Rate Control of Real-time MPEG-2 over ATM DBR Service with Bandwidth Re-negotiation

Similar documents
different problems from other networks ITU-T specified restricted initial set Limited number of overhead bits ATM forum Traffic Management

On TCP-friendly Video Transfer

TCP-friendly video transfer

Traffic Management Tools for ATM Networks With Real-Time and Non-Real-Time Services

Quality Control Scheme for ATM Switching Network

Comparison of Shaping and Buffering for Video Transmission

SIMULATION OF PACKET DATA NETWORKS USING OPNET

Testing Policing in ATM Networks

ATM Quality of Service (QoS)

Intermediate Traffic Management

Integrating Euro-ISDN with ATM Technology : Interworking Mechanisms and Services Support

Lecture 4 Wide Area Networks - Congestion in Data Networks

BROADBAND AND HIGH SPEED NETWORKS

ATM. Asynchronous Transfer Mode. (and some SDH) (Synchronous Digital Hierarchy)

A Measurement-Based CAC Strategy for ATM Networks

Designing Efficient Explicit-Rate Switch Algorithm with Max-Min Fairness for ABR Service Class in ATM Networks

Master Course Computer Networks IN2097

Compressed Video Streams: Network Constrained Smoothing

A Measurement-Based CAC Strategy for ATM Networks

Decoupling QOS Guarantees and Connection Establishment in Communication Networks.

Real-Time Protocol (RTP)

QOS in ATM Networks. Traffic control in ATM networks. Layered model. Call level. Pag. 1

Lecture 17 Multimedia Transport Subsystem (Part 3)

End-to-End Delay Analysis of Videoconferencing over Packet-Switched Networks

QoS Policy Parameters

Quality of Service (QoS)

Performance and Evaluation of Integrated Video Transmission and Quality of Service for internet and Satellite Communication Traffic of ATM Networks

Congestion in Data Networks. Congestion in Data Networks

Advanced Computer Networks

Standardizing Information and Communication Systems

Improving QOS in IP Networks. Principles for QOS Guarantees

QUALITY of SERVICE. Introduction

FAIR DELAY OPTIMIZATION-BASED RESOURCE ALLOCATION ALGORITHM FOR VIDEO TRAFFIC OVER WIRELESS MULTIMEDIA SYSTEM

Master Course Computer Networks IN2097

Traffic control in ATM networks

Simulation Study for a Broadband Multimedia VSAT Network

Master Course Computer Networks IN2097

UBR Congestion controlled Video Transmission over ATM Eltayeb Omer Eltayeb, Saudi Telecom Company

Quality Differentiation with Source Shaping and Forward Error Correction

Multimedia Networking. Network Support for Multimedia Applications

Performance Characteristics of a Packet-Based Leaky-Bucket Algorithm for ATM Networks

SIMULATION OF PACKET DATA NETWORKS USING OPNET

NEW TRANSFER METHODS FOR CLIENT-SERVER TRAFFIC IN ATM NETWORKS

11. Traffic management in ATM. lect11.ppt S Introduction to Teletraffic Theory Spring 2003

What Is Congestion? Computer Networks. Ideal Network Utilization. Interaction of Queues

Unit 2 Packet Switching Networks - II

Lecture Outline. Bag of Tricks

ATM Traffic Management

Congestion Control Open Loop

ATG s Communications & Networking Technology Guide Series This guide has been sponsored by

CS 457 Multimedia Applications. Fall 2014

Network management and QoS provisioning - QoS in ATM Networks

Introduction to ATM Traffic Management on the Cisco 7200 Series Routers

Network Support for Multimedia

Which Service for TCP/IP Traffic on ATM: ABR or UBR?

Comparative Study of Partial Closed-loop Versus Open-loop Motion Estimation for Coding of HDTV

A transport-layer approach for achieving predictable throughput for Internet applications

Traffic and Congestion Control in ATM Networks Using Neuro-Fuzzy Approach

STATISTICAL PROPERTIES OF MPEG STREAMS AND ISSUES FOR TRAMSMISSION OF VIDEO INFORMATION IN HIGH SPEED NETWORKS

Quality of Service (QoS)

11. Traffic management in ATM

From ATM to IP and back again: the label switched path to the converged Internet, or another blind alley?

