Delivering ATM-based Video-on-Demand across Asymmetrical Digital Subscriber Lines Emile Swanson Neco Ventura Technology Integration, Telkom SA Ltd Dept of Electrical Engineering, University of Cape Town Private Bag X74, Pretoria, 0001 Private Bag, Rondebosch, 7701 South Africa South Africa SwansEW1@telkom.co.za email@example.com ABSTRACT - ADSL technology provides a means for discussed in Sections III and IV respectively. Section III delivering high bit-rate services to residential subscribers. discusses the need for reliable resource management, Current service offerings are predominantly IP-based especially in the access network. Two models for ensuring this services such as high-speed Internet access. Such services are presented and evaluated in terms of their suitability for do not fully exploit the potential offered by the ADSL various situations. Section IV focuses on shaping the video technology. This paper investigates the implications of stream to ensure efficient use of the available network using the ATM technology that forms part of the ADSL resources. MPEG-1 compression is used as the video format, standard, for delivering high-quality Video-on-Demand. as it is a widely used, open standard. Three lossless shaping Various problems pertaining to delivering such a service techniques are presented and evaluated in terms of across ADSL are identified, and solutions are proposed. computational complexity and network efficiency. Section V Evaluations of these solutions show improvements in both describes the implementation of a test-bed combining one of reliability and bandwidth efficiency. Proposals are also the models from Section III with the shaping techniques from made for future research to extend these findings towards Section IV, and test results are presented for various practical / commercial implementations. situations. Section VI draws conclusions based on the findings presented in this paper, and Section VII presents directions for Keywords - ATM, Multimedia, Video-on-Demand, xDSL future research. I. INTRODUCTION II. ADSL NETWORK ARCHITECTURE Asymmetrical Digital Subscriber Line (ADSL)[1,2] technology In order to identify the factors affecting the performance of is seen as a cost-effective means of delivering high bit-rate ATM-based VoD over an ADSL link, a basic network content to residential subscribers. ADSL is an overlay network architecture is used, comprised of only those elements that leverages existing local-loop copper infrastructure, and necessary for delivering the video stream. This architecture is employs advanced digital signal processing to allow high bit- shown in Figure 1. rate operation across this infrastructure. Research has been done into the effect of using ADSL as a "last mile" technology, focussing mainly on the performance of TCP/IP over ADSL[3,4]. While these findings benefit TCP/IP-based applications such as high-speed Internet access, which are currently the main applications of ADSL, they do not fully explore the possibilities offered by ADSL. ADSL was ATM / ADSL Video-on- originally designed for streaming video , as indicated by the End-Systems DSLAM ATM Network Demand Server error-reduction techniques employed, low latencies offered, and relatively high downstream data rates offered. In addition, Figure 1: Simplified ATM/ADSL Network the use of ATM as the Layer 2 technology for ADSL further suggests that this technology is highly suited for an application From Figure 1 it can be seen that the ADSL network closely with strict network resource requirements such as Video-on- resembles a conventional ATM network, with the addition of Demand (VoD). an ADSL link in the access leg. This link serves as a physical medium for transporting the ATM traffic to the end-user. This paper explores the factors affecting the delivery of video However, it should be noted that ADSL was designed as a content across ADSL links, and proposes and evaluates transparent technology, and that the ADSL link transparently solutions for optimizing such a content delivery system, in bridges the information between the end-user's 25.6Mbps terms of network efficiency as well as received video quality. ATM link, via the ADSL line, and the backhaul link from the Section II of this paper presents a simplified ADSL network DSLAM, which may be an STM-1 connection. This means, architecture containing only those elements necessary for amongst other things, that the ATM network does not know delivering video content. From this model the factors affecting the ADSL resources, and that the ADSL equipment does not streaming video delivery are identified. These factors may be participate in connection admission control (CAC). Figure 2 addressed either at the network- or application-level, as illustrates the Q.2931 signaling for a virtual circuit being created from the end-system, through the ADSL network, to line through a direct connection to the ATU-R. The end- the video server. It can be seen that the ADSL equipment does system would then be able to integrate this information not participate at any point, and that the ADSL link's resources into its ATM stack for more reliable resource are not considered at any point in the CAC