Wireless Pers Commun (2007) 43:1195–1200 DOI 10.1007/s11277-007-9294-z
Satellite bandwidth allocation based on MPEG-4 videoconference traffic prediction Polychronis Koutsakis
Received: 5 June 2006 / Accepted: 12 March 2007 / Published online: 20 April 2007 © Springer Science+Business Media B.V. 2007
Abstract The provision of acceptable Quality-ofService for multimedia traffic over a geosynchronous earth orbit satellite network demands the existence of a well-designed Medium Access Control protocol. This paper proposes a new dynamic satellite bandwidth allocation technique which is based on accurate MPEG-4 videoconference traffic prediction. The use of the prediction in our protocol is shown via a simulation study to provide very good throughput and delay results.
1 Introduction Videoconference traffic is expected to be a substantial portion of the traffic carried by next generation networks. Video packet delay requirements are strict, because delays are annoying to a viewer; whenever the delay experienced by a video packet exceeds the corresponding maximum delay, the packet is dropped, and the video packet dropping requirements are equally strict. Therefore, a good statistical model can be very useful in evaluating network performance under various videoconferencing loads. Such a model is especially useful in satellite networks, where the propagation delay makes bursty users’ (such as video users’) current traffic profile rather useless for bandwidth allocation, since the profile will probably have changed P. Koutsakis (B) Department of Electrical and Computer Engineering, McMaster University, Hamilton, ON Canada L8S4L8 e-mail:
[email protected]
significantly by the time the bandwidth allocation is made by the Network Control Center (NCC), which we consider to be integrated in the satellite on-board device. The most well-known and used video standards for this application today are H.263 and MPEG-4. In this work, we use a model which accurately captures the behavior of multiplexed MPEG-4 videoconference movies, in order to predict the behavior of videoconference traffic in our new MAC protocol proposal for a Digital Video Broadcasting Return Channel Satellite (DVB-RCS) system.
2 MPEG-4 videoconference traffic modeling In Ref. [1], we have studied four different long sequences of MPEG-4 encoded videos, from the publicly available library of frame size traces of long MPEG4 and H.263 encoded videos provided in Ref. [2]. We have investigated the possibility of modeling the traces with a number of well-known distributions and our results have shown that the best fit among these distributions for modeling a single movie is achieved for all traces examined with the use of the Pearson type V distribution (also known as the inverted gamma distribution). The four traces are, respectively: a video stream extracted and analyzed from a camera showing the events happening within an office (“Office Cam”); a video stream extracted and analyzed from a lecture (“Lecture Room Cam”); a video stream extracted and
123
1196 Table 1 Trace statistics
P. Koutsakis
Movie
Mean frame size (bytes)
Variance-frame size
Pearson type V parameters (α,β)
Office
1,984
9,911
4701,056
(2.84,3646.28)
Lecture
1,048
7,447
828,979
(3.32,2436.92)
ARD Talk
2,763
17,062
2718,144
(4.81,10524.23)
N3 Talk
2,724
15,579
2954,121
(4.51,9568.9)
analyzed from a talk-show (“ARD Talk”); a video stream extracted and analyzed from another talk-show (“N3 Talk”). All four of these traces are movies with low or moderate motion. For each one of the four videos under study we have used the high quality coding version, in which new video frames arrive every 40 ms. The length of the videos varies from 45 to 60 min and the data for each trace consists of a sequence of the number of cells per video frame (we use packets of ATM cell size throughout this work, but our mechanism can be used equally well with packets of other sizes, as the nature of our modeling results would not be altered at all). Table 1 presents the trace statistics for each trace, as well as the parameters (α,β) of the Pearson type V distribution fit for each video. The Probability Density Function (PDF) of a Pearson type V distribution with parameters (α, β) is f (x) = [x −(α+1) e−β/x ] [β−α Γ(α)], for all x > 0, and zero otherwise. The mean and variance are given by the equations: Mean = β/(α-1), Variance = β2 /[(α − 1)2 (α − 2)]. However, although the Pearson V was shown in Ref. [1] to be the better fit among all distributions, the degree of goodness-of-fit for the Pearson V varied significantly, and even in the cases of a quite good fit, the fit was not perfectly accurate. This was expected, as the gross differences in the number of bits required to represent I, P and B frames impose a degree of periodicity on MPEG-encoded streams, based on the cyclic GOP formats. Any model which purports to reflect the frame-by-frame correlations of an MPEGencoded video stream must account for GOP cyclicity, otherwise the model could produce biased estimates of cell loss rate for a network with some given traffic policing mechanism [11]. Hence, in Ref. [1] we proceeded to study the frame size distribution for each of the three different video frame types (I, P , B), in the same way we studied the frame size distribution for the whole trace. The Pearson V distribution once again provided the best fitting results for all types of video
