INTERNET OVER DIGITAL VIDEO BROADCAST: PERFORMANCE ISSUES Hakan Yılmaz TÜBİTAK Marmara Research Center Information Technologies Research Institute Kocaeli, Turkey hy@btae.mam.gov.tr Bülent Sankur Boğaziçi University Department of Electrical and Electronics Engr. Istanbul, Turkey sankur@boun.edu.tr Abstract Digital Video Broadcasting (DVB) defines the carriage of multimedia information to clients by means of MPEG-2 Transport Streams (TS). Within MPEG-2 transport stream, it is also possible to carry defined data containers that can be used to realize new data services or to carry IP datagrams. In this paper, we analyze this for carrying best-effort IP traffic within the prioritized video traffic. We investigate the statistical properties of the traffic under different network resource configurations, i.e., bottleneck bandwidth, buffer size, and different traffic characteristics, i.e., selfsimilarity ( parameter). Keywords: Self-similar traffic, parameter, IP over DVB, throughput, delay, drop rate 1. INTRODUCTION In this work we investigate the potential of a Digital Video Broadcast (DVB) [1] system, which is normally used for television programs, as an alternative Internet access scheme. In this scenario the consumer uses his set-top box and TV screen interactively both to decode and watch TV programs and/or to be connected to the Internet. The DVB system is based on the cell-oriented packet transmission system, where the MPEG-2 Systems Standard provides the means of multiplexing several types of multimedia information into one Transport Stream (TS). The variable-bit-rate (VBR) video compression causes a bursty traffic [2,3], which requires allocation of considerably larger bandwidth to satisfy delay and bit-error-rate (BER) requirements of DVB. MPEG-2 TS also allows for data containers in addition to the audio and video data containers, on which new data services can be implemented or simply IP datagrams can be carried. An important feature of video traffic at the packet level, which has significant impact on performance, is its temporal correlation or its self-similar characteristics [2,3]. Specifically, bursts of data generated by VBR compressed video is the main cause of long-range dependence in time and hence of the self-similar traffic [4]. Such self-similar traffic is encountered in both LAN and WAN environments, for http traffic [5] and can persist across several protocol layers. The studies on the impact of self-similarity on network performance [6] have shown that network performance degrades gradually with increased heavy-tailedness while queuing delay and response time deteriorate more drastically. To mitigate performance losses one can adjust the network resources, such as increasing link bandwidth and buffer capacity, which tend to improve the performance in a superlinear (i.e., logarithmic) fashion. This is an indication that optimal resource allocation plays an important role in achieving the desired Quality of Service (QoS).
Section 2 of this paper describes the system architecture of the Internet-over-DVB scheme and the simulation setup. In Section 3 the simulation results are illustrated and their interpretation toward the engineering of the Internet access link is given. 2. SYSTEM ARCHITECTURE and SIMULATION SETUP Multimedia and web-based applications are characterized by highly asymmetrical traffic patterns, where considerably more data is received at the end-user station than what originates there. For these types of services, a configuration where the satellite provides a broadband delivery network with a complementary terrestrial Internet interaction back-channel offers a good alternative. In the Internet over DVB, the forward link is a broadband simplex channel with receive-only characteristics as seen by the client station, and it serves as the delivery channel. While the return network provides point-to-point links, the forward channel has broadcast characteristics [7]. S1 S2 G1 C1 V1 S64 Figure 1a. The simulated network configuration Figure 2b. The Internet over DVB scenario We simulate a network architecture using the UCB/LBNL network simulator NS [8]. Our network configuration that consists of 67 nodes and 66 links, is depicted in Fig. 1a. Each output link has a buffer, a link bandwidth, and an amount of latency associated with it. In this figure, S1 - S64 represent the server nodes C1 denotes the