DiffServ Architecture: Impact of scheduling on QoS

Proxy Caching Mechanisms with Quality Adjustment for Video Streaming Services

TELE Switching Systems and Architecture. Assignment Week 10 Lecture Summary - Traffic Management (including scheduling)

THE majority of traffic in broad-band networks is likely

MPEG4 VIDEO OVER PACKET SWITCHED CONNECTION OF THE WCDMA AIR INTERFACE

Network Model for Delay-Sensitive Traffic

Traffic Management. Service Categories CHAPTER

Performance Analysis of Cell Switching Management Scheme in Wireless Packet Communications

A Framework for Video Streaming to Resource- Constrained Terminals

Network Layer Enhancements

Performance of a Switched Ethernet: A Case Study

T-Spec Examples for Audio-Video Bridging. Osama Aboul-Magd

A hybrid medium access control for convergence of broadband wireless and wireline ATM networks

PROPOSAL OF MULTI-HOP WIRELESS LAN SYSTEM FOR QOS GUARANTEED TRANSMISSION

Traffic Management. Service Categories CHAPTER

Lecture 17: Distributed Multimedia

Illustration of The Traffic Conformance in ATM Network

ATM Hierarchical Shaping ATM VC into VP Shaping, page 1

3. Quality of Service

International Journal of Scientific & Engineering Research, Volume 7, Issue 1, January ISSN

Understanding the Variable Bit Rate Real Time (VBR rt) Service Category for ATM VCs

What Is Congestion? Effects of Congestion. Interaction of Queues. Chapter 12 Congestion in Data Networks. Effect of Congestion Control

9. D. Tse, R. Gallager, and J. Tsitsiklis. Statistical multiplexing of multiple timescale

Traffic Access Control. Hamid R. Rabiee Mostafa Salehi, Fatemeh Dabiran, Hoda Ayatollahi Spring 2011

Topic 4b: QoS Principles. Chapter 9 Multimedia Networking. Computer Networking: A Top Down Approach

ATM Logical Connections: VCC. ATM Logical Connections: VPC

Chapter 1 Introduction

Lecture 4 Wide Area Networks - Asynchronous Transfer Mode

Performance Analysis & QoS Guarantee in ATM Networks

Converged Networks. Objectives. References

Experiments Involving the Transmission of Layered Video over a Local ATM Network

A Method for Traffic Scheduling Based on Token Bucket QoS Parameters

Introduction to IP QoS

EVALUATION OF THREE CAC METHODS: GAUSSIAN APPROXIMATION METHOD, METHOD OF EFFECTIVE BANDWIDTH AND DIFFUSION APPROXIMATION METHOD

Reduction of Periodic Broadcast Resource Requirements with Proxy Caching

Multimedia Networking

CHAPTER 3 EFFECTIVE ADMISSION CONTROL MECHANISM IN WIRELESS MESH NETWORKS

Performance of UMTS Radio Link Control

Transcription:

Rate Control of Real-time MPEG- over ATM DBR Service with Bandwidth Re-negotiation Kentarou Fukuda Naoki Wakamiya Masayuki Murata Hideo Miyahara Department of Informatics and Mathematical Science Graduate School of Engineering Science Osaka University Toyonaka, Osaka 5, Japan Abstract The rate control method of MPEG- over the ATM DBR service class is proposed in this paper. It is intended to guarantee the Quality of Service required by video applications. The disadvantages of MPEG- Test Model 5 are resolved by introducing the bandwidth re-negotiation with the network. And user-oriented high quality video transfer can be guaranteed with our method. 1 Introduction A video transfer is now becoming a major application of high speed networks. Compared with a computer data transfer application such as ftp, a video transfer requires more strict Quality of Service (QoS). Although some data losses may be tolerated, the transfer delay must be kept small throughout the session especially in an interactive video application, e.g. video conference. To guarantee the required QoS of the video transfer over ATM networks, several service classes have been defined in the ATM standard, SBR (Statistical Bit Rate), DBR (Deterministic Bit Rate), ABR (Available Bit Rate) and UBR (Unspecified Bit Rate) service classes [1]. The ABR service class and the UBR service class are defined for computer data transfer. Only cell loss ratio is guaranteed in the ABR service class, and in the UBR service class no QoS is supported. The SBR (Statistical Bit Rate) service class is for a variable-bit-rate (VBR) data transfer and the statistical multiplexing is performed to gain a high degree of multiplexing. The video traffic has a bursty nature, and has been considered to be suited to the SBR service class. As an example, we depict traffic rate variation of a sample video sequence in Fig. 1, where MPEG- (Moving Picture Experts Group) coding algorithm [] is employed. In Fig. 1, the rate for each (Group of Pictures) is also shown with a solid line. MPEG was originally developed for the compressed video storage, but is useful enough for multimedia applications and is now widely accepted [3]. In the SBR service class the statistical multiplexing is performed based on the traffic characteristics reported from a sending end station to the network at the call setup time. More specifically, the sending end station first declares its traffic descriptors such as PCR (Peak Cell Rate), SCR (Sustainable Cell Rate) and BT (Burst Tolerance) to the network [1]. It also reports the required QoS-related parameters, by which the application can give the preferable QoS to users. If the network can accept this connection setup request without violating existing other connections, the new connection is set up. Then, the traffic the sending end station emits is tested whether it is conformed to the reported traffic descriptors, which is so called UPC (Usage Parameter Control). With this statistical multiplexing technique, it is expected that network resources can be effectively used. Although some cells would be lost within the network during the session, it can be tolerable to the video data transfer. However, there remains several problems to support a video transfer over the SBR service class. As described above, a complicated CAC (Call Admission Control) process of the SBR service class requires users to report traffic descriptors at the connection setup time. However, it is difficult to predict accurate traffic characteristics including peak rate, rate and burst tolerance of the video in advance. Especially when the video is encoded in a realtime fashion, it becomes almost impossible because traffic characteristics heavily depends on the encoding algorithm and video contents. It is also difficult to set the appropriate UPC parameters such as the bucket size and the drain rate of the leaky bucket [] for the traffic conformance test. If these parameters are inappropriately chosen, the resulting traffic through the leaky bucket would have different characteristics from the original one and it would introduce unacceptable cell losses and delays at UPC and within the network. The DBR (Deterministic Bit Rate) service class is for the continuous bit rate traffic. In this service class, the band-

