Shared access packet transmission systems for compressed digital video

Share Embed


Descrição do Produto

815

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 7, NO. 5, JUNE 1989

Shared Access Packet Transmission Systems for Compressed Digital Video KURIACOSE JOSEPH,

DIPANKAR RAYCHAUDHURI, JOEL ZDEPSKI, MEMBER, IEEE

MEMBER, IEEE,

Abstract-This paper presents an investigation of issues related to the support of real-time compressed digital video on shared access packet transmission systems. Packet transport mechanisms are motivated for compressed digital video because of their inherent ability to average data from many bursty variable rate sources, potentially eliminating the need for buffer controlled adaptation of the video coding algorithm. The feasibility of employing conventional link- and transport-level protocol services to transmit compressed video is examined by focusing on two specific practically important scenarios for compressed video transmission: 1) multipoint-to-multipoint video transmission using a 200 Mbit/s implicit token passing (ITP) fiber-optic LAN and 2) point-to-multipoint broadcast video distribution using a 90 Mbit/s packet-TDM direct-broadcast satellite channel. In order to evaluate the performance of such shared access broadband packet video systems, we develop accurate simulation models, driven by realistic “broadcast quality” AST-DPCM compressed video sources, for the example ITP-LAN and packet-TDM systems. The models are used to determine design tradeoffs between channel throughput, video quality (measured by clipping probability), and the transport-level and mediaaccess-level protocol features and parameters implemented in the packet video network interface unit (NIU).

I. INTRODUCTION N this paper, we examine the problem of designing shared access packet-transport-based transmission systems for compressed video signals. It has long been recognized that the relatively high transmission capacity of 75-100 Mibts/s required by an uncompressed PCM video signal cannot be economically supported in a variety of application scenarios. This has motivated, over the last two decades, considerable activity in the field of image source coding/compression [ 11, [2], resulting in the widespread application of intra-interframe DPCM [3]-[5] and transform coding (DCT, Hadamard, etc.) 161-[8] techniques for rate reduction. Based on the advances made in image coding, it is now possible to transmit broadcast quality motion video in the range of 15-30 Mbits/s [9][ 101, while teleconferencing quality generally requires as little as 0.384-2.048 Mbits/s [ 111. Although such image coding algorithms usually produce variable output data rates, transmission systems for compressed digital video have traditionally been designed to operate on a continuously allocated channel with a given capacity R, bits/s.

I

Manuscript received April 14, 1988; revised January 16, 1989. This paper was presented in part at GLOBECOM 1988, Hollywood, FL, 1988. The authors are with the David Sarnoff Research Center, Princeton, NJ 08543-5300. IEEE Log Number 8927589.

SENIOR MEMBER, IEEE, AND

The process of converting. the variable rate image source to a’ fixed channel rate rnvolvesrate buffer adaptive encoding [l2I9which tends to increase video dec implementation complexity while also resulting in undesirable scene-sequence-dependent image quality artifacts. The inherently variable rate output of a compressed video source motivates the use of packet transport mechanisms in which the channel capacity used by any given source can be varied on demand. In the past, packet transport of video was not considered a viable option because of the limited total bandwidth of the previous generation of packet networks which were designed for the transmission of data. However, with the recent emergence of broadband networks using fiber or satellite media, it is now more feasible to consider packet transport of video signals. For example, even with a relatively low-speed (150-250 Mbit/s) multimode fiber-optic local area network (LAN), it is possible to support simultaneously several ( N 10) “broadcast” quality or a large number ( N 100-200) of “teleconference” quality video transmissions. Another important scenario, recentky considered in 1131, relates to point-to-multipoint broadcast distribution of digital video over a high-speed packet multiplexed (TDM) fiber-optic or satellite channel. For example, a Kuband satellite transponder operating at 90 Mibts/s could be used for direct-to-home (DBS) broadcast of N - 2-5 television channels, while a 600 Mbit/s fiber would support as many as 15-30 channels. Several potential benefits are expected to accrue from the use of packet transport for compressed video, as outlined below. 1) Packet switching provides a great deal of service flexibility [ 141, [ 151 and generally involves lower switching costs, particularly when multicast or broadcast delivery is involved. 2) Subjectively desirable constant image quality can be provided due to the availability of a variable capacity channel to absorb rate variations. 3) Improved performance relative to conventional circuit switching may be achieved in terms of either a) higher efficiency of channel use at a given level of video quality (i.e., if a fixed assigned channel of speed R, can be used to provide compressed video of quality level Q,, then a broadband channel of speed R = NR, can support M > N packet video signals at the same quality level Q , ) or b) better video quality at a given transmission capacity per video signal (i.e., M = N packet video signals can be 9