123
Peak frame size (bytes)
frames’ sequences, and the modeling results were much improved in comparison with those of modeling the trace as a whole. As in Ref. [3], where a DAR(1) model with negative binomial distribution was used to model the number of cells per frame of VBR teleconferencing video, in Ref.[1] we built for each video frame type a model based only on parameters which are either known at call set-up time or can be measured without introducing much complexity in the network. DAR(1) provides an easy and practical method to compute the transition matrix and gives us a model based only on four physically meaningful parameters, i.e., the mean, peak, variance and the lag-1 autocorrelation coefficient ρ of the offered traffic (which is typically very high for videoconference sources). We proceeded with testing our models statistically (with methods from Ref. [4]) in order to study whether it produces a good fit for the trace superposition. The good fits in our results in Ref. [1] have shown that the superposition of the actual traces can be modeled well by a respective superposition of data produced by our modeling approach. In this work we will show that our accurate MPEG-4 videoconference modeling can be used very efficiently for proactive resource management in satellite systems.
3 Our MAC protocol proposal The Digital Video Broadcasting Return Channel Satellite (DVB-RCS) standard [5] develops a communication system for the return channel (uplink channel), i.e., the link from the user terminal to the network gateway. Due to the expected services features and large delaybandwidth product of satellite networks, DVB-RCS represents a proper test bed for the proposed resource allocation schemes. The most relevant elements of the DVB-RCS Network are: (a) RCSTs (Return Channel Satellite Terminals), i.e., a generic access terminal, (b) the NCC (Networks Control Center), a device in charge
Satellite bandwidth allocation based on MPEG-4 videoconference traffic prediction Table 2 System parameters Frame duration
26.5 ms
Carriers Slots/frame/carrier Bytes/slot System global rate
4 128 53 8 Mbps
of managing the access and bandwidth allocation for RCSTs, (c) Gateways and Feeders, which are the elements that receive and transmit information outside the network [6]. As in Ref. [6], our proposed satellite medium access scheme is based on a Multi-Frequency Time Division Multiple Access (MF-TDMA) approach, according to which a carrier is divided in timeslots (grouped in frames and superframes). MF-TDMA schemes are capable of providing efficient and flexible bandwidth utilization. The system parameters are taken from Ref. [6] and are presented in Table 2. The NCC allocates to each active RCST a set of timeslots, each characterized by a frequency, bandwidth, start time and duration time. Additionally to the MF-TDMA frame structure, we adopt in our work the idea that, after all requests have been satisfied, the bandwidth left is distributed freely following a certain algorithm (this approach is named in the literature as a Combined Free and Demand Assignment Scheme, CFDAMA scheme); the algorithm implemented in our scheme is a simple round-robin assignment algorithm to all RCSTs which are currently active. This idea was first proposed in Ref. [7] and is especially useful in allocating slots to video users, as the difficulty in providing them with adequate bandwidth due to their frequent changes in bandwidth needs could be somewhat alleviated by their acquiring the unused channel bandwidth freely in a round-robin manner. We use the Combined Free and Demand Assignment Scheme with Piggybacking (CFDAMA-PB) version of the protocol, which was shown in Ref. [8] to be the most efficient way of making reservations. According to the PB strategy, user stations send their capacity requests embedded in the header of their packets. The free capacity distribution performed by the protocol brings the end-to-end delay performance at low loads close to that obtained with random access protocols, while the demand-based bandwidth allocation at the beginning of each frame guarantees the protocol’s stability, robustness and
1197
efficient utilization of transmission bandwidth at high loads. A logical assumption for next generation networks is that videoconference users will be allowed to adopt one of just a few specific “modes” (each corresponding to a set of traffic parameters). This is especially plausible for videoconference traffic, as the number of variations between source bandwidth requirements is naturally restricted by the type of application (a much larger pool of “modes” would have to be used in the case of video traffic). Therefore, in this work we consider that a videoconference user can adopt one of the four “modes” which correspond to the traffic parameters of each one of the four traces under study and are presented in Table 1. Based on the good model for single videoconference traces and the highly accurate model of multiplexed traffic, we propose that the NCC should run a real-time online simulation, both for single and for multiplexed videoconference sources. Hence, based on the “mode” declared by the RCSTs at call establishment, the NCC does not need to wait for a request from the RCSTs every channel frame (which would arrive with a delay of more than five channel frames, due to the propagation