rate (Mbps) 1 1 1 maximum minimum Figure 1: MPEG video traffic (the sequence Scenery) rate (Mbps) 1 1 1 maximum allocated minimum Figure : MPEG- video traffic (Scenery) width is allocated to a session based on the single traffic descriptor, PCR (Peak Cell Rate). The peak rate bandwidth allocation can provide a simple CAC and the high quality of data transfer; no cell loss and the tightly bounded cell transfer delay as long as the cell emission rate is below the allocated bandwidth. This feature of the DBR service class can guarantee the high quality video transfer over the ATM network. However, the burstiness of video traffic causes the ineffective bandwidth usage. As shown in Fig. 1, when the bandwidth equivalent to the peak rate of video traffic is allocated (1.3 Mbps in this case), about two thirds of the allocated bandwidth is unused and wasted. In the above example, the quantization scale is fixed at. The cell generation rate is then varied according to the video content, which we will refer to a VBR encod- SNR (db) 5 35 3 5 15 5 VBR with Q limitation Figure 3: MPEG- video quality variation (Scenery) ing scheme. Another way to encode the video is a CBR encoding scheme, where cell generation rate is kept constant by controlling the video quality. This scheme seems to be suitable to the ATM DBR service class. An MPEG- Test Model 5 () [5] is a well known example to control the MPEG- traffic generation rate. By adjusting macroblock quantization and controlling the spatial resolution, the video traffic rate d over a GOP (Group of Pictures) can be kept lower than the allocated bandwidth (Fig. ). However, there are two major problems in the rate control algorithm. First, non-uniformity of the quantizer scale Q within each picture degrades the perceived video quality. The quantizer scale for each macroblock is fully dependent upon the virtual buffer fullness, which changes macroblock by macroblock. Hence, quality of all macroblocks in the picture is not treated equally because of the variations in the buffer fullness. This results in the non-uniform picture quality. Another problem in rate control algorithm is that it cannot handle scene changes properly. Figure 3 compares the picture quality variation of MPEG- video in terms of SNR; the label VBR corresponds to the VBR encoding scheme where the quantizer scale is fixed at. As shown in Fig. 1, scene changes occurring around 53th and 133th pictures have great influences on the cell generation rate in the case of the VBR encoding scheme. However, SNR is not much degraded even during scene changes (Fig. 3). On the other hand, in, the rate information of the previous picture is used for estimating the current picture rate. Then, the amount of allocated bits to the new picture is likely to be unchanged. However, the required bandwidth (a target bandwidth according to terminology) is much increased when the scene changes from simple to compli-