client node, V1 represents the VBR video source node, and finally G1 indicates the satellite gateway. The link between the satellite gateway and the client is formed by a unidirectional satellite downlink and a terrestrial interaction uplink. Although only C1 seems to be connected to the satellite link, in fact the traffic on this link is the aggregated traffic from servers to various clients, of which only one is shown. The properties of this link are set as follows: a) Non-bottleneck links, which are the uplinks between the servers and the satellite gateway, are set at 1 Mbps (duplex) and the latency of each link is set to 15 ms, to account for transmission, propagation and queuing delay. b) The VBR video server is connected to the satellite gateway via a simplex uplink of 2 Mbps bandwidth and 1 ms propagation delay. c) Uplink from the client to the gateway/server site is provided by means of a simplex link with 28.8 kbps bandwidth and 3 ms propagation delay. d) The downlink between the satellite gateway and the
client, which figures as the bottleneck link in our configuration, has 28 ms propagation delay. The bandwidth of the satellite link is parametrically varied in different simulations. Internet Traffic: Differently seeded/independent Pareto on/off sources have been used to simulate Internet server nodes. In these nodes inter-request times (idle time/off time) from the servers are drawn from a Pareto distribution while each server node possesses a Pareto-like file size (burst time/on time) distribution. Each file, however, is split into segments before transmission where, according to the DVB specifications, the maximum segment size is taken as 2 bytes. The shape parameter α of the Pareto on/off sources is parametrically varied in order to obtain traffic at different levels of self-similarity in simulations. The location parameter k of the Pareto on/off sources is adjusted meanwhile to guarantee the same mean rate at different α (α : 1.95, 1.65, 1.35, 1.5) shape values. The mean rate of the Pareto on/off sources has been calculated as 426 kbps for the aggregated traffic at the input of the satellite gateway. Meanwhile the self-similarity index H for the mentioned shape values is estimated using the Whittle estimator as: H:,,,, respectively. Video Traffic: Both actual video trace files and synthesized traces have been used to simulate VBR video traffic. The trace file of MPEG-1 coded Star Wars movie, coded by Mark Garrett (Bellcore) [3], has been used in most of our simulations. Some simulations have been repeated with trace file of MPEG-4 coded Jurassic Park movie for consistency check. Both trace files have been tested for self-similarity using the R/S plot method and Whittle estimator method. The parameter for Star Wars trace and Jurassic Park trace were, respectively, estimated as.9989 and.75. Finally the mean bit rate of Star Wars and Jurassic Park traces were calculated as 365 kbps and 783 kbps. Compound Traffic: Packets from VBR video source and Pareto on/off (Internet) sources are aggregated on the gateway node to constitute the compound traffic. Class Based Queuing (CBQ) is implemented on the gateway so that video traffic is prioritized. Consequently Internet packets are carried on the available bit rate (ABR) or unspecified bit rate (UBR) left over from the VBR video traffic but without affecting the video traffic. The buffer size for the queue is varied on different simulations to analyze its impact. Methodology: Simulations are run for 12 seconds. The first seconds interval is ignored to capture only the steady-state performance of the system. Simulations with the Star Wars video trace have shown that this trace requires either very large buffers or high link bandwidth to satisfy the loss rate requirements for MPEG video. We tried to establish the minimum effective bandwidth over which this video could be carried within the delay variation and bit error rate specifications. For broadcast quality real-time video these bounds are specified as 1 ms delay variation and 1-5 bit error rate [9]. A simple calculation proves that these bounds are met when a buffer size of 1 packets and a link with minimum ~8 kbps bandwidth is used. This configuration is used as a baseline for other simulations presented in this paper. Two transmission protocols, respectively connection-oriented and connectionless, are used. To see the effect of self-similarity at source level, an open-loop greedy connectionless transport protocol (User Datagram Protocol: UDP) is used. On the other hand, as a case of connection-oriented transport protocol, TCP Reno, as being the most