process. ATM management. This option may not always be available, as Quality-of-Service (QoS) can therefore not be guaranteed it is heavily dependent on whether the ATU-R supports when operating across an ADSL network. Section III of this such access to its management information. Current ATU- paper presents methods of addressing this shortcoming at the Rs also have various ways of providing this access, so the network level, to ensure ATM QoS in an ADSL network. implementation would be tied to a specific set of ATU- Rs. ADSL End- ADSL ATM ATM VoD System Access Network Switch A Switch B Server (ATM) 2. Querying the DSLAM via in-band management. The end-system may be able to access the management SETUP SETUP information stored on the DSLAM, and use this SETUP information to update its own ATM stack. Disadvantages CALL of this approach are that it compromises security by CALL PROCEEDING CALL PROCEEDING PROCEEDING CONNECT allowing access to the ADSL network management system, and that the end-system needs detailed CONNECT CONNECT ACK information about which line's details to query, e.g. shelf, slot and port number. CONNECT CONNECT ACK 3. Querying the DSLAM via out-of-band management. This method is similar to that described above, with the CONNECT ACK main difference being the use of out-of-band management. This method uses IP-based SNMP to query Figure 2: Q.2931 Signalling for successful SVC setup the DSLAM. In addition to the disadvantages for the in- band method, this method also needs the IP networking It should also be noted that an ADSL network imposes stricter stack to be initialized before the ATM resource bandwidth constraints than other fixed-line ATM networks, management can be fully initialized. Fortunately, IP is a particularly for bursty traffic such as compressed video. best-effort service, and therefore typically uses a UBR Research has addressed the requirements of delivering VoD in class-of-service across an ADSL link. This means that the ATM networks[8,9,10,11,12,13,14]. However, much of this research end-system's ATM resource management stack may be focuses on achieving statistical multiplexing gains. This passed an updated set of line parameters without research is not applicable to ADSL, as the limited bandwidth negatively affecting the performance of the existing IP available on an ADSL link (up to 8Mbps downstream) does connectivity. not allow for multiplexing video streams. Instead, the traffic stream should be shaped to minimize the peak bandwidth The primary advantage of using a distributed solution is that it required. Ideally this shaping should be done at the application may be implemented in software, by extending the ATM level, as this will allow more intelligent shaping based on the networking stack at the end-system. This type of solution is characteristics and requirements of the traffic. Section IV suited to existing networks, as it does not require costly addresses this problem by presenting and evaluating lossless hardware replacement or upgrades. Software implementations shaping techniques that impose minimal additional may also be developed and deployed more quickly than computational load on both the server and the client systems. hardware-based solutions. III. NETWORK-LEVEL RESOURCE MANAGEMENT B. Centralized Solution As discussed above, current ADSL equipment does not In this solution, the DSLAM is extended to contain more participate in resource management activities such as CAC, complete ATM switching functionality than current systems, making it impossible to guarantee ATM QoS. This problem effectively evolving into an ATM switch. The DSLAM now may be overcome by including the ADSL resources in the uses the information about the specific ADSL line, in CAC already being performed at one of the participating ATM conjunction with information about the backhaul ATM link, nodes, using either a distributed or a centralized solution. when creating new virtual connections. The DSLAM therefore acts as a proxy for each end-system, assisting with the CAC A. Distributed Solution process. An additional advantage is that, as an ATM node, the In this solution, the ATM resources of the ADSL line are DSLAM would also be able to form an active part of an ATM managed by the end-system's ATM stack. The end-system multicast tree, improving the efficiency of multicasting to determines the available resources on the line in one of three ADSL end-users. While this solution offers the advantage of ways, listed below in decreasing order of preference: resource management at a single, central point, it comes at the cost of increased hardware complexity in the DSLAM. This 1. Querying the ATU-R. solution would therefore be more suited to new networks than Ideally the end-system would be able to access existing installations. information about the available resources on the ADSL In this research, the solution presented in Section III.A.3 has A. Short-Term Shaping been implemented and evaluated. Details about this are When studying an MPEG-1 stream, two things are noticed. provided in Section V. Firstly, the I-frames are typically largest and therefore contribute most to the peak bandwidth requirement of the IV. APPLICATION-LEVEL