Self-Adjusted Network Transmission for Multimedia Data

Transcription

1 Self-Adusted etwor ransmission for Multimedia Data Mei-Ling Shyu Department of Electrical and Computer Engineering University of Miami Coral Gables, FL 334 USA Shu-Ching Chen School of Computer Science Florida International University Miami, FL 3399, USA Hongli Luo Department of Electrical and Computer Engineering University of Miami Coral Gables, FL 334 USA Abstract High bandwidth requirements in multimedia transmission mae the efficient use of limited networ resource a challenging tas, especially when multiple clients mae their requests to the server simultaneously. In this paper, we propose a self-adusted networ transmission mechanism for multiple clients. Instead of assigning a fixed bandwidth for each client, the server determines the transmission rate for each client connection according to the buffer pacets and playbac rate at each client. ransmission rates are adusted when the total requested bandwidth is larger than the networ bandwidth. In addition, the proposed mechanism can minimize the bandwidth allocation and maximize the client buffer utilization. A simulation is performed and the simulation results show that the proposed mechanism can dynamically change the transmission rate for each client to avoid overflow of the client buffer, and achieve the optimal utilization of the limited networ resource in multiple client networ environments.. Introduction he development and use of distributed multimedia applications are growing rapidly in these years. ypical examples are video-conferencing, video-on-demand, digital library and distant learning. High networ bandwidth is usually needed in order to provide high quality delivery of information. Efficient utilization of networ resources is essential in the provision of costeffective multimedia services where quality-of-service (QoS) [4] requirements are met. Parameters for QoS are reliability, bandwidth, itter, and etc. Multimedia information in general has a highly timevarying bandwidth requirement since media data are variable bit rate (VBR) in nature due to the coding and compression technologies applied [4][7]. For example, in an MPEG- movie, the average frame size is usually less than 5% of its maximal frame size [9]. here are several approaches to address the bandwidth requirements of multimedia transmission. One approach is the static resource reservation schemes that allocate a constant bit rate (CBR) channel to transmit the VBR stream by the pea data rate. With large variations in bandwidth requirements of multimedia data, static allocation usually results in considerable wastage of networ resources. Another approach is rate adaptation [5] that adusts the bandwidth used by a transmission connection according to the existing networ conditions. he adaptive approach can better utilize the available networ resource that changes with time in comparison with the static allocation approach. Media synchronization is also needed to guarantee itter-free playbac requirement [][3]. Specifically, the delivery of videos imposes two challenges for multimedia transmission, namely a high-bandwidth requirement and a real time delivery constraint. In order to provide itter-free video playbac, each video frame must arrive at the client buffer before the time it is scheduled to be displayed. he data arrives at the client and is stored temporarily in the client buffer. If the buffer is full when a pacet arrives, this pacet will get lost and overflow occurs. If the pacet cannot arrive at the client before the schedule time, underflow occurs and itters in playbac happen. he client buffer acts as a reservoir to regulate the difference between the transmission rate and the playbac rate. It is an important resource for users to prevent playbac itters. Since most of the clients have limited buffer space, the networ must control the transmission rate to avoid overflow or underflow to guarantee the end-to-end QoS. Much wor has been done in the resource allocation, traffic control, and multimedia synchronization [5][6][0] [6] in transmission in multimedia networ. he provision of an efficient and reliable transmission rate scheduler that can optimize the utilization of buffer and bandwidth at the