cated pictures. It then results in the quality degradation of the new scene as shown in Fig. 3 (the case of is labeled as ). In depicting this figure, we set the allocated bandwidth at.9 Mbps which is equivalent to the rate to a VBR case (Fig. 1). When the scene changes from complicated to simple ones, on the other hand, unnecessarily enough bandwidth is provided. One may think that the problem of can be resolved by introducing a limitation on a range of a picture quality variation, i.e., a range of a quantizer scale variation. As shown in Fig. 3 (labeled as with Q limitation ), quality degradation in terms of SNR can be sustained. However, when we limit the range of Q according to the required video quality, the rate control algorithm cannot perform properly and the resulting traffic generation rate goes beyond the allocated bandwidth as shown in Fig.. To avoid this, we do not reduce the coding rate (as in ), but increase the allocated bandwidth by the bandwidth renegotiation with the network in this paper. The bandwidth re-negotiation can be performed through the signaling protocol []. Or, we can utilize a Fast Reservation Protocol [or ABT (ATM Block Transfer)] [7] for fast bandwidth renegotiation. Note that in the case of FRP, the concept of burst cannot be applied to our case and we may need some modification to FRP. A further study is required about the implementation aspect of such a variant of FRP. In this paper, we propose the rate control method which solves the problems of and guarantees the required video quality. And with our proposed method, the effective bandwidth usage can be performed. This paper is organized as follows. In Section, we present our rate control algorithm which is an extension of rate control. Evaluation results are then given through comparing with the original in Section 3. In Section, we give the brief introduction and evaluation of delay control mechanism on our proposed method. Finally, we conclude our study in Section 5. MPEG- transfer over ATM with bandwidth re-negotiation In this section, we propose a rate control method based on MPEG- suitable to the ATM environment. Our method incorporates the bandwidth re-negotiation protocol available in the ATM networks. As having been shown in Fig., when we apply the limitation, Q in the MPEG- encoding algorithm, the resulting traffic generation rate sometimes goes beyond the allocated bandwidth, and the excess traffic finally causes the cell buffer overflow at the sending end station. As shown in Fig. 5, the maximum queue length of the coding scheme with Q limitation is about one hundred times as large as that of original. To avoid this queue length explosion, we do not reduce the coding rate (as in ), but increase the allocated bandwidth by the bandwidth re-negotiation with the network. The bandwidth re-negotiation can be performed through the signaling protocol [] or Fast Reservation Protocol (FRP) [or ABT (ATM Block Transfer)][7]. Note that in the case of FRP, the concept of burst is not applied in our case and we need some modification to FRP. Namely, the use of RM cells should be extended to allow the rate changes. If on-the-fly FRP signaling is employed, the rate changes can be attained in the very short time. However, the further study is required about the implementation aspect of such a variant of FRP. rate (Mbps) 1 maximum allocated 1 1 minimum Figure : MPEG- video traffic in with Q limitation (Scenery) Queue length of I picture (kbits) with Q limitation Figure 5: Queue length transition in with and without Q limitation (Scenery)

MPEG- Encoder Q information Sending Station Variable Bit Rate Buffer Bandwidth Controller Constant Bit Rate Bandwidth Re-negotiation Figure : Rate controller model Data Signal Network rate (Mbps) Scenery Starwars Live Comedy R(Q) There are some researches on the video transfer over the bandwidth reservation based networks where QoS requirement of application varies during a session [, 9,, 11]. For example in [11], they also propose the mechanism to provide the QoS guarantee by introducing bandwidth renegotiation and quantizer scale control. However, received video quality fluctuates and the effective bandwidth usage cannot be expected. Our method, together with, can utilize most of the allocated bandwidth for video data and achieve the highly effective bandwidth usage. In Subsection.1, we introduce the rate control method and then a control part of our method, i.e., determination method of the requested bandwidth in re-negotiation is discussed in detail in Subsection.. The quality of the proposed rate control method is evaluated in Section 3..1 Bandwidth re-negotiation mechanism The proposed model is shown in Fig.. Solid lines in the figure stand for cell stream and dashed lines control signals. An MPEG- encoder in the figure is an MPEG- Test Model 5 video encoder. A coded video stream is stored in the buffer and UPC regulates the cell emission rate according to the allocated bandwidth. There are two alternatives in monitoring the bandwidth usage to see whether the allocated bandwidth is sufficient or not. The first approach is to utilize buffer occupancy information. When the cell queueing goes over the predefined threshold, the rate controller perceives the shortage of the bandwidth and begins to negotiate the bandwidth re-allocation with the network. This buffer based approach seems effective to avoid the buffer overflow. However, the queue length varies with different time ranges, i.e., macroblock by macroblock, picture by picture and scene by scene. Thus, it would be difficult to decide an appropriate threshold value on the queue length. Another approach is to use the Q values within a picture. If the bandwidth is insufficient, the rate control algorithm reduces the amount of coded bits by applying the larger quantizer scale Q. Thus, when the large portion within the picture are coded with large Q, it indicates 1 5 15 5 3 35 quantizer scale Q Figure 7: Q-R relationship that the allocated bandwidth is insufficient. Namely the bandwidth controller re-negotiates to increase the allocated bandwidth when the Q within a picture goes over the pre-defined threshold. The most influential picture type is I pictures and next P. These pictures have relatively large sizes and sensitive to the scene changes. Thus, the quantizer scale of I and P pictures is monitored in our control method. It is then sent to the bandwidth controller so that the bandwidth controller can decide whether the allocated bandwidth should be re-allocated or not. When the bandwidth is re-negotiated successfully, the amount of newly allocated bandwidth is informed to UPC and its conformed rate is changed.. Determination of required bandwidth In our method, the application first decides the required video quality by the largest quantizer scale Q max which the application can tolerate. Or, the application may use the lowest video quality in the form of SNR instead of Q max. In either way, the MPEG- controller decides the amount of required bandwidth to satisfy the requested video quality. Then, the connection setup is requested to the network. In determining the required bandwidth, we may employ the Q-R relationship function R(Q). In Fig. 7, we show the relationships between the quantizer scale and the traffic generation rate of three sample video sequences. As expected, the rate is decreased as the quantizer scale Q is set to be larger. For example, we can change the rate from Mbps to 1. Mbps by changing the quantizer scale from 1 to in the Scenery sequence, at the cost of video quality. It should be noted here that increasing quantizer scale beyond 3 is little useful, because essential information such