-

0733-8716/89/0600-0815$01.OO

-

O 1989 IEEE

816

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 7. NO. 5, JUNE 1989

supported on a broadband channel of speed R = NRo with quality level Q > eo).The potential for such improvements in capacity and/or video quality is substantiated by recent results in [ 161, [ 171. 4) Lower codec implementation complexity may result because the variable transmission capacity potentially eliminates the adaptive rate control logic and large rate buffers required in constant rate video codecs. While the above potential benefits of packet video are quite attractive, it is clear that in many cases standard packet network interfaces and parameters may not support the unique requirements of compressed video transmission. First of all, the real-time nature of video signals tends to limit transmission of packet video to certain classes of broadband networks with a high accessible bandwidth and low delay variance. Even when consideration is restricted to such suitable classes of broadband networks, it is necessary to develop new transport protocols, or at least appropriately modify existing transport-, network-, and link-level (including media access and error control) parameters, in order to support packet video successfully. Thus, an important problem, which is addressed here, is to determine the extent to which conventional broadband networks and their protocols apply to packet video and, where necessary, to suggest easily implemented modifications that would help to accommodate compressed video. Fig. 1 shows the generic structure of a packet video system which uses a broadband packet network for transmission. The packet video network interface unit (NIU) serves as the interface between a compressed video encodeddecoder and the network. The function of the NIU is to implement appropriate transport-, network-, and link(including media access) level protocols’ for the satisfactory delivery of compressed video packets from encoder to decoder. For a given link access protocol, support of real-time video sources introduces unique transport-level (i.e., packetization, buffering, flow control, and error control) considerations, which are different from conventional data applications since packets must be delivered to the decoder strictly on a periodic basis. In many systems, the link access protocol must also be tailored to the unique characteristics of video to avoid problems at the transport level. The first step in determining the suitability of a given set of broadband network protocols for the transmission of compressed video is to develop analytical and/or simulation models for performance evaluation. Using appropriate performance evaluation tools, video quality measures such as the packet loss (clipping) probability and/or packet delay can be determined as a function of channel throughput for a given NIU functionality. The eventual goal is to specify network interface protocol parameters which achieve the desired video performance ‘In this paper, consideration will be limited to link- and transport-layer aspects for single-hop broadcast-type fiber and satellite systems. In general (particularly for store-and-forward networks), network-layer protocol issues also need to be addressed.

I 7 ) 1

(Packet Video Transmission System

Packet data Packet Video Periodic video elements (e.g., lines) L

z

~

I

e

Fig. 1. Schematic representation of a packet video transmission system.

criteria with an NIU complexity/cost that is consistent with the application under consideration. In this>paper, we present performance evaluation results for specific broadband packet video system examples to display the typical tradeoffs between video quality measures such as the packet loss (clipping) probability and the network interface protocol algorithms and parameters. The two broadband packet video examples that will be used to illustrate design issues are 1) multipoint-to-multipoint high-quality video services using a 200 Mbit/s implicit token passing (ITP) fiber-optic LAN and 2) pointto-multipoint broadcast quality video distribution using a packet-TDM multiplexed 90 Mbit/s satellite channel. For each of these cases, we examine the link-access-level and transport-level protocol implications, leading to an understanding of typical packet video NIU functions. Overall, our results will demonstrate that when N >> 1 video sources share a common broadband channel, reasonable video packet loss (clipping) rates can be achieved at high channel utilization levels (without adaptive control of the encoding modes required on fixed capacity channels) within the framework of existing data communication protocol services, provided appropriate packetization and buffering parameters are employed at the transport level. 11. SYSTEM MODELS In this section, the components of our performance evaluation model for the two broadband packet video systems under consideration are outlined. First, we describe the video source model, which is common to both system evaluations. A . Video Source Model The results presented in this paper are based on a “broadcast quality” intra-interframe DPCM compressed video source. The encoding algorithm, known as adaptive spatiotemporal (AST) DPCM, has been described by Ng and Hingorani in [ 181 and is representative of the general class of advanced DPCM-based compression techniques. As shown in the AST-DPCM encoder/decoder block diagrams in Fig. 2, the algorithm adaptively switches between temporal and spatial predictors on a pixel-to-pixel

JOSEPH er

U/.:

817

TRANSMISSION SYSTEMS FOR COMPRESSED DIGITAL VIDEO 0 003

Sampled-data

I

I

I

I

I

out

l o l l

R a t e (bits/line) Predictor

1 I d r p

Fig. 3 . Rate distribution (number of bits per line) for the given AST-DPCM video source encoder. Encoder

Predictor

Sampled-data video out Variable rate e n d e d video in

+%qv Predictor

Decoder

Selection

I I

Predictor

Fig. 2 . Block diagram of AST-DPCM encoder and decoder.

basis. This has the effect of using the strongly correlated temporal predictor most of the time, while switching to a moderately effective spatial (intraframe) predictor in areas of high motion. Details on the coding procedure are given in [ 181 and will not be repeated here. For the “broadcast quality” application envisioned here, an appropriate set of encoding parameters has been selected. Images are sampled at 14.3 MHz and processed through appropriate two-dimensional filters to bandlimit the spatial frequencies. The diagonal spatial frequencies are attenuated to permit resampling with a quincunx kernel at an effective rate of 7.1 MHz. The quincunx pattern, specifically the field quincunx, is an offset (diamond-shaped) sampling pattern which provides full horizontal and vertical resolution after appropriate interpolation processing is performed, at the expense of reduced diagonal resolution. The AST-DPCM prediction error signal was quantized with a 171-level uniform quantizer with a step size equal to 3.2 The function Q in Fig. 2 maps the input level to an index j in the range [ - 85, 85 1, while the Q - represents the reverse operation. The quantizer output is then converted to a sequence of nonzero elements and zero run lengths, which are statistically encoded using variable word length Huffman tables generated from appropriate training sequences. The average data rate for the luminance ( Y ) component of typical images at this ‘‘broadcast quality” is approximately 20 Mbit/s. A probability histogram for the number of bits x in an 6 x ) versus x for encoded line (i.e., prob ( x Ix Ix 6x = 4 bits) obtained from simulations (on the David Sar-



+

’Use of a single quantizer results in a fixed level of distortion, yielding a fixed qualityivariable rate encoder.

