
In [38] the authors study the rate allocation problem in streaming over wireless networks
with heterogeneous link speeds. The focus of their work is on how to allocate the rate
between multiple video streaming sessions on heterogeneous links to maximize the aver-
age quality among all users, while the quality enjoyed by a single user is not explicitly
considered.
The authors of [1] addressed the problem of flow rate control for different types of
traffic flows and heterogeneous wireless links, and employs an H-infinity optimal rate
controller to achieve efficient utilization of all channels while taking the requirements of Chapter 2. Literature Review 12 different flow types in to account. Both of them assume that all the access networks in the integrated network are available all the time, while our assumption is that the user is moving and the trajectory is not always covered by both sub-networks.
The authors of [35, 36, 37] considered the video streaming problem in an integrated
3G/WLAN network from a monetary cost point of view. Their system setting of the heterogeneous networks is the most similar one to ours, yet the objective and control actions are completely different. Different streaming strategies are proposed to decide how much data to be streamed (i.e. the transmission rate) in each individual network as well as when to hand off to the other network, so that the monetary cost of streaming the data is minimized. While in our system model we also consider the cost effects of each network, we assume that it is always good to fully utilize the ABR and that the MT will switch to the higher-rate, lower-cost network whenever it is available, as this strategy is simple and already implemented in commercial systems such as the iPhone.
It is worth mentioning that, some of the techniques mentioned above, such as [33] and
[5] have similar objectives as ours, i.e. minimizing the variations in adapted video rate caused by variations in network transmission rate. However, in homogeneous wireless networks, the variation caused by fading and other short-term effects are quite different from the variation caused by handoff between different access technologies in both time scale and magnitude. Hence these techniques cannot be directly applied to our problem.
Furthermore, because the time scale of variation in our problem is longer, we may have more ways to predict the variation, such as utilizing geographical information.
2.3 Buffering in Heterogeneous Wireless Networks
For streaming pre-encoded video, buffering is another technique to overcome the mis- match between video source rate and channel bit rate. By caching enough data in the client buffer ahead of time, continuous high-rate playback can be sustained when the Chapter 2. Literature Review 13 channel throughput is low. Buffering schemes for streaming VBR video over heteroge- neous wireless networks are studied in [15]. These schemes include fixed/jointly optimized schemes based on buffering delay, buffered playout data, and playout time. Analysis on both the jitter frequency and the buffering delay are conducted for these schemes.
However, without rate adaptation buffer underflow happens frequently when the av- erage channel throughput is lower than the average video source rate, leading to playback jitters. Hence we propose to combine rate adaptation with buffering for long-term varia- tions to smooth out the streaming process in heterogeneous networks. We choose only the playout time based buffering scheme, as when we introduce rate adaptation the buffering delay become meaningless, and the buffered playout data become highly variable with the changing video rate.
To the best of our knowledge, this thesis is the first work to propose the combination of rate adaptation and buffering to address the problem of smooth video playback in heterogeneous networks. Nevertheless, our work is inspired by and based on the related works listed here in that the scheme shall utilize a generic rate adaptation techniques mentioned above to perform the control actions. Chapter 3
Problem Statement
In this chapter, we explain in details the problem we study in this thesis. We first present the application scenario, then introduce our way modeling of the system based on this scenario with some practical assumptions. A mathematical formulation of our problem is then presented based on these common assumptions.
3.1 Application Scenario
We study video streaming over the heterogeneous wireless networks with the overlapping of two networks (or two zones), as shown in Figure 3.1. The Tier-1 Network (T1N) is assumed to provide universal coverage with low bit rate, while Tier-2 Network (T2N) covers limited areas around the Access Points (APs), with high bit rate. In reality, T1N is usually more costly than T2N. A proper example of T1N can be the 3G cellular network, while T2N can be Wireless LAN. A user prefers to access the Internet through T2N due to its high bandwidth and low cost, thus whenever he/she enters a T2N covered area, the mobile device switches to T2N for transmission. While our scheme is independent of lower layer (e.g. PHY/MAC layer) implementations and can actually handle the simultaneous transmission over both sub-networks, we maintain the assumption of using only one sub-network at the same time since it is more practical to do so in reality.