delay); instead, it can start allocating resources to the videoconference terminals, by simulating the single source models with the sources’ declared mean rate as a simulation start point, and by computing the free slots in each channel frame [using the DAR(1) models for multiplexed videoconference traffic, and subtracting the estimated used slots from the total number of slots in the system] in order to allocate the estimated number of free slots in a round-robin manner to all active RCSTs. With this slot allocation scheme, the RCST will not need to send frequent requests to the NCC; it will only need to send a “corrective” request every superframe (defined in our work as equal to six channel frames, to account for the propagation delay). The reason for sending this request will be for the RCST to help the NCC correct any mistakes (due to either slots overassignment or underassignment) of the models produced at the NCC via online simulation. After receiving the request, the NCC will resume its simulation with the current RCST state (in terms of bandwidth requirements) as a start point. As it will be shown from our results, this approach, which minimizes the need for signaling among the video RCSTs and the NCC, provides very good results in terms of videoconference users’ QoS requirements satisfaction, due to the
123
1198
quality of the prediction made by our video modeling scheme.
4 Conceptual comparison with relevant work The subject of allocating bandwidth to video users in a satellite system has not been widely studied in the relevant literature. Most of the research on satellite MAC protocols has focused on handling data traffic and on integrating voice and data traffic over satellite uplink channels of limited capacity, and, to the best of our knowledge, no previous work has studied the problem of transmitting MPEG-4 videoconference traffic over GEO satellite links. However, there are a few schemes in the literature which have addressed the issue of providing acceptable QoS to video users. We will compare our work conceptually with two such schemes, the ones proposed in Ref. [9,10]. In Ref. [9], the authors propose a scheme which uses n levels of allocation to MPEG-2 video users. Each of the levels represents a threshold for the user’s throughput; if the data currently transmitted by the user is below a threshold, the user gives up a portion of its allocated bandwidth in favor of other users which are currently in need of more bandwidth resources; if the user’s throughput exceeds a threshold, the user regains the bandwidth which was initially allocated to it. The disadvantages of this scheme concern: (a) the signaling delay for giving up and regaining bandwidth, (b) the large number of levels needed for ideal control of slot allocation and deallocation (the authors produce results with various numbers of levels, reaching up to 100 levels), which add a very significant computational load to the system (c) the authors use for their results only one, synthetically generated MPEG-2 compressed movie (whereas we use actual video traces in our work); the implementation of their mechanism for a large number of movies (i.e., for users transmitting various movies at various qualities) would need a definition of a different large number of levels for each one of the movies, as the good choice of thresholds is of vital importance to the proposed mechanism, hence yielding this approach impractical. In Ref. [10], the authors propose a satellite MAC protocol for transmitting MPEG-1 video over satellite channels. Their work is based on the assumption that the video stream at the Group of Pictures (GOP) level could be thought of as a sequence of scenes, and a
123
P. Koutsakis
scene change occurs when a significant change occurs in the GOP size; therefore, a new bandwidth allocation should take place only at scene changes, hence decreasing the frequency of signaling between user terminals and the NCC. For this reason, the authors propose the extraction of statistics of the slow time process and the fast time process of MPEG-1 movies in order to derive a metric for allocating bandwidth to video users. However, the packet flow metric used by the authors depends on certain parameters (e.g., the time window in which the metric is measured), the ideal value of which will actually depend on the burstiness of the movie, therefore these parameters need to be different for different movies, as pointed out in Ref. [10]. This complexity limits the practicality of the authors’ scheme, at least in its ideal form (it could be implemented with a compromise on a fixed time window length for all movies transmitted and at all qualities for each movie, but this would lead to inferior system performance). Also, the basic idea behind the authors’ scheme, i.e., the idea of tracking scene changes, does not apply in our work which is focused on videoconference traffic (the usual case in this type of traffic is that there are minimal or no scene changes). Finally, the authors in Ref. [10] conclude that if video packet loss as high as 3% is tolerable, their scheme achieves considerably lower delays than other known satellite MAC protocols. However, the maximum video packet dropping of 3% is quite high; our scheme will be shown to provide very good throughput results for a significantly stricter maximum allowed video packet dropping of 0.1%.