noff Research Center’s VAX-780-based digital video facility (DVF) 1191) over 30 representative live video segments of 2.675 (80 frames) each is given in Fig. 3. The results in Fig. 3 are for luminance ( Y ) only, but we have observed in our experiments that it is not unreasonable to determine the rate for a color image by appropriate scaling by a constant in the range 1.25-1.5. The line length varies over a wide range, from as few as about 500 bits to as many as 2400 bits with a mean and standard deviation of 1220 and 300 bits, respectively. While the global line length distribution given in Fig. 3 is useful for a general assessment of the capacity needed, it does not contain sufficient information to drive a video source model, which is a nonstationary stochastic process with periodic arrivals. We made some attempts to obtain simple Markovian characterizations from the raw AST-DPCM simulation data as in [ 131, 1201, [2 11, but the results were not encouraging. Accordingly, in order to preserve the nonstationary source characteristics, we chose to drive the broadband channel simulations directly with the raw output data records obtained from the AST simulation sequences. B. Broadband Channel Models 1 ) ITP Fiber-optic LAN: The first broadband channel example is a 200 Mbit/s (multimode) fiber-optic LAN which uses ITP to support shared access multipoint-tomultipoint video communication by a number of stations. This scenario is representative of future applications of broadband LAN’s for high-quality videophone, teleconferencing, or multisource video distribution within a building or a campus. As discussed in the previous section, a packet video NIU serves as the interface between the video codec and the fiber LAN. We consider here an approach by which a “standard” link-level protocol (ITP) is used for packet video traffic, so that the burden of ensuring adequate video quality must be assumed by the transport-layer functions shown in the NIU. a ) Link-Level (Media Access) Protocol: ITP protocols for shared access of a common broadband fiber channel have been widely proposed for multipoint-to-multipoint local area communications. Typical protocols of this class are Expressnet [22], Fasnet 1231, D-net [24], etc. In each case, either a folded bus or a dual-bus topology, as shown in Fig. 4, is used to implement a contention-free

818

IEEE JOURNAL ON SELECTED AREAS I N COMMUNICATIONS, VOL. 7. NO. 5 , J U N E 1989 packet format sensing delay R

Station 111

propagation delay

R

R

Station

Station

1 2

113

Station

p-

transmission cycle

-q

Fig. 5 . Illustration of channel activity on a representative ITP-LAN (Dnet).

existing link access protocol (i.e., ITP) which does not automatically support the special requirements of compressed video, it is clear that certain measures will have to be taken by the video NIU at the transport level. The basic transport-level functions of the NIU (illustrated schematically in Fig. 6) are 1) packetization, 2) variable packet delay compensation, 3) flow control, 4) error control/concealment, and 5 ) packet resequencing (if necesSlation Statlo” Station SIat ion sary). Note that these functions are consistent with the X N 1 2 113 X1 class of services normally associated with transport-level protocols. A brief discussion of the specifics of 1)-5) is given below. i ) Packetization: The transport NIU accepts variable length video units (typically lines or spatial blocks) at a periodic interval (every T seconds). Before the video data enter the ITP-LAN, it is necessary to form transport packets which are suitable for transmission. For the ASTDPCM source described earlier, the natural unit is a line round-robin type of channel access, without the explicit of data. However, since each channel access by a station use of a token. The basic principle is to create a sequence involves a certain amount of overhead in terms of propaof packet “trains” to which each station may append its gation delays, addressing, detection times, and acquisitransmission (if any) when an end-of-train (EOT) condi- tion preambles, the overhead may be reduced by grouping tion is detected on the channel. The details of the ITP K > 1 lines into a single packet for transmission. Another protocols described in [22]-[24] differ mainly in terms of less convenient option is to form packets of a specified the manner in which new trains are generated after each minimum size using the appropriate number of lines retransmission cycle is completed. An example of the re- quired. Fixed length transmission packet formats are not sulting channel activity for Expressnet or D-net, in which desirable in view of the resulting subdivision of some line stations (or a designated station) generate a “locomo- elements. tive” to lead a new train after detecting the EOT, is shown ii) Delay Compensation: At the start of transmisin Fig. 5. sion, the transmit station sends packets over the network The service policy in ITP systems is implicitly round as they are accepted by the link-level protocol. However, robin (unless special measures are taken). We assume un- at the receiver NIU, arriving packets are initially buffered constrained exhaustive service for each station, which for a duration equal to B video unit durations from the means that when a station is given access to the channel, time of the start of transmission (i.e., BT s where T i s the it transmits all packets contained in that transmit buffer. duration of a unit), before being transferred periodically Note that alternative link-level service stategies can also (i.e., every T s ) to the decoder. This has the effect of be considered for the video transmission. One example is creating a total backlog of B units between the encoder a nonexhaustive service policy where only a fixed maxi- and the decoder, resulting in a constant delay of BT s mum number of data packets and/or bits may be trans- (which cannot be too large for perceptual reasons), while mitted by each user on any given access. In general, adap- serving the purpose of absorbing variations in traffic-detive service policies based on observed channel congestion pendent network transit delay. A prolonged burst of and the criticality of each station’s traffic can be devised. congestion (due to high traffic volume at one or more staFor video transmission systems, there may be some ben- tions on the channel) may cause the transmit NIU buffer efit from interaction between the link- and transport-level to fill up, a situation that will result in packet loss (clipprotocols (described below), so that higher link access ping). priority can be given to stations with video buffers nearing iii) Flow Control: A full transmit NIU buffer rethe clipping limit. quires flow control action. conventional data networks b) Transport-Level Protocol: Since we are using an often use input buffer limit (IBL) flow control at the

819

JOSEPH er a l . : TRANSMISSION SYSTEMS FOR COMPRESSED DIGITAL VIDEO P a c k e t Video Transport Level Protocol Video Buffer (B lines)

N e t w o r k I n t e r f a c e U n i t (NIU)

-

Media Aa'ess Level

Physical Level

Packet. izatior

+

out Transmitter

t

Compressed Video Line From encoder Every 63.5 usec

Packet Data

4 to channel

Control Transmit

Functions

Control

Media Amess Level

Physical

Signal Transport Level Protocol Video Buffer

Video out To decoder

(B lines)

-llmcF+F*

-

Packet Data

-

Receive

De-

4

+-

Buffer

in

Receiver

from channel

Reseq.