2 same time is still a challenge. In our previous wor [], we proposed a closed-loop framewor for multimedia transmission and investigated the single client situation. Linear Quadratic (LQ) tracer was used to obtain the optimal transmission rate for one client connection. It was proved that the proposed closed-loop framewor efficiently utilizes the networ resources such as the bandwidth and the client buffer. In this paper, a selfadusted transmission mechanism that dynamically changes the transmission rates for multiple clients is designed. A networ model is introduced to capture the relationships among the transmission rates, buffer pacets, and playbac rates. Simulation results show that the proposed mechanism can minimize the bandwidth allocation and maximize the pacet size stored in client buffer at every time instant. he organization of this paper is as follows. he proposed networ framewor and the self-adusted networ transmission mechanism for multiple clients are presented in Section. Simulation results are given in Section 3. Section 4 concludes this paper.. Self-adust networ transmission In this section, the self-adusted networ transmission mechanism for multiple clients and the networ model are presented... etwor model for a single client he networ model for a single client (e.g., the th client) is given in Figure. Define the variables as below: : time instant; Q(): pacet size in buffer at time instant ; R(): pacets transmitted from the server at time instant ; R (): pacets arriving at the client buffer at time instant ; L(): pacets used for playbac at time instant ; and Q r : allocated client buffer size at the setup of the connection. In this framewor, we consider the situations that the propagation delay is quite small and can be negligible. So it can be assumed that the pacets transmitted from the server at a time instant are equal to the pacets received at the client. herefore, we have R ( ) R( ) () and R() is used to represent R () in this paper. Consider the relationship among Q(), R(), and L(). he buffer pacet at the time instant is Q( ) Q( ) R( ) L( ) () Let Q() be the variable we want to trac, and e() be the difference between the traced Q() and the allocated Q r. e( ) Q( ) (3) Q r In order to mae fully use of the client buffer, e() should be minimized. On the other hand, the transmission rate R() should also be minimized to save the bandwidth. Figure. Proposed networ model for a single client (e.g., the th client). he Multimedia Augmented ransition etwor (MA) model that was proposed in our previous research [][][3] can provide the playbac schedule for the multimedia video before the transmission starts. At the time of creation of a multimedia presentation, the MA model can be used to specify the temporal constraints among various multimedia data that must be observed at the time of playbac. he multimedia presentation is modeled by MA and stored at a server. hat is, the value for L() is nown. When the client requests the information from the server, the playbac schedule is obtained by the server and can be used to determine an optimal transmission rate to satisfy the playbac requirement. Instead of transmitting the pacets in a fixed rate, transmission rate can be changed automatically according to the buffer pacets and playbac rates in the clients. In order to maximize the buffer utilization and minimize the bandwidth allocation, Linear Quadratic (LQ) tracer [8] is used to design the transmission scheduler to get the optimal transmission rate for each client connection... Discrete linear quadratic tracer First, an optimal control law [8] that forces an obect to trac a desired reference traectory over a specified time interval is presented. Given an obect that can be represented in a linear status space form, and a nown disturbance d, we have

3 x Ax Bu d (4) where A and B are constant matrices describing the obect, u is the control input, and x is the status of the obect to be traced. Let the cost function J of the system be ( x Gx u Hu ) J x S x (5) with S 0, G > 0, H > 0, where G and H are the weighting matrices, S is for the boundary conditions, and [, ] is the period we want to trac the obect. he optimal control u is given by ( B S B H ) B S A, S given, A S ( A B ) G, ( A B ) v ( A B ) S d S v v F (6) (7) 0, ( B S B H ) B x F v, (8), (9) u (0) where is the feedbac gain, F is the feedforward gain, and S and v are auxiliary sequences when calculating the optimal u. Applying the control input sequence u to the obect, we can get a sequence of obect status that minimizes the quadratic cost function J..3. Optimal rate transmission mechanism for multiple clients Assume that there are m clients that request data from the server simultaneously. For the th client, let J be the cost function, Q r, be the allocated buffer size of the th client at the connection, e () be the difference between the allocated buffer size and the buffer pacet of the th client at the time instant, Q () be the buffer pacet at the time instant, and R () be the transmission rate of the th client. Define the difference between the allocated buffer size and the buffer pacets as e ( ) Q ( ) Qr, () o maximize the client buffer Q () is to minimize the e (). he optimization function of the server side should be ( e ( ) R ( )) m m J J () Since the m clients are independent, if the J of the th client is minimal, then the sum of the J functions is also the minimal. ( e ( ) R ( ) ) J (3) For each client, the LQ tracer is used to achieve the optimal transmission rate. After the server calculates the optimal bandwidth for each client, an effective bandwidth W for the connection for the th client should satisfy the requirement that W BW (4) where BW is the networ bandwidth. Since the transmission rate is optimized individually, it is possible that the sum of all the requested bandwidth is greater than the networ bandwidth. herefore, we need to reallocate the bandwidth to each client connection. o provide fairness among all clients, the bandwidth reallocated should be proportional to the real requirement of each client. An algorithm for the multimedia server is given below in the form of the pseudo code. Let rate( ) be the transmission rate for the th client and otal_rate be the sum of the transmission rates for all the m clients. Pseudo code for rate adustment at time instant on the sever is given as follows. m otal_rate rate( ); if otal_rate > BW for to m do rate()rate()*(bw/otal_rate); endfor endif Sometimes when the playbac rate is low and the transmission rate is large, overflow may occur. Since the playbac schedule is nown, we can predict the situation and avoid it. he buffer size can be increased automatically at the possible overflow instant to accommodate those pacets and be decreased to the normal value when overflow period ends.

