rate of a mobile network make different feedback techniques ineffective [3]. It is difficult to ... packets to serve the requests from multiple mobile clients based ..... of IEEE International Conference on Multimedia Computing and Systems, Florence ...
1
A Fair and Adaptive Scheduling Protocol for Video Stream Transmission in Mobile Environment Joe Yuen, Kam-Yiu Lam and Edward Chan Department of Computer Science City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong {csjyuen|cskylam|csedchan}@cityu.edu.hk
Abstract In this paper, we propose an Adaptive Buffer Sensitive (ABS) protocol for scheduling the transmission of video streams over a mobile network. Two important considerations in the design of ABS are: (1) to minimize the impact of transient overloading and communication errors on video playback; and (2) to minimize the impact of a video playback by other video requests. In ABS, the allocation of network bandwidth for serving client requests is divided into two phases. In the first phase, a minimum bandwidth is allocated to serve each request with an aim to minimize the impact of a video request on the performance of other requests. The allocation of bandwidth in the second phase is based on the playback buffer levels of the clients with an objective to make it more adaptive to the playback situation of individual clients. Extensive simulation experiments have been performed its performance especially under different network failure probabilities.
environment, how to provide a fair service to serve the video requests from different types of clients is an important concern. In this paper, we have designed a protocol, called Adaptive Buffer Sensitive (ABS), for scheduling the transmission of video packets to serve the requests from multiple mobile clients based on the work-in-progress work in [10]. It is assumed that video frames are packed into packet before transmission. Each packet consists of a group of pictures (GOP) compressed in MPEG standard. Two important concerns in the design of the scheduling protocol is to minimize the impact of transient overloading and communication errors on the performance of individual video playback, and at the same time to provide a fair schedule to serve the requests such that the impact of a video request on the performance of other video requests can be minimized. Transient overloading may occur due to connection errors and variation in video traffics.
2 Keywords: mobile multimedia, mobile network, real-time scheduling and buffering
1
Introduction
In last decade, the research in distributed video player systems has received a lot of interests [1, 6, 9]. Various efficient techniques, i.e., video frame buffering, feedback control mechanisms, and video stream smoothing techniques, have been proposed [1, 3]. However, due to the unique characteristics of mobile machines and mobile networks, these techniques may not be suitable to mobile video player systems. The asymmetric bandwidth property of a mobile network poses challenges on the flow-control mechanisms for video stream transmission. Limited bandwidth together with high error rate of a mobile network make different feedback techniques ineffective [3]. It is difficult to guarantee the playback quality in such a poor and unstable network environment. At the same time, mobility of clients may seriously affect the distribution of workload in the system and the effectiveness of admission control mechanisms in maintaining the quality of video playbacks. The pre -buffering techniques [7] proposed for solving the variable workload problem of a video need further enhancements before they can be effectively applied in a mobile multimedia system since some of the mobile devices, i.e., PDAs, may only have very limited buffer spaces. Even though the bandwidths of mobile networks are improving, the demand on high video quality is also growing. It is believed that even with the support of the 3G mobile networks, mobile network is still a big concern on the perfo rmance of a mobile video player system. Furthermore, different types of mobile clients may request videos with great differences in sizes, resolutions and workload characteristics. In such as a mixed workload and dynamic
Related Works
In the design of scheduling methods for multimedia systems, one of the most important concerns is to meet the urgency of each video packet since missing the playback deadline will make it useless. Therefore, many priority-based real-time scheduling algorithms, i.e., earliest deadline first and rate monotonic, have been extended for scheduling the transmission of video packets in distributed multimedia systems. Other important concerns, which are specific important for mobile multimedia systems, are fairness in services, workload distribution problem due to mobility of mobile clients , and error problems in video packet transmission. In [8], the earliest deadline first scheduling algorithm is extended for scheduling of video packets in a mobile environment. In [5], the Server Based Fairness (SBFA) approach is proposed in which it uses a long-term fairness server to save the bandwidth of bad channels. Channel-Condition independent fair queuing (CIF-Q) [4] is another fair queuing approach for error posed mobile networks. It creates an error-free system on each system as a reference and can use any well-known algorithm as the fair queue algorithm. Session selection is based on the virtual time of each system in the ideal system, i.e., an error-free system. In [2], a hierarchical admission scheme is proposed to solve the admission problem when the system is a cellular system and the mobile clients have high mobility. In this paper, we will concentrate on the design of efficient scheduling methods based on the demand of each request and the buffer levels of the clients.