Table 1: Traffic characteristics of test videos Name Mean Maximum PtoA Scenery Size 17 Kbits 57 Kbits 3. Rate 5. Mbps 17. Mbps Starwars Size 5 Kbits 3 Kbits.11 Rate.15 Mbps.9 Mbps Live Size 99 Kbits 35 Kbits 3. Rate.97 Mbps.7 Mbps Comedy Size 3 Kbits 33 Kbits.13 Rate 5.95 Mbps.7 Mbps as headers give limitations on further rate reduction. In the case of the video sequence Scenery, the following equation can be derived using fit function of Mathematica [, 13]; 3:7 R(Q) = :7 + + :5 (1) Q Q It is true that Q-R relationship varies video by video, but they show similar characteristics as shown in Fig. 7. Thus, this equation R(Q) may be applied to other video sequences. In our method, we adjust R(Q) function to the actual video traffic by controlling the first terms of Eq. 1. Through simulations, we have revealed that about 9% of the quantizer scales used in the video sequence coded by are within the range up to the quantizer scale plus. We may use this characteristics of quantizer scale distribution. For example, to satisfy the maximum quantizer scale Q max =, the sending end station requests the network to be allocated the bandwidth R() such that the quantizer scale Q =can be obtained. 3 Simulation results In this section, we evaluate the effectiveness of our proposed rate control method for the MPEG- video transfer with the rate control algorithm. As described in Section, we can provide guaranteed QoS to users by giving the limitation on the range of the quantizer scale variation. However, as shown in Figs. and 5, the quantizer scale limitation causes the cell generation beyond the allocated bandwidth and the queue length increase. To avoid this, we introduce the rate control algorithm with the bandwidth renegotiation method. We employ four MPEG- video sequences for quality evaluation, Scenery, Starwars Live and Comedy. Each video sequence is seconds long and consists of pictures. We should note that we included the magnified figure in this paper to see the detailed behavior. Each picture consists of x pixels, i.e., macroblocks. A structure is IBBPBB. There traffic characteristics are summarized in Table 1. rate (Mbps) 1 maximum allocated 1 1 minimum Figure : MPEG- video traffic in with re-negotiation (Scenery) quantizer scale of I,P picture 3 5 15 5 Figure 9: MPEG- quantizer scale variation in with re-negotiation (Scenery) 3.1 Basic property of control method In Fig., the resulting traffic generation rate of Scenery is shown where the bandwidth re-negotiation is employed. In this figure, we assume that the bandwidth re-negotiation is completed in a time length of a macroblock, i.e., 1/3 msec. Note that one may think 1/3 msec re-negotiation time is too short. It is perhaps true and we will investigate this aspect later. Our intention of this parameter setting is to explore the basic property of our method. For this we assume that there is no possibility of the bandwidth re-negotiation failure in evaluations. The required QoS is Q max in these figures. In obtaining Figs. through