TRAFFIC SHAPING traffic stream. Secondly, the I-frames are always preceded by Applications operating across limited capacity media such as B-frames, which are smallest. It should therefore be possible ADSL should minimize their network resource requirements, to reduce the peak requirement of the stream by buffering and specifically their peak bandwidth requirement. This should transmitting the average of an I-frame and its preceding B- ideally be done at the application level, as the application frame during their original transmission periods. The sequence possesses detailed information about the traffic stream, and now becomes …B(BI)'(BI)'BBPBBPBBPBBPB(BI)'(BI)'B…, can therefore perform more intelligent traffic shaping than the where (BI)' refers to the averaging of the I- and preceding B- network. MPEG-1 compressed video, for example, has a frame. Figure 4 shows the effect of applying this technique on repetitive structure, as shown in Figure 3. the sequence in Figure 3. The figure shows a noticeable reduction in the peak bandwidth requirement, from 25610 25000 bytes per frame (in the GoP in Figure 3) to 16090 bytes per frame (in the smoothed GoP in Figure 4). Specific details on 20000 improvements obtained with this technique are provided in Section V. 15000 16,000 10000 14,000 12,000 5000 10,000 0 8,000 I B B P B B P B B P B B P B B 6,000 Figure 3: Relative frame sizes and sequence for MPEG-1 4,000 2,000 As can be seen, the stream consists of three frame types (I-, P-, - and B-frames), with a regularly repeating sequence, arranged (BI)' B B P B B P B B P B B P B (BI)' to form Groups-of-Pictures (GoPs). This sequence is described in terms of M and N parameters, referring to the spacing Figure 4: Reduced peak bandwidth after short-term between consecutive I-frames, and I- and P-frames smoothing respectively, e.g. the sequence …PBBIBBPBBPBBPBBPBBIBBP… etc would be described B. Medium-Term Shaping as M=15, N=3. These frames can also be characterized relative to each other in terms of size. I-frames are typically largest, Further investigation of the GoP used in Figures 3 and 4 followed by P-frames, with the B-frames being smallest. suggest that further reductions in peak requirements may be Example ratios for I:P:B frame-sizes are 6:2:1 (for low-motion obtained by using the remaining B-frames to reduce the peaks streams) or 4:4:1 (for higher-motion streams). The average bit- due to the large I-frames. In addition, the peaks due to the rate for an MPEG-1 compressed stream is also known. medium-sized P-frames may be reduced in the same manner. Together, these parameters may be used to describe an This method effectively buffers the entire GoP and transmits MPEG-1 stream. A video server may use the information at the much lower average rate of the GoP. Figure 5 illustrates provided by these parameters to shape the traffic stream the reduced peak requirement of the smoothed GoP. Note that Figure 5 spans 3 GoPs, to show the smoothing per GoP. Note As stated previously, an advantage of performing traffic also that "GoP A" in the figure refers to the GoP used in shaping at the application level is that the server can use a Figures 3 and 4. priori information (e.g. frame-type sequence, relative frame sizes) to perform intelligent shaping. While other researchers A potential disadvantage of using this form of traffic shaping use lossy shaping techniques, in this research we have is that delays are required for the buffering of the stream prior elected to use only lossless shaping techniques, thereby to transmission. During the implementation and evaluation of ensuring no degradation of video quality due to the presence this technique, however, it was found that it is possible to of the limited-bandwidth ADSL link. employ the existing de-jittering buffers in standard MPEG-1 servers and clients without imposing additional delays or Based on the structure of an MPEG-1 compressed stream computational load. Details of the specific improvements (described above), the following three types of lossless obtained with this technique are provided in Section V. shaping techniques may be used, in increasing order of the expected reduction of peak requirements: GoP B: Peak =7414 GoP C: Peak =7452 The test-platform was built around a Siemens XpressLink D bytes bytes 8000 v1.0 DSLAM, connecting two ATM/ADSL clients to an GoP A: Peak =6491 bytes ATM-based MPEG-1 video server via an ATM network. A 7000 development version of the Linux operating system (kernel 6000 version 2.3.41) with ATM extensions (version 0.66) was used on the client machines. The resource management extensions 5000 were integrated into the device driver software installed at the 4000 ATM/ADSL end-systems. All three smoothing techniques were implemented in a custom-developed MPEG-1 video 3000 server-client system, developed with the SDL multimedia 2000 development libraries. In order to create switched virtual circuits (SVCs) from either of the end-systems, paths were 1000 mapped through the DSLAM to the edge ATM switch, and ILMI was activated on the additional paths. SVCs were 0 used to confirm the acceptance or rejection of call-setup Figure 5: Reduced peak variations after medium-term