Receive Functions

Fig, 6 . Functions of the

transport level to deal with similar situations. Typically, in data systems, no additional data are accepted from the source; for video, however, an appropriate strategy may be to discard the oldest packet@) in the transport buffer, while accepting the new video packets which arrive every T s. This is because of the fact that when the buffer fills up, it is already too late for the oldest packet to reach the decoder, whereas the new packet still has a chance of reaching the receive NIU on time. Alternative flow control strategies that are tailored to video traffic may anticipate congestion and discard lines which are old and long, thus attempting to minimize the number of lines clipped. Note that if there were a local interaction between the transport. and presentation layers, this clipping might also be compensated for at the encoder itself to prevent errors from propagating to subsequent frames due to the nature of the coding. Concealment of clipping errors may be further aided by other flow control procedures: as an example, discarding widely spaced lines at the transmit buffer when an overflow condition is anticipated may provide some subjective advantages. iv) Error ControUConcealments: When errors are detected, it seems reasonable to avoid the delay problems associated with retransmission and to treat these conditions in the same way as packet loss conditions due to buffer overflow. Error concealment can be performed either at the transport level or within the video codec. Implementation at the transport level has the advantage of simplifying the video codec to a simple compression-codeto-video-picture-element converter without any special logic. The strategy to be used for error concealment depends strongly on the type of video compression algorithm used. For example, for interframe DPCM with line units (as is the case here), it may be helpful to insert an all-zero unit. On the other hand, for an intraframe block DCT system, interpolation within the video codec may be

packet

NIU

preferable. Note that in the error concealment context, K = 1 lines per packet is a preferable packetization strategy since loss of a packet is then less destructive to image quality. Alternatively, a strategy of using K widely spaced lines within the frame to form a transport-level packet would reduce the subjective impact of packet loss. v) Resequencing: Resequencing of packets at the transport level may be required in certain systems which utilize alternative routes within the LAN. This is not a factor in the ITP-LAN. 2) Packet-TDM Satellite Broadcast: The second example considered is based on a 90 Mbit/s satellite channel in which a central transmit station packet multiplexes several ( N 2-4) AST-DPCM video signals for pointto-multipoint distribution. This scenario, illustrated in Fig. 7, corresponds to an alternative digital technology for direct-broadcast satellite (DBS) distribution of broadcast quality video, which has traditionally been based on analog FM techniques. Typically, the 90 Mbit/s channel speed considered here could be supported with QPSK modulation on a conventional 54 MHz Ku-band (FSS) satellite transponder. By sending compressed digital video instead of analog FM, there is the potential that the capacity (i.e., number of channels per DBS transponder) of such systems could be increased, while simultaneously providing the advantage of the improved link budgets that are associated with digital transmission. Of course, for the DBS application, low-cost receivers for packet demultiplexing and video decoding are critical for economic viability. In this scenario, the transmit (head-end) packet video NIU's serve the link-level function of controlling channel access for several channels ( N > 1 ) of packet video, while providing transport-level buffering against traffic-load-dependent delay variations. At the receive NIU, packets for each video channel being displayed (typically no more than two or three for most applica-

-

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 7. NO. 5, JUNE 1989

820

Channel

#I Encoder Packet-

m Channel Encoder

Vdeo Broadcast Head-end

Broadcast TDM Channel (satellite, fiber, etc.)

NIU

\ ‘c. /

Channel Encoder

Video Decoders

Rcvr.

w

Receive NIU

Fig. 7. Packet-TDM distribution of multiple video channels. Packet format Packet Header

\ time

Video bits (variable)

/ +

Fig. 8. Broadcast TDM signal containing N

tions) must be demultiplexed and buffered to produce the periodic output required by the corresponding video decoder. a) Link-Level (Media Access) Protocol: An example of channel activity for the simplest implementation of a point-to-multipoint TDM broadcast packet video network with round-robin exhaustive service is shown in Fig. 8. Video packets enter the common buffer shown in the order generated and are served by the channel (at a constant bit rate) in an FCFS manner. Although the results in this paper are limited to the simple round-robin case, in principle more complex service policies which jointly optimize the video performance of the N channels supported can be visualized. Note here that in comparison to the multiaccess ITP-LAN case, for the broadcast case there is much greater flexibility in the definition of link access protocols (TDM multiplexing rules). This is because collocation of the video sources at the “head end” permits arbitrarily complex service policies jointly based on the instantaneous requirements of each video source; also, somewhat higher-complexity implementation can be tolerated at the single head-end NIU. b) Transport-Level Protocol: The transport NIU functions (illustrated in Fig. 9) are similar to those for the multipoint-to-multipoint LAN case discussed earlier (i.e., packetization, variable packet delay compensation, flow control, packet resequencing, and error control/concealment). Each function is discussed briefly below.

=

3 packet multiplexed video sources.

i) Packetization: In this case, the overheads associated with each packet are due to addressing and error checking CRC’s. Since these are typically quite small, it does not appear necessary to pack K > 1 lines into a single transmission unit as in the ITP-LAN case. ii) Delay Compensation: As the figure shows, in general, N > 1 channels are encoded and multiplexed at the transmit station (head end) for broadcast transmission. A constant total backlog of B units is maintained in the transmit and receive NIU’s per channel supported. The buffer design, while qualitatively similar to that for the ITP-LAN, probably requires a lower packet loss rate. Also, if the service priority for the video channels can be varied adaptively (as a function of global observation of buffer occupancies, incoming video rates, etc.), it may be possible to acheive a high level of performance without very large buffers. iii) Flow Control: The flow control issues are similar to those discussed for ITP-LAN. In this case, in addition to discarding incoming video units, congestion situations may be handled by adaptively altering the transport priority for the different channels served at the transmit NIU. iv) Error Control/Concealment: The considerations for TDM broadcast are similar to those discussed earlier in the ITP-LAN. v) Resequencing: This is not an issue for the single-channel broadcast application under consideration.