3
Admission Control and Fairness in Scheduling
In order to minimize the probability of overloading the system, the use of an admission controller is important. It decides whether a video request can be accepted or has to be rejected
2 based on the existing video workload in the system. A simple way to define the admission condition is based on the average workload of each video request, i.e.:
n ∑ BW_Consume i i =1 ≤1 BW_Total
(1)
where BW_Consume i is the average bandwidth requirement of video stream i. BW_Total is the current total bandwidth available of the mobile network. With Eqn. 1, each client will receive sufficient packets for playback of its requesting video on average over a long period of time. However, transient overloading may still occur due to variable bandwidth requirements of videos. The impact of transient overloading can be resolved by buffering video frames at the clients. A client normally defines a buffer level such that initially the playback of a video will start only after the buffer level has been reached. We call this as preferred buffer level (PBL) . Different clients may have different PBL since they may have different buffer sizes. Based on the play rate of a video, we can determine the buffer playback duration (BPD), which is the period of time that the buffer will become empty if no frame arrives during that period of time. Since a video server may serve several video requests at the same time, it is important to define a fair schedule to serve them. However, the definition of fairness in such a mixed and dynamic environment is not trivial. For example if a simple deadline driven scheduling algorithm is used, i.e., earliest deadline first (EDF), the performance of the playback of a video, which workload is very light, could be seriously affected by the playback of other videos which workloads are heavy [8]. The client who is requesting a heavy workload video will use up most of the bandwidth. It is obvious that this is not fair to the light workload clients. If the allocation of bandwidth is according to frame (or packet), i.e. the server serves the requests in a roundrobin manner and every time the server transmits a video frame (packet) for each client, the playback performance of the videos with light workloads will be seriously affected by the playback of videos with heavy workloads. This policy is also unfair to light workload clients. Two other ways, which seem to be comparatively fairer, are (1) to allocate same bandwidth to serve each request using a round robin policy such that each client receives the same amount of bandwidth; and (2) to use rate monotonic such that the higher workload request receives a greater among of bandwidth. The allocation is fixed for each client. However, these methods will make the scheduling less adaptive to the playback situation of individual request. The consequence is that it will make the playback more sensitive towards the effect of transient overloading and errors in data transmission.
4
Adaptive Buffer Sensitive (ABS)
One of the objectives of the Adaptive Buffer Sensitive (ABS) protocol is to provide a fair scheduling to server the video requests from different mobile clients. To achieve this, we allocate a minimum bandwidth to each request such that the requests will have sufficient video frames for playback for a sufficient long period of time on average. After providing minimum bandwidths to all the requests, the remaining bandwidth will be allocated in an adaptive way such that the request which playback situation is poorer will receive more
bandwidth with an attempt to improve its playback status. The situation of a video playback can be defined based on the buffer playback duration (BPD) of the client buffer. If the situation is so poor that after allocating additional bandwidth to serve a request still cannot solve the problem, the server may start to drop packets at the server side by using a selective transmission scheme with an attempt to increase the buffer playback duration at the client. Although this will affect the smoothness in the playback temporarily, it is hope that after the selective transmission, the playback status and buffer level can be restored to the preferred level.
4.1 Admission Control in ABS Eqn. (1) does not consider the problem of high error rates of a mobile network. Due to the errors, the actual among of video packet transmitted could be much less than the bandwidth allocated to serve the request. In order to include the problem of re-transmission, an error factor ξ is introduced to estimate the average percentage of bandwidth allocated for retransmission of lost and corrupted packets. Therefore, the available bandwidth for packet transmission excluding those error re-transmissions, BW_Availiable, can be defined as: BW_Availiable = BW_Total × (1- ξ) (2) Therefore, Eqn. (1) can be re -defined as: n
∑ BW _ Consumei
i =1 ≤ 1 (3) BW _ Total× (1 − ξ ) A new video request j will be accepted if adding its average workload, BW_Consume j to the total workload of all the existing streams in the system can still satisfy the above inequality. Otherwise, it will be rejected and has to wait until the condition is true. In order to prevent the problem of indefinite postponement, the queuing order is first come first serve. Since the error rate of a mobile network is not a constant and is highly affected by the surrounding conditions of the clients, the value of the error factor á can be dynamically adjusted based on the error rates of the clients within a period of time using a control feedback mechanism. In a cellular network, the number of mobile clients within a cell is changing all the time and is determined by the mobility of clients. In order to prevent the interruption of a video playback while a client is moving, a video request has to be continued even if the client has entered into another cell. However, the consequence will make the total workload at the cell significantly higher than the original admission controlled value and the probability of long overloading will be higher. In order to prevent this, we can apply a predictive method together with a hierarchical scheme for admission control. We may predict the future location of a client based on its previous movement pattern. Note that the mobility of clients in a short period of time is not difficult to predict although it will change over time gradually. For example, it will affect by the connection of the roads. When we have defined the path of a mobile client, in the admission control of its future site, we will include the workload of the client in the admission of other requests into the cell.
4.2 Bandwidth Allocation After a request has been admitted, a buffer will be allocated for the request at the video server. Video frames will be retrieved from disk storage into the buffer and they will be organized into
3 packets for transmission. The allocation of bandwidth for scheduling of packet transmission is divided into two phases.