SNR (db) 5 35 3 5 15 5 with renegotiation Figure : MPEG- video quality variation in with re-negotiation (Scenery) allocated rate (Mbps) 1 1 delay delay 1 Figure : Comparison of amount of allocated bandwidth in with bandwidth re-negotiation (Scenery) Queue length of I picture (kbits) with renegotiation Queue length of I picture (kbits) 1 delay delay 1 Figure 11: Queue length transition in with and without bandwidth re-negotiation (Scenery) 13, we have used and for upper and lower threshold values to invoke the bandwidth re-negotiation. We will investigate this aspect later. The allocated bandwidth is shown with the solid line and the actual bit rate is shown with the dashed line (labeled as ). Comparing Fig. with Fig. 1, it is shown that the coded bit rate with our method has a bursty nature and the enough bandwidth is allocated to a scene-changing picture, while the scene changing is not treated properly in (Fig. ). The allocated bandwidth over the entire sequence is almost same in both and our method (Figs. and ). The required QoS is guaranteed as shown in Fig. 9. The quantizer scale is kept under through an entire sequence as Figure 13: Comparison of queue length in with bandwidth re-negotiation (Scenery) requested by users. As a result, the video quality is kept more uniform and higher than in as shown in Fig. In Fig. 11, the queue length transition in our method is shown together with original. As shown in Fig. 11, the queue length is slightly increased by about Kbits (5 cells; about msec). 3. Effect of large re-negotiation delay In Figs. through, we have assumed that it takes one macroblock time (i.e., 1/3 msec) for the network to perform the bandwidth re-negotiation. It means that the required bandwidth is allocated successfully 1/3 msec

rate (Mbps) 9 7 5 3 fast up-fast down fast up-normal down normal up-normal down slow up-slow down 3 5 7 Figure 1: Allocated bandwidth comparison for four combination of threshold values (Scenery, magnified) later than the emission of RM cell for the bandwidth renegotiation. In Figs. and 13, we next investigate the effect of the bandwidth re-negotiation delays. The bandwidth re-negotiation delays include the propagation delay and the processing delay of the RM cell. In those figures, we compare two cases where the processing delay is one macroblock time (1/3 msec, 1/ picture time) with, macroblock time (5/1 sec msec, about picture time). As shown in Fig., the bandwidth allocation is delayed in the case of, macroblock renegotiation delay. The delayed bandwidth allocation results in slightly different traffic characteristics, because the cell generation rate depends on the amount of the allocated bandwidth in. As a result, the queue length of the case of, macroblock re-negotiation delay is larger than that of one macroblock delay as shown in Fig. 13. The maximum queue length is about 1.5 Mbits (3,7 cells) when the bandwidth re-negotiation delay is, macroblock time. 3.3 When the bandwidth should be re-negotiated? It is important to decide the appropriate timing of the bandwidth re-negotiation. As already described, quantizer scale within each picture is monitored to detect the shortage or surplus of the allocated bandwidth. When the values of Q within the picture goes over the predefined threshold value, Q up, the bandwidth controller begins to re-negotiate the bandwidth with the network. In the case where the network load is high, the bandwidth re-negotiation may fail and the sending end station should Queue length (kbits) 7 5 3 fast up-fast down fast up-normal down normal up-normal down slow up-slow down 3 5 7 Figure 15: Queue length comparison for four combination of threshold values (Scenery, magnified) retry the bandwidth re-negotiation. Therefore, the sending end station also stores the excessive cells in the buffer until the sufficient bandwidth is allocated. The sending end station should detect the shortage of the bandwidth and begin the bandwidth re-negotiation as fast as possible to avoid the buffer overflow. To begin the bandwidth re-negotiation quickly, the threshold Q up should be small. However, because the Q distribution within the picture has a large variation, the small threshold causes the bandwidth controller to re-negotiate the bandwidth too frequently. When the value of Q of some picture is too small, the allocated bandwidth is too much and the bandwidth re-negotiation is required to make the allocated bandwidth smaller. From the point of view of the network efficiency, the allocated bandwidth should be reduced as soon as and as much as possible, to achieve the multiplexing gain on the link. In our method, when the value of Q within the picture goes under the pre-defined threshold (Q down ), the bandwidth controller begins the bandwidth renegotiation with the network to reduce the allocated bandwidth. Figure 1 investigates the appropriate threshold values for the bandwidth re-negotiation. The maginified figure is shown to see the details. The traffic rate variation when the fixed quantizer scale is applied is shown in Fig. 1. The pairs of thresholds (Q up Q down ) we have used are (9,7), (9,5), (,) and (11,5), which are referred to as fast up-fast down, fast up-slow down, normal up-normal down and slow up-slow down, respectively. For example, in fast up-fast down (9,7), the sending end station starts the bandwidth re-negotiation to increase the allocated