5 Results and discussion At the start of our simulation study, we let each videoconference user choose one of the four traffic parameter sets (“modes”), with equal probability. The average video packet dropping is set to a maximum of 0.1% in our work and the maximum end-to-end video packet delays is set to 0.6 seconds, which is especially strict considering that, for each possible failure of our prediction due to underassignment, the respective packets which would have to wait for a new assignment will have a minimum end-to-end video packet delay of 0.54 s, due to the propagation delay in a GEO satellite channel. Two MAC schemes are compared in this work. The first is our proposed scheme, and the second is an
Fig. 1 Average video packet dropping versus system utilization
Video Packet Dropping (%)
Satellite bandwidth allocation based on MPEG-4 videoconference traffic prediction 1,000 0,900 0,800 0,700 0,600 0,500 0,400 0,300 0,200 0,100 0,000
1199
Our M AC Protocol Ideal Assignment
10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94
Fig. 2 Average End-to-end video packet delay versus system utilization
Average End-to-End Video Packet D elay (s econds)
Normalized Throughput (%)
1,000 Our Mac Protocol
0,800
Ideal Assignment
0,600 0,400 0,200 0,000 10
15
20 25
30
35
40 45
50
55
60 65
70
75
80 85
90
95
Normalized Throughput (%)
“ideal” scheme, as we want to compare our protocol with a similar one in which the NCC would “know,” without any information exchange (therefore, no contention is necessary among video RCSTs), exactly what the video RCSTs’ bandwidth demands for the next video frame will be. Figure 1 presents our simulation results for the average video packet dropping metric versus the system utilization. Utilization indicates the traffic load normalized to the uplink capacity, e.g., a traffic load equal to 40% represents 40% of the 8 Mbps uplink capacity, i.e., 3.2 Mbps system throughput. As it is shown in the Figure, the difference in video packet dropping between our scheme and the “ideal” case is so small that it can be considered almost negligible for all normalized video traffic loads. Our scheme can handle up to 68% system load while at the same time satisfying the strict QoS requirement of maximum video packet dropping equal to 0.1%; the respective maximum system load which the “ideal” scheme can handle is 74%. The reason that none of the two schemes can achieve a higher throughput is the high burstiness of video traffic; in certain channel frames, video bursts from more than one RCST happen to take place simultaneously in the uplink channel. Although our traffic modeling scheme can often predict such bursts, the total amount of requested bandwidth in certain channel frames may surpass the system’s available capacity; this will lead to inevitable video packet dropping, as some of the packets may will
not be sent within the roughly one and a half channel frames which pass before the arrival of the next video frame (when a new video frame arrives, all packets of the previous video frame which have not yet been sent are discarded). Figure 2 presents our simulation results for the average end-to-end video packet delay versus the system utilization. The results are generally similar in nature with those of Fig. 1, denoting that our scheme’s results are very close to the ones achieved by the “ideal” scheme. However, it should be pointed out that: (a) as the system load increases, the “ideal assignment” scheme achieves a lower delay of about 0.15–0.2 s in comparison to our scheme, due to the lack of contention (and therefore, lack of collisions) in the ideal scenario, (b) by studying Figs. 1 and 2, it is clear that for a 68% system load only the video packet dropping metric surpasses its maximum set value in our scheme (the average video packet delay is lower than 0.6 s); similarly, for a 74% system load only the video packet dropping metric surpasses its maximum set value in the “ideal” scheme (the average video packet delay is again lower than 0.6 s). Therefore, since the strictest metric is shown to be that of the maximum allowed packet dropping, our future work will focus on the subject of introducing an efficient Forward Error Control (FEC) mechanism, which will allow for larger maximum video packet dropping without adding a large overhead to the information transmitted in the uplink.