~

JOSEPH er al.: TRANSMISSION SYSTEMS FOR COMPRESSED DIGITAL VIDEO

82 1

modeled by a sequence of AST simulation records, each randomly selected from the available set of 30 records -* B unit from different representative conference video scenes with a duration of 80 video frames. The decision to use 30 PacketTo Flow Channel #2 izalion Channel records was based on experimental evidence related to the Control Encoder a - - + convergence of the global line length histogram (as in Fig. Service Video Buffer 3). Of course, given the inherent unpredictability of video processes, it would be inappropriate to claim that the ensemble of image sequences considered is completely adequate. However, we expect the set used to result in fairly compressed video accurate modeling of the video process, consistent with (variable rate) the storage and processing speed limitations we worked 3-channel Transport Level receive NIU I I under. As a point of reference, we observe that the total of 2400 image frames is a significantly larger database I I l r..... n 13 unit Video Buffer Dethan the 300 frames used for similar purposes in [ 131. 1 ]Packet- I I b) Transmit Process: This process accepts the genization. Channel #2 erated video data lines (for which the starting instants are a error randomized from station to station to model the expected lack of synchronism between stations) and transmits them on the channel using the specified link- and transport-level 1 I functions. Consistent with the protocols described in SecFig. 9. Transmitter and receiver NIU’s for packet-TDM broadcast tion 11, packets are generated by combining data from K lines, resulting in a variable duration transmission on the 111. PERFORMANCE EVALUATION channel. Appropriate overheads that account for the acIn this section, we present numerical results for the quisition bits and the addressing and CRC check bits are multiaccess ITP-LAN and the broadcast TDM packet appended to the packet. Packets are queued at the stations video systems described in the previous section. The re- and enter the channel based on the timing established by sults in this section are obtained from direct-event simu- the channel process. lation of the packet video channel access and transport c) Channel Process: This describes the events takprotocols, using the methodology we have outlined in [25] ing place on the channel and, as alluded to earlier, is login the context of satellite networks. The objective of the ically similar for the ITP-LAN and TDM cases. The packet video system performance evaluation presented round-robin servicing (i.e., access) sequence for the N next is to identify the design tradeoffs between perfor- stations is based on the known physical ordering of stamance measures (such as video clipping probability), ef- tions on the line for the ITP-LAN and is arbitrarily preficiency of channel use (i.e., throughput), and the trans- determined for the TDM case. The record of channel port-level and media-access-level protocol features and events is formed as an appropriate sequence of transmisparameters used. Ideally, we seek a combination of op- sion intervals, propagation intervals, and sensing delays erating parameters that provides reasonable video packet (both propagation and sensing times are zero for the TDM loss rate at high channel throughput without requiring ex- case). The delay between the time at which a packet is cessive buffering and/or complex media access strategies. generated and the time at which it enters the channel can thus be computed. Clipping takes place if this delay exA. Simulation Approach ceeds the total delay variation permitted at the transport For the present study, we have implemented a simula- level. When a packet enters the channel, statistics are coltion model of a generic ITP channel as described in [ 2 6 ] . lected for the overall system and for the particular station. The same simulation program, with a modification that These are used at the end of the simulation run to compute eliminated the propagation and sensing delays between appropriate performance measures on a global and/or stastations, was used to realize a broadcast TDM simulation. tion-by-station basis. The basic principle of the simulation is to operate simuld ) Conjidence Levels: The accuracy of the results taneously N independent video packet generation and from a simulation run are a function of the length of the transmit processes, corresponding to each of the N sta- simulation and the correctness of the source and channel tions on the ITP-LAN or TDM channel, together with an models. With regard to confidence levels, standard techoverall channel observation process that globally moni- niques such as those described in [27] are usually emtors and records the events taking place on the channel. ployed to determine an appropriate simulation run time. These processes are functions of both the link-level and Since the simple methods used for independent experitransport-level protocols defined and are discussed in the ments do not readily apply to network simulations, a frefollowing paragraphs. quently used technique is to identify “regenerative” a) Packet Generation: As mentioned in Section II- states, which separate the run into uncorrelated segments. A, video sources used in the network simulation were An example of a suitable regenerative state is the

-

Channel #I Encoder

0

0

Dnlirv

Y

822

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 7. NO. 5, JUNE 1989

“empty” condition in the simulation of a Markovian queue. Unfortunately, this approach cannot be applied to the nonstationary and nonexponential arrival process associated with compressed video. Because of these difficulties, we adopted an empirical approach to establish the accuracy of the simulation by running the simulation programs for incrementally longer periods of time until the differences in the simulation outputs of interest converged to appropriately small values. In general, we found that by running the simulation in the range of 106-107 packet arrivals per station, adequate accuracy for the average and distribution-“tail”-based performance measures of interest could be obtained. Some examples which illustrate the levels of confidence obtained for key performance metrics are given in the Appendix. 1) ITP Fiber-optic LAN: We consider a 200 Mbit/s ITP-LAN (D-net or Expressnet) with the folded bus structure shown in Fig. 4(a). The total span of the fiber from the first to last station is 5 km, corresponding to a propagation delay of 25 ps. There is an additional 25 ps of propagation delay (t,,) in the fiber between trains of data transmitted due to the time for the train to return to the first station. The sensing overhead ( t d ) is 8 bytes. Other overheads in each packet include 8 bytes of acquisition preamble and 9 bytes for addressing and CRC. In order to determine capacity limits for such a channel, we first assume the buffer size B is large and determine the average packet delay d as a function of the number of stations N . As described earlier, we consider the exhaustive service channel access policy with the constraint that no transmitted packet may be smaller than K 2 1 lines of compressed video. Fig. 10 shows curves of d versus N for K = 1 , 2 , 5 , and 10. While each packet represents a fixed number of compressed lines of video, the actual packet length in bits is variable. An average packet would be K * 1218 bits, where 1218 is the mean of the histogram in Fig. 3. As expected, each of the curves exhibits relatively low delay when lightly loaded, followed by a region in which the delay increases rapidly due to channel congestion. The results show that, as speculated, using K > 1 increases the channel’s capacity by controlling access overheads. However, the use of too large a K is not beneficial since latency delays at low load outweigh the small incremental capacity advantage at heavy load. Overall, we see that with the best choice of K ( = 5 or l o ) , a maximum of about eight or nine stations can be supported, depending on the degree of buffering permissible and the subjective impact of lipp ping.^ This corresponds to a channel usage efficiency (throughput) factor S = (average video rate * N * (one-clip probability )/(channel speed ) in the region of 0.75-0.95. Observe that the throughput (which accounts for both the inherent multiaccess and variable rate multiplexing inefficiences in the system) approaches the ideal value of unity even with the relatively ’A study of the effect of clipping on image quality is beyond the scope of this work. However, AST-DPCM with encoding on a line basis is very amenable to error concealment.