4.2.1 First Phase: Even or Minimal Bandwidth allocation In the first phase, the bandwidth allocated to serve a client request is the minimum of the bandwidth that is equally divided from the available bandwidth, and the mean bandwidth requirement of its requested video:
BW _ Ad i= min(
BW _ Available , BW _ Consumei ) n
where BW_Ad i is the bandwidth allocated to serve request i and n is the number of concurrent requests. The rationale of such allocation scheme is that we want to provide a fair treatment to all the clients independent of the workload characteristics of their requested videos. However, if a client is in high error status, i.e., high re -transmission probability, the system will not allocate bandwidth to it in this phase. The purpose is to reduce the amount of bandwidth wasted on packet re -transmission. If the average required bandwidth of a client is smaller than the mean value, (BW_Avaliable/n), we only provide its average bandwidth. Although the size of the next packet for transmission may be larger than the average workload of the video, allocating average bandwidth of the video to serve a client should not significantly affect the playback if the client buffer level is high enough. The buffer level should be able to be maintained at the preferred level in a sufficient long period of time when the client receives the average workload of its requested video. The impacts of transient overloading and communication errors on its performance will be resolved by the adaptive bandwidth allocated in the second phase.
4.2.2 Second Phase: Allocation Based on Buffer Level The total bandwidth allocated at the first phase, BW_Ad FR, is:
BW _ Ad FR =
∑ min(
i∈N
BW _ Available , BW _ Consume i ) n
where N is the set of clients which are not in error state. After first phase allocation, the remaining bandwidth, if any, will be allocated to the clients based on their buffer playback duration, which is determined as the playback time of the latest packet at the client buffer minus the current time. Obviously, the playback will be less affected by transient wo rkload if its playback duration is longer. Therefore, in the allocation of the bandwidth in the second phase, more bandwidth will be allocated to the client which playback duration is shorter:
BW_AS i =
(Buffer_Playback_Dur ation i) −1 n
−1
×
( ∑ Buffer_Playback_Dura tion i ) i =1
( 1 − BW_Ad FR)
It is expected that by allocating more bandwidth to serve a client whose video buffer is low, we can restore its buffer level to the preferred buffer level.
4.3 Selective Transmission If the buffer playback duration is lower than a certain pre defined threshold value even though more bandwidth has been allocated to it in the second phase, the server may need
selectively to send packets to the client for playback, i.e., some packets will be skipped. The selection of next packet for transmission is based on the resource consumption rate (RSR) of the packets within a selection window, which is a pre -defined period of time: RSR i = size of packeti × (play rate of the video / number of packets in the packet) In the selective transmission scheme, the set of packets within a selection window will be examined and the packet with the smallest RSR will be selected. Once a packet has been selected, the packets in front of the selected packet will be skipped. By skipping packets, it is hoped that the buffer playback duration at the client buffer can be extended. The selective transmission scheme will stop when the buffer level is restored to the preferred level. The size of fixed number of packets is a tuning parameter, and its value will be adjusted to a larger value if the playback situation of the client continues to go worse.
5
Performance Evaluation and Results
In order to investigate the performance characteristics of ABS, we have developed a simulation program using a simulation language called CSIM , to simulate a mobile video system, which consists of a video servers and a number of mobile clients. The clients are divided into three groups and each group of clients has different buffer sizes and will request videos with different workload characteristics. The followings show the key model parameters and their baseline setting values. Mean Stream Length 10,000 GOPs Network bandwidth 1Mbps Number of Client Groups 3 Mean bandwidth of video 0.2Mbps requested by Class 1 clients (BW_Class1) Mean bandwidth of the video to BW_Class1 × 0.5 be requested by Class 2 clients Mean bandwidth of the video to BW_Class1 × 0.01 be requested by Class 3 clients Number of Clients per group 50 Think Time of the clients between 100 ~ 300 sec each request uniformly distributed Error Possibility 0.5 Error Duration 100 sec Preferred buffer level (PBL) 5 sec Drop packet buffer level 1.5 sec Simulation length 500,000 sec Figure 1. Model parameters and baseline values We have modeled a system with three classes of clients and each class of clients has different workload requirements as defined in Figure 1. When the buffer level at a client is lower than the drop packet buffer level, the selective transmission scheme will be invoked. We have included an error model to model the mobile communication problems between the clients and the server in the simulation model. It consists of two components, error probability and error duration. The error probability is used to model the probability of a client in communication error state. It may be a result of the interferences from the surrounding buildings or a result of being too far away from the server. The error duration is the mean duration of a client in error state in each time. In addition of modeling ABS, we also have developed two other simulation programs to simulate the same system using earliest deadline first (EDF) and
4 rate monotonic (RM) for comparison purposes. Both EDF and RM are optimal real-time scheduling algorithms. EDF schedules the transmission of packets based on their playback deadlines. A higher priority is assigned to the packet with the closest deadline. In RM, the allocation of bandwidth to server each request is proportional to the bandwidth requirements of the each request. As shown in Figure 2, the performance of ABS is consistent much better than EDF and RM when we vary the error duration. As expected, the frame lost rate increases with error duration. The better performance of ABS is due to a higher buffer level maintained at the mobile clients such. As shown in Figure 3, the average buffer level of the clients in ABS is consistent much higher than that in EDF and RM. If we look at the performance of individual class, we find that although the performance of Class 1 clients (heavy workload clients) under ABS is worse than EDF and RM, the performance of Class 2 and Class 3 clients under ABS is much better than under EDF and RM. The reason is that under ABS, the impacts from higher workload requests on lighter workload requests are much smaller. Figure 4 and 5 shows the performance results when we vary the error probability. Consistent with the results shown in Figure 2 and Figure 3, ABS gives a much better performance and the buffer level of the clients is higher. We also have investigated the impacts of number of clients on their performance. The results are consistent with those shown in Figure 2 to Figure 5.