4 For each client, we have the following buffer difference equation (5) of the client buffer that can be used to obtain the optimal transmission rate. Since for each client, we follow the same computation procedure, in order to simplify the notation, e(), R(), and L() are used to represent the corresponding values of the th client at time instant. of the clients and the total transmission rate of the server are depicted in Figure. In Figure, the transmission rates,, and 3 indicate the optimal transmission rates for client connections,, and 3 after the rate adustment. As can be seen from the figure, at some time instant, the optimal transmission rate should be adusted when the e( ) e( ) R( ) L( ) (5) According to the corresponding equations (4) and (5), we set the values A, B, S 0, G and H. hen, the optimal transmission rate R() can be obtained by solving the following equations. ( S ) S, S 0, (6) S S ( ), (7) ( ) v ( ) S L, v 0, ( S ) v (8) F (9) herefore the optimal transmission rate at time instant is R (0) ( ) e ( ) F v and the pacets in the buffer at time instant is Q ( ) e( ) () In the proposed algorithm, the transmission rate for the whole transmission period can be calculated before the transmission starts. Lots of the parameter calculation can be done off-line, which simplifies the implementation of the transmission rate controller. hat is, the sequences S, and F can be computed off-line before the transmission rate applies. he gains and F can be stored for use when the actual transmission proceeds. he only wor left to do is to compute the optimal control R() with equation (0). 3. Simulation results In this paper, a simulation is performed to illustrate how the proposed self-adusted networ transmission for multiple clients can be achieved. For simplicity, assume that there are 3 clients simultaneously connecting to the host and requesting the services. he simulation is run within the interval [, 50] seconds with the time increment of second and uses the randomly generated playbac rates to simulate the real playbac scenario. Assume that the playbac rate is between 0.MB per second (MBps) and MBps, the allocated client buffer size is MB, and the networ bandwidth is MBps. he transmission rates Q r Figure. ransmission rate changes for each client (clients,, and 3) and the server (networ bandwidth MBps, t to 50 second, and increment second). total transmission rate of all the clients is larger than the networ bandwidth (for example, at time instants 34 th second and 37 th second). After the rate adustment, the total transmission rate will be equal to the networ bandwidth, which means the limited bandwidth is utilized maximally to satisfy the playbac requirements of all the clients. If the transmission rate is not large enough to satisfy the playbac rate L() after the adustment, the pacets in the client buffer will be consumed so that the difference between the allocated buffer size (Q r ) and the pacet size in the buffer (Q()) increases. he later optimized transmission rates will be automatically obtained to alleviate this situation and satisfy the playbac requirement. Each client has its own playbac schedule and has its pea rate at different time instants. If we allocate the bandwidth according to the pea rate, it is a waste of bandwidth since most of the time the playbac requirement is much below the pea rate. In addition, there will be fewer clients that can be serviced by the server. In our approach, the proposed self-adusted networ transmission mechanism dynamically allocates

5 the bandwidth to each client according to its playbac requirement and buffer pacets. At a certain time instant, more bandwidth can be assigned to those clients whose playbac requirements are high and less bandwidth can be assigned to those clients with low playbac requirement, under the constraint that the sum of bandwidth does not exceed the networ bandwidth. Under this mechanism, the playbac requirement of each client can be satisfied and the underflow situation can be avoided. At the same time, service for more clients can be provided with the same limited networ bandwidth. Compared with the traditional fixed bandwidth allocation approaches, our approach is more efficient in utilizing the networ bandwidth. transmission rate increases when the buffer pacets cannot provide enough pacets for the playbac requirement. It needs to be noted that at the current stage, the proposed mechanism does not put any constraint on the final time instant in our optimization function. herefore, it is possible that overflow occurs at the final time instant. From the simulation results, it is shown that our proposed self-adusted networ transmission mechanism can avoid the loss of pacets, and at the same time achieve the minimal bandwidth allocation and maximal utilization of the client buffer. he advantage of the proposed mechanism is that the transmission rate is dynamically adusted according to the playbac rates of all the clients. he transmission rate is increased automatically when the playbac rate is high. Consequently, the client buffer can be fully utilized under the limited bandwidth to provide itter free playbac. According to the performance index function we try to minimize, when the minimal index function for each client is achieved, the minimal index function at the server side is obtained, so that the bandwidth utilization is optimized. herefore, more connections can be admitted under the same networ bandwidth. 4. Conclusions Figure 3. ransmission rate changes with the pacet sizes in the client buffer and the playbac rates for client (Q r MB, t to 50 second, and increment second). o illustrate how the rate changes according to the playbac rate and buffer pacet for every client connection, the changes of the transmission rates, pacet sizes in the client buffer, and the playbac rates for one single client (i.e., client ) are shown in Figure 3. Usually, the transmission rate is determined according to the playbac rate requirement and the existing buffer pacet. he rate is determined so as to satisfy the playbac requirement while at the same time to eep the rate as minimum as possible. For example in Figure 3, at the time instant 6 th second, since the client buffer is full and the existing buffer pacets can provide for the playbac requirement, the transmission rate can be adusted at a low level to save the bandwidth. Also, the playbac rate begins to increase and eeps at a high level starting from the time instant 6 th second, and the When there are multiple clients requesting data from a multimedia server simultaneously, how to efficiently allocate the bandwidth to each client and satisfy the QoS requirements of the client applications is a challenging tas. In this paper, we propose a self-adusted networ transmission mechanism that can obtain the maximum utilization of the networ resource. Instead of giving a fixed transmission rate for each client, the transmission rate for each client is determined dynamically to achieve the maximal utilization of client buffer and the minimal allocation of bandwidth. he optimized transmission rate can be adusted to satisfy the constraint of the networ bandwidth. With the limited bandwidth resource of the networ, QoS requirements for multiple clients can be met using the proposed mechanism. In addition, since lots of the parameter computation can be done off-line, the proposed mechanism is also practical and efficient. 5. References [] S.-C. Chen, M.-L. Shyu, and R.L. ashyap, Augmented ransition etwor as a Semantic Model for Video Data, International Journal of etworing and Information Systems, Special Issue on Video Data, Vol. 3, o., pp. 9-5, 000. [] S.-C. Chen, M.-L. Shyu, C. Zhang, and R.L. ashyap, Obect racing and Augmented ransition etwor for Video Indexing and Modeling, the th IEEE International Conference on ools with Artificial