0

0 02

Number of Users

(NI

Fig. 10. Delay performance of the ITP-LAN. 1

Number of lin:s/pkt 1

0

-

1

1

(K) = 5 1

Delay (milliseconds)

Fig. 1 1 . Delay distribution plots as the channel load is vaned.

small number of video sources in this example, thus demonstrating the effectiveness of statistical multiplexing provided by shared packet access. Example probability histograms for delay (i.e., prob ( d Id Id 6 d ) versus d for 6d = 10 ps) corresponding to the average delay results in Fig. 10 are given in Figs. 11 and 12. Fig. 11 shows the effect of increasing the number of stations ( N ) on the ITP-LAN for a constant value of K ( = 5 ) . The results in Fig. 11 show that at N = 7 the delay is essentially bounded to less than 1 ms. However, as N is increased to approach the channel’s “capacity” of N = 9, the delay corresponding to an appropriately defined probability level (such as increases quite sharply. Fig. 12 illustrates the effect of the parameter K at a constant value of N ( = 9). As expected, it is observed that K = 5 provides better peak delay than K = 2 or 3. For specified values of channel loads N, it is possible to determine the clipping probability, which is a measure of received video quality, as a function of the buffer size B . Such curves are shown (with parameter K ) for N = 8 and 9 in Fig. 13(a) and (b), respectively. The results show that, as expected, the clipping probability goes down with both B and K. The curves in Fig. 13 suggest that the use of K > 1 is clearly beneficial in terms of delivered video quality and that K = 5 may be a reasonable choice at both N = 8 and 9. Note that, to provide an image with only a few seconds of clipping per hour, a clipping rate in the region of 10-3-10-4 is required. The curves show that at N = 8 and 9 a clipping probability of imples that B is approximately equal to 200 and 800 lines, respectively, B is about while for a clipping probability of 2 * 350 for N = 8 and greater than 1000 lines for N = 9. These values of B are also consistent with acceptable transfer delay since they are in the region of 0.01-0.1 s.

+

823

JOSEPH er o / . : TRANSMISSION SYSTEMS FOR COMPRESSED DIGITAL VIDEO 1

I 0- 1

1

I

Number of w e n (N) = 9

No ,of Users = 9

.

K ,= 5 10-1

1105

Buffer Occupancy in b i t s

Fig. 14. Distribution of buffer size in bits for varying line storage capacity (B). 100

I

100

Channel Speed

No o f users = 8

10-44 I

.

. . . ..

..I

. .

,

..

. . . I

.

100

10

'

1.--I I 000

Transmit buffer l i m i t (lines)

1000

90 nbps

10000

Transmit buffer l i m i t ( 8 )

Fig. 15. Clipping performance in the packet video broadcast scenario.

(a)

10

100

z

1 00

IO00

Transmit buffer l i m i t (lines)

(b)

Fig, 13. (a) Clipping performance of ITP-LAN with N = 8 users. (b) Clipping performance of ITP-LAN at a load of N = 9.

Also, while the typical required values of buffer B are moderately large in terms of lines, the storage required for compressed lines may not be very great. To estimate the actual buffer storage capacity required, in Fig. 14 we show for K = 5 and N = 9 the buffer occupancy histogram (i.e., prob(b Ib Ib 6b) versus b for 66 = 2 kbits) for several values of maximum buffer size B . Observe that the monotonic decline in the buffer occupancy curves is due to similar properties in the corresponding network by delay distributions. The results show that the 800 lines of storage needed for N = 9 at a clipping probability of lop3 correspond to a buffer requirement of 700 kbits, which appears reasonable by current standards. 2) Packet-TDM Satellite Broadcast: Consider a packet-TDM broadcast scenario example based on a 90 Mbit/s satellite channel. In this case, N video channels share the TDM capacity, with each line transmitted as a separate packet ( K = 1 ) using the format shown in Fig. 8. The packet overhead is assumed to be 9 bytes (for video channel identification and CRC). There are no access overheads in the TDM system.

+

In this case, we first note that, since the average data rate of each compressed video source is the region of 20 Mbits/s, a maximum of N = 4 video channels can be supported. Accordingly, in Fig. 15 we examine the clipping probability versus buffer size characteristics for N = 3 and 4, which correspond to channel utilizations of approximately 0.64 and 0.85, respectively. The results show can that for N = 3 a clipping probability of about be obtained with a buffer size of 1000 lines. When N = 4, the load is high enough to require a transmit buffer of about 10 000 lines in order to achieve the same video quality; however, a clipping probability in the region of lop2can be obtained with 1000-line buffers. The change in buffer size requirements as N goes from 3 to 4 is quite dramatic and is a consequence of the increasing delay variance at channel utilization approaching unity. When compared to the results for the 200 Mbit/s ITP-LAN, we observe that, since the 90 Mbit/s TDM system offers statistical averaging from fewer sources, larger tansport-level buffers are required for equivalent quality and channel utilization. Overall, the results for the packet-TDM case show that high channel efficiency may be associated with rather large buffers (on the order of 2000-3000 lines per video channel) at receive stations. Since in TDM broadcast applications, it is desirable to trade receiver cost against head-end complexity, it may be worthwhile to explore further the adaptive channel access (packet multiplexing) policies alluded to earlier. Suitable access policies combined with appropriate subjectively based flow control methods offer the potential for substantial improvement in video quality versus buffer size characteristics at high channel utilization levels.