widely used transmission protocol on the Internet today, is used as our common benchmark protocol. 3. ENGINEERING OF THE INTERNET ACCESS The two resources considered are the channel bandwidth (the downlink from satellite gateway to the client) and the buffer capacity on the gateway. The parameters are the traffic shape as determined by its self-similarity measure, i.e., the parameter. Finally the performance parameters we investigate are the packet loss probability and the queuing delay. Effect of Self-Similarity: To capture the effect of Internet traffic self-similarity on network performance at source level, UDP-based transmission driven by a greedy transport agent is used and the shape parameters of Pareto on/off sources are varied. When Internet traffic is transferred on UDP via a bottleneck link, without video traffic, with link bandwidth equal to the mean-rate of Internet traffic sources, i.e., 427 kbps (lower group of curves), Internet packet drop rate increases with parameter. Then the compound traffic (prioritized video traffic + Internet traffic) is transferred, when the bottleneck link bandwidth is set to 792 kbps (365 kbps + 427 kbps), which is the sum of the mean rates of video and Internet traffic. In this case, the drop rate curves suggest that the best-effort traffic is desensitized to the self-similarity measure of the Internet traffic when the latter is carried in the presence of prioritized self-similar video traffic. Similarly we use Reno TCP to evaluate the effect of self-similarity when the transport layer implements reliable communication with congestion control. Fig. 2a depicts the reliable throughput of TCP without video traffic on a bottleneck link channel of 427 kbps. These curves indicate that the effect of short-range statistics become manifest in larger buffer configurations where a gradual drop can be seen with increasing values of the H parameter. On the other hand a smaller buffer is easily filled up even at low self-similarity levels. Fig. 2b depicts TCP traffic average queuing delay for packets that have not been dropped. This figure suggests that self-similarity at application/source level is directly transferred to the lower layers, which is then seen in the transport and network layers. Effect of Bottleneck Link Buffer Capacity: Fig. 3a presents the reliable throughput of TCP Internet traffic versus bottleneck link bandwidth for different self-similarity index values (H:,,, ). Packet queuing delay for the network configuration is presented in Fig. 3b. The TCP throughput achieves the mean Internet source rate above a critical buffer size for any degree of self-similarity. This nice performance plateau comes, however, only at the cost of excessive queuing delay. As seen in Fig. 3b, for small H values, average queuing delay, as representative of mean queue length, follows a sublinear, roughly logarithmic dependence on buffer capacity. Whereas for large H values close to one, the dependence of queuing delay becomes linear, that is the highly fractal traffic finds always a means to fill up any size buffer. Effect of Bottleneck Link Bandwidth: We evaluate the effect of bottleneck link bandwidth on network performance by conducting simulations over different bottleneck link bandwidth values. Fig. 4a and 4b illustrate the best-effort traffic reliable throughput and the concomitant delay of Internet packets in the presence of Star Wars video traffic. Performance differences between different H valued Internet traffic classes is minimal. Two important remarks at this moment are as follows: First one can observe that the bandwidth resource has a much more critical impact on the viability of the Internet access link over DVB as compared to the buffer resource. Second, the