14 Chapter 3. Problem Statement 15
T1N T2N
AP
BS T2N
Figure 3.1: Integrated two-tier network
Two types of handoffs take place in this network: intra-technology handoff or Hori- zontal Handoff (HHO) in which the mobile terminal (MT) switches between two Access
Points (AP) or Base Stations (BS) using the same access technology, e.g. from one
WLAN AP to another, and inter-technology handoff or Vertical Handoff (VHO), which occurs when the MT roams between different access technologies, e.g. switching from 3G to WLAN when entering a WLAN covered area. VHO affects different system perfor- mance metrics, such as the signaling load, resource utilization and user perceived QoS. In particular, the available bit rate (ABR) in our model may vary by one order of magnitude after any VHO. Both types of handoffs may cause extra delays in the transmission, but we do not consider the extra delays here and assume that there is a seamless handoff handling scheme which can eliminate the delays caused by both types of handoffs (which may be achieved at the cost of ABR). In practice, a handoff handling scheme such as the one presented in [6] can be employed to satisfy this assumption.
A video streaming session is running on the device while the MT traverses through this integrated two-tier network. The streaming server lies outside this wireless network, but the bottleneck of the connection is the last hop - the wireless link between the
MT and the BS/AP. The MT keeps sending control messages to request the server to adjust the video source rate and transmission rate. The server then makes adjustments accordingly and transmits the data to the MT. The MT has a buffer, which stores the Chapter 3. Problem Statement 16 received data before they are used for playback. Our goal is then to develop a control scheme to determine how to choose the video source rate and how much to buffer ahead of time given some statistical and observed channel information. The streaming session is assumed to be very long and we analyze the performance of our scheme on a time-average basis.
3.2 Models and Assumptions
3.2.1 Rate Adaptation and Playback
We model the variation in the wireless channel, the error control scheme, and the handoff handling effects all into the random ABR of the network, denoted by R(i), which is always positive. In reality, ABR is the amount of error-free video data received by the MT at each time slot.
We divide the whole streaming session, which consists of several “in-T1N” and “in-
T2N” intervals, into cycles, and denote each cycle by its sequence number in the whole streaming session, j, where j = 1 represents the first cycle in this session. Each cycle
starts at a T1N-to-T2N VHO and contains one “in-T2N” interval followed by one “in-
T1N” interval, the lengths of which are denoted as T2 and T1 respectively. (Note that, here we assume the streaming session always starts in T2N, i.e. the high rate network. This is
reasonable since there is not much we can optimize before the first VHO if it starts within
T1N, and we are considering the average performance in the whole streaming process,
so edge effect at the beginning can be ignored.) We further model the video streaming
process as time-discrete with time slots of equal length, and denote each time slot by its
sequence number in the current cycle, i, where i = 1 represents the first time slot in
the current cycle. The control actions are decided and performed at the beginning of
each time slot.
We assume that we have original video streams with very high quality. The average Chapter 3. Problem Statement 17 rate of video can be higher than the highest network rate, and we can adjust the source encoding rate to any level at any granularity up to the original rate. This assumption of rate adaptation at any granularity can be accommodated by quantization in practice. By making this assumption, we eliminate the possibility of playback stop, since transmission rate R(i) is assumed to be always positive.
The original source rate of the whole video sequence as well as the rate-quality rela- tionship of the video are transmitted to the MT before the streaming starts, thus the MT can utilize this information to adjust the rate of “future” parts of the video which are being streamed. Here we do not define the specific technique used to adapt the source rate. The transcoding or SVC technique mentioned in Chapter 2 may be employed in reality to perform rate adaptation.
It is worth noting that, although control actions are performed at the beginning of each time slot, it is not necessarily effective for only one time slot in playback time.
This is because control decisions are made at every time slot of transmission time, in which period more than one time slot of video data in the playback sequence may be transmitted. We illustrate the relationship between transmission time and playback time in Figure 3.2.
We use a “quality level” defined as q(k) = f(r(k), r0(k)) to control the source rate adaptation at time slot k, where f(x, y) is monotonously increasing with x, r0(k) is the original video rate at time slot k, and r(k) is the adapted video rate at time slot k. Hence for each time slot, the perceived image quality increases with q(k). While there are several methods to characterize the quality of the picture, e.g. PSNR, distortion, etc, we do not choose a specific metric, but assume a generic form of the rate-quality relationship. For simplicity, we assume the original video stream is constant-quality encoded (whether it is CBR or VBR), thus the perceived quality only depends on the adjusted bit rate and the original bit rate, and q(k)= f(r(k), r0(k)) = f(r(k)).
0