~

824

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 7, NO. 5 , JUNE 1989

TABLE I COMPARISON OF SIMULATION RESULTS FOR VARIOUS RUNLENGTHS Run Length (packets/station)

1o6

2 x lo6

5 x lo6

Average delay ( m s )

1.08

1.01

1.11

Throughput

0.856

0.857

0.858 0.197 0.330 0.706 3.72

Percentile of delay (ms)

50 percent 75 percent 90 percent 95 percent

0.193 0.318 0.640 2.86

0.195 0.323 0.677 3.04

Clipping probability with E =

10 100 500 1000

1.97 x l o - ’ 3.62 x lo-’ 7.61 X 8.53 x 1 0 - ~

2.02 3.63 5.31 4.26

IV. CONCLUDING REMARKS In this paper, we have addressed the issue of using conventional shared access packet transport channels for the transmission of compressed digital video. The relationship between the link- and transport-level protocols has been discussed in the context of video transmission, and a generic description for a packet video network interface unit has been presented. Specific attention has been given to identifying appropriate transport-level protocol functions and parameters required for the support of real-time compressed video. Simulationlresults for example fiber LAN and satellite TDM broadcast scenarios were presented to demonstrate the viability of the packet video concepts introduced here. The results show that with appropriate transport-level buffering, reasonable video packet loss rates can be achieved at high channel utilization levels, without the need for buffer controlled adaptive encoding techniques. While our results are somewhat preliminary, it appears that there is the potential that packet transport of compressed digital video may offer a costeffective alternative to conventional analog and circuit assigned digital video transmission systems.

APPENDIX DISCUSSION O F SIMULATION CONFIDENCE LEVELS As stated in Section 111, acceptable confidence levels for performance measures of interest were obtained by running the simulation in the range of 106-107 packet arrivals per station. The actual value required for a particular run increases with channel load, although a minimum of lo6packets was used in all cases. Taking as an example the heavy load ITP-LAN case of N = 9 and K = 5, we illustrate the empirical process of determining the appropriate run length in Table I. The table indicates that a run of lo6 packets per station is adquate for average performance measures such as delay and throughput. For other key measures, such as percentiles of delay and clipping probability, the values become less accurate as they tend towards the tails of the curves from which they are obtained; i.e., the 50 percent delays are more accurate than the 95 percent delay, while the clipping probability for B = 100 is more accurate than for B = 1000, etc. Based on

x lo-’ x lo-’ x IO-) x

2.07 x l o - ’ 4.08 x lo-’ 6.7 x io-’ 2.58 x

observations such as those in Table I, run lengths consisting of 2 X lo6 data packets for each station were used when measuring the clipping probabilities for N = 8 and 9. In view of the decreasing accuracy for rare events, values of clipping probability below lop4are not given in Figs. 13(a) and (b) and 15. Also, to ensure accuracy in Fig. 14 (which goes down to buffer occupancy probabilities in the region of a run length of 10 packets per station was used.



REFERENCES [ l ] J. 0. Limb, C. B. Rubinstein, and J. E. Thompson, “Digital coding of color video signals-A review,” IEEE Trans. Comrnun.,vol. COM25, pp. 1349-1385, NOV. 1977. [2] A. K. Jain, “Image data compression: A review,” Proc. IEEE, vol. 69, pp. 349-389, Mar. 1981. [3] K. A. Prabhu, “Predictor switching scheme for DPCM coding of video signals,” IEEE Trans. Commun.,vol. COM-33, pp. 373-379, Apr. 1985. [4] P. Pinch, “Adaptive intra-interframe DPCM coder,” Eel1 Syst. Tech. J . , vol. 61, no. 5, pp. 747-764, May-June 1982. [5] T. Ishiguro and K. Iinuma, “Television bandwidth compression by motion-compensated interframe coding,” IEEE Commun. M a g . , pp. 24-30, NOV.1982. [6] A. Habibi and P. A. Wintz, “Image coding by linear transformation and block quantization,” IEEE Trans. Commun.,vol. COM-19, pp. 50-62,Feb. 1971. [7] J. A. Roese, W. K. Pratt, and G. S . Robinson, “Interframe cosine transform image coding,” IEEE Trans. Commun.,vol. COM-25, pp. 1329-1339, NOV. 1977. [8] H. Murakami, Y. Hatori, and H. Yamamoto, “Comparison between DPCM and Hadamard transform coding in the composite coding of NTSC color TV signal,” IEEE Trans.Commun., vol. COM-30, pp. 469-479, Mar. 1982. [9] H. Yamamoto, Y. Hatori, and H. Murakami, “30 Mbit/s codec for the NTSC color TV signal using and interfield-intrafield adaptive prediction,” IEEE Trans. Commun.,vol. COM-29, pp. 1859-1867, Dec. 1981. [lo] H. Murakami, S. Matsumoto, Y. Hatori, andH. Yamamoto, “15/30 Mbits/s universal digital TV codec using a median adaptive predictive coding method,” IEEE Trans. Commun.,vol. COM-35, pp. 637645, June 1987. [ l l ] S. Sabri and B. Prasada, “Video conferencing systems,” Proc. IEEE, vol. 73, pp. 671-688, Apr. 1985. [12] J. Zdepski, D. Raychaudhuri, and K. Joseph, “Statistically based buffer feedback policies for constant rate transmission of compressed digital video,” IEEE Trans. Commun.,submitted for publication. [13] B. Maglaris, D. Anastassious, P. Sen, G. Karlsson, and J. Robbins, “Performance analysis of statistical multiplexing in packet video communications,” IEEE Trans. Commun,vol. COM-37, pp. 834844, July 1988.