bandwidth required to guarantee satisfactory Internet access performance on a Class Based Queuing system still remains somewhat above the mean traffic rate, with provisioning of an adequate buffer capacity. For instance, the Star Wars sequence plus the Internet traffic add up to a bit rate of 792 Kbps, while the bandwidth for a low delay high throughput scheme necessitates about 1 Mbps. 4. CONCLUSION We have shown by means of simulation analyses that the MPEG pipe in the VBR context provides a viable Internet access scheme provided the system is judiciously dimensioned. Compressed video, due to its self-similar nature, requires more bandwidth than its mean rate. This extra bandwidth is not in service, however, all the time and it is fully utilized to accommodate bursts of traffic. The potentially wasted excess bandwidth can be used to implement best-effort IP services. We have remarked that, even when the parameter is.99, prioritized video requires about 12% more bandwidth than its mean traffic rate. The Internet traffic, transported via TCP, can be nicely fitted into this excess bandwidth. Thus the two traffic streams, that is, the prioritized video and the background Internet, can be transmitted over the same link within the bandwidth that solely video traffic would have demanded. Notice that Star Wars, represents very severe network conditions. In reality most video programs, such as news, exhibit much lower burstiness, implying that the proposed Internet-over-DVB scheme should perform even better. In the final analysis of a network design, we can conclude that when the source traffic is susceptible of long-range dependence, increasing buffer capacity alone is ineffective since it carries a significant queuing delay penalty while improving on packet losses only gradually. A bandwidthdriven network resource allocation policy is more effective since it improves both throughput and queuing delay. 5. References 1. EBU/ETSI, "Digital Video Broadcasting (DVB); Implementation guidelines for the use of MPEG-2 Systems, Video and Audio in satellite, cable and terrestrial broadcasting applications", ETR 154, September 1997. 2. Beran, J., R. Sherman, M. S. Taqqu and W. Willinger. Long-Range Dependence in Variable-Bit-Rate Video Traffic, IEEE Transactions on Communications, Vol.43, pp.1566-1579, 1995. 3. Garrett, M. W. and W. Willinger, Analysis, modeling and generation of self-similar VBR video traffic, Proceedings of the ACM Sigcomm '94, pp. 269-28, 1994. 4. Leland, W. E., M. S. Taqqu, W. Willinger and D. V. Wilson, On the self-similar nature of Ethernet traffic (Extended version), IEEE/ACM Transactions on Networking, Vol.2, pp.1-15, 1994. 5. Crovella, M. E. and A. Bestavros, Self-Similarity in World Wide Web Traffic: Evidence and Possible Causes, Proceedings of the 1996 ACM SIGMETRICS, International Conference on Measurement and Modeling of Computer Systems, May 1996. 6. Şahinoğlu, Z. and Ş. Tekinay, On Multimedia Networks: Self-Similar Traffic and Network Performance, IEEE Communications Magazine, pp.48-52, January 1999. 7. Clausen, H. D., H. Linder and B. Collini-Nocker, Internet over Direct Broadcast Satellites, IEEE Communications Magazine, pp. 146-151, June 1999. 8. Faal, K. and K. Varadhan, The NS Manual, http://www.isi.edu/nsnam/ns/ns-documentation.html 9. Fluckinger, F., Networking Requirements of Audio and Motion Video, Understanding Networked Multimedia: Applications and Technology, Prentice Hall, 1995.
Internet Traffic Queuing Delay: BW 427 kbps Throughput [Mbps].43.425.42.415.41.45.4 Internet Traffic Throughput: BW 427 kbps Buffer 1 2 35 75 Queuing Delay [ms] 4 4 3 3 2 1 Buffer 1 2 35 75.395 Throughput [Mbps].39.385.38.6.7.75.8.85.9.95 1 Parameter Figure 2a. TCP traffic reliable throughput with no video traffic present (Bottleneck link buffer size, packets: 1, 2, 35,, 75,, ).43.425.42.415.41.45.4.395.39.385.38 Internet Traffic Throughput: BW 427 kbps 1 Buffer Size [Packets] Fig. 3a. TCP traffic reliable throughput with no video traffic present ( parameter, H:,,,.4.35 Internet Traffic Throughput.6.7.75.8.85.9.95 1 Parameter Figure 2b. TCP traffic queuing delay without video traffic (Bottleneck link buffer size, packets: 1, 2, 35,, 75,, ). Queuing Delay [ms] 4 4 3 3 2 1 Internet Traffic Queuing Delay: BW 427 kbps 1 Buffer Size [Packets] Fig. 3b. TCP traffic average queuing delay with no video traffic present ( parameter, H:,,, ) 4 4 Internet Traffic Queuing Delay.3 3 Throughput [Mbps].25.2.15 Queuing Delay [ms] 3 2 1.1.5.5 1 1.5 2 2.5.5 1 1.5 2 2.5 Bandwidth [Mbps] Bandwidth [Mbps] Fig. 4a. TCP traffic reliable throughput in the presence of Star Wars video traffic (Buffer: 35 packets, parameter, H:,,, ) Fig. 4b. TCP traffic average queuing delay in the presence of Star Wars traffic (Buffer: 35 packets, parameter, H:,,, )