requests based on available resources. smoothing In the first stage of testing, the end-systems attempted to create SVCs exceeding the available resources on the ADSL C. Long-Term Shaping link. Before enabling the improved resource management, all Long-term shaping of compressed video traffic is discussed in SVCs were accepted by the ATM network. Data transfers , with the emphasis on renegotiating the allocated network across these links were possible, but data losses and reduced resources to match the requirements of the smoothed traffic. throughput due the associated retransmissions were observed. This method allows the reserved bandwidth to be dynamically With the resource management mechanisms enabled, requests tailored to match the traffic stream, in order to obtain high for resources in excess of those available on the link were bandwidth efficiency through the network for compressed rejected by the network, ensuring reliable operation of streams in general. In this research however, we focus connections across the link. specifically on MPEG-1 compressed video, and can therefore simplify our long-term smoothing by using our knowledge of Having confirmed correct operation of the underlying the stream's structure. network, the second stage of testing focussed on evaluating the performance of the application-level shaping mechanisms. MPEG-1 compression works by reducing spatial and temporal redundancy at a frame- and GoP-level. However, each GoP is The shaping techniques described in Section VI were effectively encoded independent of the rest of the stream, for implemented on the video server and clients. Two source error-resilience amongst other reasons. An advantage of this streams were used (SeatSpin.mpg and AlienSong.mpg), as compression method is that the average bandwidth samples of high-motion and low-motion streams respectively. requirement of each GoP does not deviate from that of the "SeatSpin.mpg" was also specifically chosen to represent a file overall stream as much as the individual frames do. It should having high instantaneous requirements, to determine how therefore be possible to transmit the entire stream at the effectively the smoothing techniques would handle large overall average rate without imposing heavy buffering peaks. More detailed information on these streams are requirements as suggested in . This method effectively provided in Tables 1 and 2. converts the original (highly bursty) variable-bit-rate (VBR) video stream to a constant- bit-rate (CBR) for transporting MPEG Movie Info (SeatSpin.mpg) across the ADSL link, thereby minimizing the peak Audio Bitrate (average) 176400 requirements of the stream. The evaluation in Section V shows Picture Size 352 x 288 that this technique is practically applicable, both for low- and Video Bitrate (Average) 1152000 high-motion streams. Frame rate 25 File Size 3897998 V. IMPLEMENTATION AND PERFORMANCE EVALUATION I:P:B ratio 4:4:1 In order to verify the operation of the resource management Table 1: Stream information for SeatSpin.mpg solutions described in Section III, the technique described in Section III.A.3 was implemented and evaluated on an ADSL MPEG Movie Info (AlienSong.mpg) test-bed in the Communications Research Group at the Audio Bitrate (average) 88200 Picture Size 320 x 240 University of Cape Town. This technique was chosen for Video Bitrate (Average) 480000 implementation due to limited time, limitations on the Frame rate 24 available equipment and the lack of access to the hardware File Size 3177818 specifications required for implementing the DSLAM-based I:P:B ratio 6:2:1 solution. The implementation and evaluation of the remaining Table 2: Stream information for AlienSong.mpg solutions are left for future research projects. Statistical analysis of the streams used showed that the peak two categories, namely network-level and application-level, buffer requirements for each of the smoothing techniques and then addressed as such. This paper ensures reliable never exceeded the size of one GoP (for the two streams network operation (in terms of QoS guarantees) by proposing used). However, due to the small sample set used, it was and implementing the network-level solutions, before decided to dimension the buffers to accommodate up to four addressing the application-level problems. By integrating both (4) GoPs, with the operating target being 50% buffer classes of solution into a system, we have shown that it is utilization. Table 3 provides information on the buffer sizes indeed possible to reliably deliver high quality ATM-based (in bytes) and associated delays introduced (in seconds). streaming video across ADSL links, without imposing significant computational or storage requirements on either the Buffer size Latency server or the client systems. (bytes) (sec) SeatSpin.mpg 398520 1.20 VII. DIRECTIONS FOR FUTURE RESEARCH AlienSong.mpg 177563 1.25 The remaining models proposed in Section III should be implemented and evaluated relative to the model used in Table 3: Buffer requirements and latencies / delays Section V. All the models should also be evaluated to introduced for two test streams determine applicability to various scenarios. Factors to be