6 Intelligence (ICAI 000), Vancouver, British Columbia, Canada, pp , ovember 3-5, 000, [3] S.-C. Chen, S. Sista, M.-L. Shyu, and R.L. ashyap, Augmented ransition etwors as Video Browsing Models for Multimedia Databases and Multimedia Information Systems, the th IEEE International Conference on ools with Artificial Intelligence (ICAI 99), Chicago, IL, USA, pp. 75-8, ovember 9-, 999. [4] W. Effelsberg, O. Spaniol, A. Danthine, and D. Ferrari, editors. High-speed etworing for Multimedia Applications. luwer Academic Publishers, 996. [5] J. Y. Hui, Resource Allocation for Broadband etwors, IEEE Journal on Selected Areas in Communication, pp , Vol. 6, o. 9, 988. [6] R. Jain, Congestion Control and raffic Management in AM etwors: Recent Advances and a Survey, Computer etwors & ISD Sys., Vol. 8, o. 3, pp , 996. [7] H. anaia, P. Mishra, and A. Reibman, Adaptive Congestion Control Scheme for Real-time Pacet Video ransport, SIGCOMM Symposium on Communications Architecture and Protocols, San Francisco, California, pp. 0-3, ACM/IEEE, Sept [8] F. Lewis, and L. Syrmos, Optimal Control. John Wiley & Sons, IC., 995. [9] M. Garrett, and W. Willinger, Analysis, Modeling and Generation of Self-similar VBR Video raffic, ACM SIGCOMM, pp , August 994. [0] S. Ramanathan, and P. V. Rangan, Adaptive Feedbac echniques for Synchronized Multimedia Retrieval over Integrated etwors, IEEE/ACM ransactions on etworing, Vol., o., pp , April 993. [] J. Roberts, U. Mocci, and J. Virtamo, editors. Broadband etwor eletraffic, Final Report of Action Cost 4, Springer, 996. [] M.-L. Shyu, S.-C. Chen, and H. Luo, Optimal Resource Utilization in Multimedia ransmission, IEEE International Conference on Multimedia and Expo (ICME), Waseda University, oyo, Japan, pp , August - 5, 00. [3] R. Steinmetz, Human Perception of Jitter and Media Synchronization, IEEE Journal on Selected Areas in Communications, Vol. 4, pp. 6-7, January 996. [4] M. atipamula, and B. hasnabish, editors. Multimedia Communications etwors echnologies and Services. Artech House, pp , 998. [5] X. Wang, and H. Schulzrinne, Comparison of Adaptive Internet Multimedia Applications, IEICE rans. Commun., Vol. E8-B, o. 6, pp , June 999. [6] M. Woo,. U. Qazi, and A. Ghafoor, A Synchronization Framewor for Communication of Pre-Orchestrated Multimedia Information, IEEE etwor, Vol. 8, o., pp. 5-6, January/February, 994.