rate (Mbps) 3 5 15 5 maximum allocated minimum Figure 1: MPEG- video traffic in with bandwidth re-negotiation (Starwars) rate (Mbps) 1 1 1 maximum minimum Figure 17: MPEG- video traffic (Starwars) bandwidth when the quantizer scale within a picture goes beyond 9. In obtaining the figure, Q max is fixed at and the resulting quantizer scale should be around in our method. Thus, when we apply fast up-fast down, both shortage and surplus of the bandwidth will be detected fast. With slow up-slow down threshold pairs the allocated bandwidth is kept unchanged for some duration, because it is insensitive to the bandwidth starvation and surplus. As a result, the queue length of slow up-slow down is the largest among four threshold pairs as shown in Fig. 15. The other three threshold pairs show similar behavior in queue length. From these results, we can observe that normal up-normal down is preferable strategy in terms of low frequency of bandwidth re-negotiation and smaller delays at the buffer. Queue length of I picture (kbits) with renegotiation Figure 1: Queue length in with bandwidth renegotiation (Starwars) In Fig. 1, the processing delay of the bandwidth renegotiation within the network is fixed at one macroblock time (1/3 msec). This processing delay (displayed as D proc ) includes the propagation delay in the network and the processing delay of the signaling message. It is a rather ideal situation, and we have also investigated the case where the processing time D proc is, macroblock (5/1 sec ' msec), and we confirmed that the same tendency can be observed. The effect of the bandwidth re-negotiation failure where the sending end station retries the bandwidth re-negotiation repeatedly can be evaluated through this investigation on the large processing delay case. 3. Applicability of proposed control method As described in Subsection., we have employed the Q-R relationship function R(Q) given by eq.1 to determine the amount of requested bandwidth. While this equation is derived from a video sequence Scenery, Q-R relationship show similar characteristics in both Scenery and Starwars as illustrated in Fig. 7. We investigate the adaptability of R(Q) function to another video sequence Starwars. We assume that the bandwidth re-negotiation delay is one macroblock time in simulations. In Fig. 1, the cell generation rate and the allocated bandwidth are shown. Compared with Fig. 17, the traffic generation rate is kept almost same, which means the scene changes are treated properly. The allocated bandwidth is close to the actual traffic and this results in the non-explosion of the queue length as shown in Fig. 1. Thus, R(Q) derived from a specific video sequence Scenery may be applicable to the other video se-

Table : Allocated bandwidth comparison (Scenery, Mbps) D proc =1 D proc =, Q max Ave Max Min Ave Max Min.31 13.5 3.9.37.7 3.93..9.95.7.5. 1 3.5 9.7.9 3.75..19 Table 3: SNR comparison (Scenery, db) D proc =1 D proc =, Q max Ave Max Min Ave Max Min.7 35... 3.9.3 7. 33.9 19. 7. 33.9 19.3 1. 3.7 1..5 3.3 1. quence in our proposed method. Of course, more experiments are necessary to derive such a conclusion, which is left as a further study. 3.5 Quality comparison with Test Model5 In this section, we compare the quality of our proposed method with original MPEG- Test Model 5. In Tables and 3, we summarize the quality of our proposed method for different processing delay and required maximum quantizer scale. In these tables, the threshold pair of normal up-normal down (Q up =, Q down =) is employed. As shown in Table, the appropriate bandwidth allocation is performed according to the required video quality. As shown in the table, the resulting video quality is higher than. For example, when the bandwidth of is fixed to be 3. Mbps (which corresponds to the rate of Q max = 1), the minimum SNR of video stream is.5, which is smaller than. of our proposed method. We have also applied our proposed method to the other sequences shown in Table 1 and confirmed that our method is applicable and effective. Maximum delay guarantee Through Sections 3 and 3.5, we have shown that we can achieve effective and high quality video transfer over ATM networks with our proposed method. However, the bandwidth re-negotiation does not always succeed when the network load is high and there is no available bandwidth. Even if the bandwidth re-negotiation fails and the allocated bandwidth is insufficient, MPEG- encoder in our method produces the video stream which satisfies the required quality. As a result, the cell generation rate goes beyond the allocated bandwidth and it causes the buffering delay at the sending end station. In such a case, QoS violation in terms Queue length of I picture (kbits) with delay control without delay control Figure 19: Effect of queue length control (Starwars) of delay may happen, especially when the application operates in a real time fashion. For such applications which requires strict guarantees on data transfer delay and does not persist in the each picture quality, we introduce the mechanism to guarantee the delay bound. In this mechanism, considering the sudden growth in queue length tends to occur during bandwidth re-negotiation process, the Q limitation is removed for a while when the queue length goes beyond the threshold. When there is no Q limitation, normal encoding is performed and the queue length can be kept at the appropriate level. For example in Fig. 19, we show the effectiveness this delay control in terms of queue length at the sending end station. In this figure, when the buffering delay exceeds msec during the bandwidth re-negotiation process, the Q limitation is removed. The bandwidth re-negotiation request is repeated in the next picture s time slot. When the re-negotiation succeeds, the Q limitation is introduced again. Compared with no delay control case (labeled as without delay control in Fig. 19), it is clearly shown that the maximum queue length is kept smaller. For example, the buffering delay becomes about 9 msec from msec around 13th picture. However, the received video quality is degraded from 19.3 db to 1. db because of removal of Q limitation. It is obvious that there is trade off between picture quality guarantee and delay guarantee, and the application may decide the preferred policy. 5 Conclusion In this paper we have proposed the rate control algorithm of MPEG- video traffic, which is based on MPEG-