6
Conclusions
In this paper, we propose an Adaptive Buffer Sensitive (ABS) scheme for scheduling of frames from multiple video streams to use the limited mobile bandwidth. Our objective is to minimize the impact of transient overloading and errors in communications, which is very common in a mobile network, on the playback quality of a video and at the same time to minimize the interferences between the playbacks of different videos. By using a two phases bandwidth allocation scheme, we find that the buffer levels at the mobile clients can be maintained at a higher level and the performance of the whole system can be significantly improved. Frame Lost Rate
0.8 0.6 0.4
RM ABS
0.2
EDF
0 0
20
40 60 80 Error Duration (sec.)
100
Figure 2. Frame Lost Rate Vs. Error Duration Buffer Level (Frames)
80
RM EDF ABS
60 40 20 0 0
20
40 60 80 Error Duration (sec.)
Figure 3. Buffer Level Vs. Error Duration
100
Frame Lost Rate
0.8 0.6 0.4
RM
0.2
EDF
ABS
0 0
0.2
0.4 0.6 0.8 Error Probability
1
Figure 4. Buffer Level Vs. Error P robability Buffer Level
40 30
RM ABS
20
EDF
10 0 0
0.2
0.4 0.6 Error Probability
0.8
1
Figure 5. Buffer Level Vs. Error Probability
References [1] R. Agarwal and A. M. K. Cheng, “Reducing Variation in BitRate Produced by Encoder in MPEG Video”, in Proceedings of IEEE International Conference on Multimedia Computing and Systems, Florence, Italy, June 1999. [2] Sunghyun Choi and Kang G. Shin, “A comparative study of bandwidth reservation and admission control schemes in QoS-sensitive cellular networks”, ACM Wireless Networks, vol. 6, no. 4, pp. 289-306, 2000. [3] Shanwei Cen, Calton Pu and Richard Staehli, “A Distributed Real-time MPEG Video Audio Player”, in Proceedings of the 5th International Workshop on Network and Operating System Support of Digital Audio and Video, 1995. [4] T. S. Ng, I. Stoica, H. Zhang, “Packet Fair Queuing Algorithms for Wireless Networks with Location-Dependent Errors”, INFOCOM ’98, vol. 3, 1998. Pg 1103-1111. [5] P. Rammanathan, P. Agrawal, “Adapting Packet Fair Queuing Algorithms to Wireless Networks”, 4th Annual ACM/IEEE international conference on Mobile Computing and Networking, Oct 1998. [6] S. Rao and A. M. K. Cheng, “Scheduling and Routing of Real-Time Multimedia Traffic in Packet-Switched Networks”, in Proceedings of IEEE International Conference on Multimedia, July -Aug. 2000. [7] Cormac J. Screenan, Jyh-Cheng Chen, Prathima Agrawal, B. Narendran, “Delay Reduction Techniques for Playout Buffering”, IEEE Multimedia, vol 2, no. 2, pp. 88-100, 2000. [8] S. Shakkottai, R. Srikant, “Scheduling Real-time Traffic with Deadlines over a Wireless Channel”, in Proceedings 2nd ACM Wireless Mobile Multimedia, August 1999. [9] David K. Y. Yau, and Simon S. Lam, “Adaptive RateControlled Scheduling for Multimedia Application”, IEEE/ACM Transactions on Networking, vol. 5, no. 4, 1997. [10] Joe Yuen, Kam-Yiu Lam, and Edward Chan, “Adaptive Playback Scheduling for Transmitting Videos Streams in a Mobile Multimedia Systems”, in Proceedings of 7th IEEE Real-Time Technology and Applications Symposium (work in-progress paper), pp.125-126, May 2001, Taipei.