JOSEPH

CI

ol : TRANSMISSION SYSTEMS FOR COMPRESSED DIGITAL VIDEO

J . P. Vorstermans and A. P. DeVleeschouwer, “Layered ATM systems and architectural concepts for subscribers’ premises networks,” IEEE J . Selecr. Areas Commun., vol. SAC-6. pp. 1545-1555, Dec. 1988. H. Suzuki, T. Takeuchi, F. Akashi. and T. Yamaguchi, “Very high speed and high capacity packet switching for broadband ISDN,” IEEE J . Select. Areas Commun.. vol. SAC-6, pp. 1556-1564, Dec. 1988. N. Ohta. M . Nomura, and T. Fujii, “Variable rate video coding using motion-compensated DCT for asynchronous transfer mode networks,” in Proc. ICC 1988, pp. 38.6.1-38.6.5. W. Verbiest. L. Pinnoo. and B. Voeten. “The imoact of the ATM concept on video coding,” IEEE J . Select. Areas Commun., vol. SAC6 , pp. 1623-1632, Dec. 1988. S . B. Ng and R. Hingorani, “Adaptive spatio-temporal DPCM for videoconferencing,” in Proc. GLOBECOM 1986, Houston, TX, p. 8.7.1-5. C . H . Strolle, T . R. Smith, G . A. Reitmeier, and R. P. Borchardt, “Digital video processing facility with motion sequence capability.”in ICCE Dig. Tech. Papers, June 1985, pp. 178-179. T. Fujii, M. Nomura, and 0. Ohta, “Characterization of variable rate-inter-frame video coding for ATM-based networks,” in Proc. GLOBECOM 1988, pp. 33.1.1-33.1.5. 1 S . 3 . Huang, “Source modelling for packet video,” in Proc. ICC 1988, pp. 38.7.1-38.7.6. 1221 F. A. Tobagi, F. Borgonovo, and L. Fratta, “Expressnet: A high performance integrated services local area network, IEEE J . Select. Areas Commun., vol. SAC-I, pp. 898-913, Nov. 1983. [23] J. 0. Limb and L. E. Flamm, “A distributed local area network packet protocol of combined voice and data transmission,” IEEE J . Select. Areas Commun., pp. 926-934, Nov. 1983. [24] C.-W. Tseng and B.-U. Chen, “D-Net, a new scheme for high data rate optical local area networks,” IEEE J . Select. Areas Commun., vol. SAC-I, pp. 898-913, Apr. 1983. [25] K. Joseph and D. Raychaudhuri, “Simulation models for performance evaluation of satellite multiple access protocols,” IEEE J . Select. Areas Commun., vol. SAC-6, pp. 210-222, Jan. 1988. 1261 P. Rodrigues, L. Fratta, and M. Gerla, “Tokenless protocols for fiber optic local area networks.” IEEEJ. Select. Areas Cornmun, vol. SAC3, pp. 928-940, Nov. 1985. [27] J . F. Kurose and H. T. Mouftah, “Computer-aided modeling analysis and design of communication networks,” IEEEJ. Selecr. Areas Cnmmun.. vol. SAC-6, pp. 130-145, Jan. 1988. ”

Kuriacose Joseph (SM’82-M’85) received the B.Tech degree in electrical engineering (electronics) from the Indian Institute of Technology, Madras, India, in 1980 and the M.S and Ph D. degrees in electrical engineering from the State Univesity of New York, Stony Brook, in 1981 and 1985, respectively Since 1985 he has been working a4 a Member of the Technical Staff in the Communications Research Laboratory of the David Sarnoff Re\earch Center, Princeton, NJ. He has been involved in

825

analytical research on satellite communication systems, primarily in the design and evaluation of satellite multiple access systems. He is currently working on video coding algorithms for teleconferencing applications. His research interests include video coding and transmission, packet switching networks for data and video, and satellite multiple access and mobile communications systems. Dr. Joseph was the recipient of the RCA Laboratories Outstanding Achievement award in 1986.

Dipankar Raychaudhuri (S’78-M’79-SM’87) received the B. Tech degree (honors) in electronics and electncal communication engineenng from the Indian Institute of Technology, Kharagpur, India, in 1976 and the M S and Ph D degrees in electncal engineering from the State University of New York, Stony Brook, in 1978 and 1979, respectively. Since January 1979 he has been with the Communications Research Laboratory at the David Sarnoff Research Center (formerly RCA Laboratories), Princeton, NJ, where he is currently a Senior Member of the Technical Staff He has been involved in analytical research on various aspects of satellite and computer communications and, most recently, was responsible for a research program on the design of very small aperture terminal (VSAT) data networks His research interests are in the areas of packet networks, satellite communications, multiple access, and performance evaluation, and he has authored six U.S patents and approximately 40 journal and conference papers in these areas Dr Raychaudhuri is a member of Sigma Xi He has been active in IEEE Communications Society conferences and committees and is currently serving as an Editor for the IEEE COMMUNICATIONS MAGAZINE and as ViceChair of the Data Communication Systems Committee. He was a recipient of RCA Laboratories Outstanding Achievement Awards in 1981, 1984, and 1986

Joel Zdepski (M’89) received the B S E E. and M.S E E. degrees from Rutgers University, New Brunswick, NJ, in 1981 and 1986, respectively After 1981 he was employed by Dranetz Technologies Inc., where he designed data acquisition equipment. In 1984 he returned to Rutgers University to pursue graduate studies full time In 1986 he joined the Video Communications Research group at the David Sarnoff Research Center, Princeton, NI, as a Member of the Technical Staff His activities include compression algorithm design for image data compression, robust video formats for satellite transmission, and video communications systems design He is presently pursuing the Ph D degree, also at Rutgers. Mr. Zdepski is a member of Tau Beta Pi and Eta Kappa Nu.

Lihat lebih banyak...

Comentários

Copyright © 2017 DADOSPDF Inc.