taken into account include network size and structure, number The results of tests under various ADSL line conditions were of users, traffic volumes, redundancy and reliability, and cost as expected. SVC setup requests were correctly accepted or implications. rejected based on the availability of reserve capacity on the ADSL link, with no cell losses due to congestion on the ADSL A more detailed study of the statistical properties of MPEG-1 link. The smoothing techniques also performed as expected, streams is also required. The implementation described in this with the long-term shaping providing the greatest reduction in paper makes assumptions based on descriptive statistics of the peak requirements, down to 18% of the original peak stream and achieves acceptable results. However, the accuracy requirements for the "AlienSong.mpg" stream. Error detection of the buffer size calculation becomes more critical as the size was enabled at the video client applications, but no stream of the network being served by the video server grows. It errors were detected. It was originally intended to conduct a would therefore be desirable to have an empirical formula to subjective assessment of the delivered video stream as well, accurately calculate the minimum required buffer sizes for an but this was deemed unnecessary since the objective measures arbitrary stream. were reporting error-free operation. It was therefore accepted that the implementation was performing as intended, VIII. REFERENCES combining reliable resource management and intelligent  International Telecommunication Union, "ITU-T traffic shaping to deliver high-quality video across an ADSL Recommendation G.992.1 - Asymmetrical Digital link. Tables 4 and 5 list the results obtained when using the Subscriber Line (ADSL) Transceivers" "SeatSpin.mpg" and "AlienSong.mpg" respectively.  American National Standards Institute, “ANSI T1.413 Issue 2” Smoothing PCR Relative Reduction  Hari Balakrishnan, Venkata N. Padmanabhan, "How None 13339 - - Network Asymmetry Affects TCP", IEEE Communications Short-Term 9790 73% 27% Magazine, April 2001, pp60-67. Medium term 4247 32% 68%  Ryoichi Kawahara, Hiroshi Saito, "Performance of Long-Term 3665 27% 73% TCP/IP over ATM over an ADSL", IEICE Transactions on Communication, Volume E83-B, No.2, February 2000, Table 4: Reduced peak bandwidth requirements (in pp140-154. cells/sec) after smoothing SeatSpin.mpg  ADSL Forum, "ADSL Forum: Technical FAQs", available at http://www.dslforum.org/aboutdsl/faq.html Smoothing PCR Relative Reduction  International Organisation for Standardisation, "MPEG-1 None 8625 - - Coding of moving pictures and associated audio for Short-Term 4668 54% 46% digital storage media at up to about 1,5 Mbit/s" Medium term 1739 20% 80%  International Telecommunication Union, "ITU-T Long-Term 1564 18% 82% Recommendation Q.2931 - Digital Subscriber Signalling System No. 2 - User-Network Interface (UNI) layer 3 Table 5: Reduced peak bandwidth requirements (in specification for basic call/connection control" cells/sec) after smoothing AlienSong.mpg  Marwan Krunz, "Bandwidth Allocation Strategies for Transporting VBR Video Traffic", IEEE Communications VI. CONCLUSIONS Magazine, January 1999, pp40-46.  Pedro Cuenca, Antonio Garrido, Fransisco Quiles, Luis Although a number of impairments still prevent the proper Orozco-Barbosa, “Error-Resilient Video Transmission operation of ATM over ADSL links, it is possible to address over ATM Networks”, IEEE Communications Magazine, these problems. The problems may be broadly classified into December 1999, pp106-111.  Marwan Krunz, Satish K. Tripathi, "Exploiting the Temporal Structure of MPEG Video for the Reduction of Bandwidth Requirements", IEEE Infocom'97 Proceedings Volume 1, pp67-74.  T. V. Lakshman, Partho P. Mishra, K. K. Ramakrishnan, "Transporting Compressed Video over ATM Networks with Explicit Rate Feedback Control", IEEE Infocom'97 Proceedings Volume 1, pp38-47.  Marcel Graf, "VBR Video over ATM: Reducing Network Resource Requirements through Endsystem Traffic Shaping", IEEE Infocom'97 Proceedings Volume 1, pp48- 57.  Patrick Vine, Neco Ventura, "Maximising the simultaneous users of a Video on Demand Server", SATNAC98 Proceedings, pp329-336.  Wu-chi Feng, Jennifer Rexford, "A Comparison of Bandwidth Smoothing Techniques for the Transmission of Prerecorded Compressed Video", IEEE Infocom'97 Proceedings Volume 1, pp58-66.  Nicholas Yeadon, "Supporting Quality of Service in Multimedia Communications via the use of Filters", BT Labs, Ipswich/ Lancaster University Internal Report  ATM Forum, "Integrated Local Management Interface (ILMI) Specification", af-ilmi-0065.000 Author's Biography: Emile Swanson Emile Swanson completed his B.Sc(Eng) degree at the University of Cape Town in December 1998. He joined the Communications Research Group in the Department of Electrical Engineering at the University in 1999, and is studying towards an M.Sc(Eng) degree, focussing on ATM and ADSL technologies. He is currently employed as an electrical engineer at Telkom SA, working with ATM integration issues.
Pages to are hidden for
"Delivering ATM Video on Demand across ADSL"Please download to view full document