123
1200
References 1. Lazaris, A., Koutsakis, P., & Paterakis, M. (2006) On Modeling Video Traffic from Multiplexed MPEG-4 Videoconference Streams. In Proceeding of the 6th international conference on next generation teletraffic and wired/wireless advanced networking (NEW2AN) Vol. 4003, 2006, pp. 46–57. St. Petersburg, Russia, published in Lecture Notes in Computer Science, Springer-Verlag. 2. http://www-tkn.ee.tu-berlin.de/research/trace/trace.html. 3. Elwalid, A., Heyman, D. P., & Lakshman, T. V., Mitra, D., & Weiss, A. (1995). Fundamental bounds and approximations for ATM multiplexers with applications to video teleconferencing. IEEE Journal on Selected Areas in Communications, 13(6), 1004–1016. 4. Law, A. M., & Kelton, W. D. (1991). Simulation modeling & analysis (2nd ed.,). New york:McGraw Hill Inc. 5. Std., ETSI EN 301 790 V1.2.2 (2000-12). 6. Chiti, F., Fantacci, R., & Marangoni, F. (2005). Advanced dynamic resource allocation schemes for satellite systems. In Proceedings of the IEEE international conference on communications (ICC) 2005 (Vol. 3, pp. 1469–1472). Seoul, Korea. 7. Krishnamurthy, S. V., & Le-Ngoc, T. (1995). Performance of CF-DAMA protocol with pre-assigned request slots in integrated voice/data satellite communications. In Proceedings of the IEEE international conference on communications (ICC) 1995 (Vol. 3, pp. 1572-1576). Seattle, USA. 8. Le-Ngoc, T., & Krishnamurthy S. V. (1996). Performance of combined free/demand assignment multiple access schemes in satellite communications. International Journal of Satellite Communications, 14(1), 11–21. 9. Celandroni, N., Ferro, E., & Potorti, F. (2002). A multi-level satellite channel allocation algorithm for real-time VBR data. International Journal of Satellite Communications, 20(1), 47–61. 10. Connors, D., Ryu, B., Pottie, G. J., & Dao, S. (2002). A medium access control protocol for real time video over high latency satellite channels. Mobile Networks and Applications, 7 (1), 9–20. 11. Frey, M., Ngyuyen-Quang, S. (2000). A gamma-based framework for modeling variable-rate video sources: The GOP GBAR Model. IEEE/ACM Transactions on Networking, 8(6), 710–719.
123
P. Koutsakis Polychronis Koutsakis was born in Hania, Greece, in 1974. He received his 5-year Diploma in Electrical Engineering in 1997 from the University of Patras, Greece and his MSc and Ph.D. degrees in Electronic and Computer Engineering in 1999 and 2002, respectively, from the Technical University of Crete, Greece. He was a Visiting Lecturer at the Electronic and Computer Engineering Department of the same University for three years (2003–2006). He is currently an Assistant Professor at the Electrical and Computer Engineering Department of McMaster University, Canada. His research interests focus on the design, modeling and performance evaluation of computer communication networks, and especially on the design and evaluation of multiple access schemes for multimedia integration over wireless networks, on call admission control and traffic policing schemes for both wireless and wired networks, on multiple access control protocols for mobile satellite networks, wireless sensor networks and powerline networks, and on traffic modeling. Dr. Koutsakis has authored more than 55 peer-reviewed papers in the above mentioned areas, has served as a Guest Editor for an issue of the ACM Mobile Computing and Communications Review, as a TPC member for conferences such as IEEE GLOBECOM 2006, IEEE GLOBECOM 2007, IEEE LCN 2006, IEEE LCN 2007, and IEEE PerCom 2006, as Session Chair for the IEEE GLOBECOM 2006 Symposium on Satellite & Space Communications and the IEEE WCNC 2007 and serves as a reviewer for most of the major journal publications focused on his research field. He is a member of the IEEE.