rate control algorithm. We have employed the bandwidth re-negotiation method to guarantee the requested QoS, and to improve the uniformity of the quality within the picture and over the entire video sequence. Through simulations, we showed that two problems of, i.e., nonuniformity of the video quality and the scene change handling, are much improved by our proposed method. We have also investigated the effect of the processing delay required to perform the bandwidth re-negotiation. Our method can also be applied to other protocols in which the bandwidth is reserved to the connection and the allocated bandwidth can be changed through the negotiation with the network. One such example is RSVP (Resource reservation Protocol)[1] which is recently developed for the Internet. As further studies, we should examine the effectiveness of our control method by applying the other MPEG- video sequences. In this paper, we have assumed that the bandwidth re-negotiation always succeeded. For more realistic evaluation where the bandwidth re-negotiation fails according to the network traffic load condition, we need to investigate the appropriate set of parameters for delay control mechanism. Acknowledgements This work was supported in part by Research for the Future Program of Japan Society for the Promotion of Science under the Project Integrated Network Architecture for Advanced Multimedia Application Systems. References [1] ITU-T, ITU-T recommendation I.371, traffic control and congestion control in B-ISDN, International Telecommunication Union, 199 revised in 1995. [] ISO/IEC DIS 131-, MPEG- video, ISO standard, 199. [3] D. Gall, MPEG: A video compression standard for multimedia applications, Communications of the ACM, vol. 3, pp. 35 313, April 1991. [5] ISO/IEC/JTC1/SC9/WG11, Test model 5, must be removed, April 199. [] ITU-T Draft Standard Q.93, Preliminary Draft. International Telecommunication Union, 1995. [7] F. Guillemin and P. Boyer, ATM block transfer capabilities: The special case of ABT/DT, Proceedings of IEEE GLOBECOM 9, pp. 7 7, November 199. [] A. Eleftheriadis and D. Anastassiou, Meeting arbitrary QoS constraints using dynamic rate shaping of coded digital video, Proceedings of 5th Intenational Workshop on Network and Operating System Support for Digital Audio and Video, pp. 95, April 1995. [9] A. Campbell, D. Hutchison, and C. Aurrecoechea, Dynamic QoS management for scalable video flows, Proceedings of 5th Intenational Workshop on Network and Operating System Support for Digital Audio and Video, pp. 7 11, April 1995. [] A.Adas, Supporting real time VBR video using dynamic reservation based on linear prediction, Proceedings of IEEE INFOCOM 9, pp. 17 13, March 199. [11] D. J.Reininger, D. Raychaudhuri, and J. Y.Hui, Bandwidth renegotiation for VBR video over ATM networks, IEEE Journal on Selected Areas in Communications, vol. 1, pp. 7 5, August 199. [] K. Fukuda, N. Wakamiya, M. Murata, and H. Miyahara, QoS mapping between user s preference and bandwidth control for video transport, Proceedings of Fifth IFIP International Workshop on Quality of Service, pp. 91 3, May 1997. [13] S. Wolfram, Mathematica. Wolfram Research, Inc., 199. [1] L. Zhang, S. Deering, D. Estrin, S. Shenker, and D. Zappala, RSVP: A new resource reservation protocol, IEEE Network Magazine, September 1993. [] M. Butto, E. Cavallero, and A. Tonietti, Effectiveness of the leaky bucket policing mechanism in ATM networks, IEEE Journal on Selected Areas in Communications, vol. 9, pp. 335 3, April 1991.