NI Broadcasting (MPEG-2, CD) Email, file transfer. RealAudio .... Sender. OC-3 (155 Mbps). Figure 3. Network topology. multimedia streams. The motion JPEG clips, ... trolled cross traffic consists of batches of cells injected into the network at ...
On QOS Mapping in Multimedia Networks Jean-Franc¸ois Huard and Aurel A. Lazar Department of Electrical Engineering and Center for Telecommunications Research Columbia University, New York, NY, 10027-6699 http://comet.ctr.columbia.edu/˜fjfhuard, aurelg Abstract A framework for studying QOS mapping between various levels in the information transport protocol stack and a platform for evaluating the mapping rules by performing concurrent network and application level measurements of QOS are presented. An empirical loss mapping rule between the application (frame) level and network (cell) level is given. The rule was tested using two 10 minutes motion JPEG video clips and a 30 minutes MPEG-2 VBR video clip under 55 different network load conditions and gives results within a 3% error margin.
1. Introduction The ability to satisfy quality of service (QOS) is one of the key requirements of ATM networks. The task of guaranteeing end-to-end QOS requires the understanding of the various levels of QOS of the protocol stack. For example, switch vendors are typically concerned only with per-hop, per-class cell-level QOS while applications with multimedia streams are only concerned with end-to-end, per connection frame-level QOS. It is the end-to-end QOS, rather than the per-hop QOS, that is perceived and important to the users. Furthermore, network operators use the latter to guarantee QOS to their users and need to translate user QOS requirements into network QOS requirements. Each of these QOS terminologies deal with different statistical quantities that have to be mapped into each other. The objective of this paper is to lay down a framework for QOS mapping and to provide mapping rules whenever possible. In terms of mapping rules, this paper focuses on application-to-network loss mapping. The data unit considered on the application level is the frame (video frame or audio packet) while the ATM cell is used on the network level. In short, the focus is on understanding the relation between frame loss and cell loss. Application-to-network
QOS mapping is needed to reserve the appropriate amount of network resources at connection establishment time. Furthermore, good mapping rules are essential for avoiding the reservation of too much (or too little) resources. An empirical loss mapping rule between the application (frame) level and network (cell) level is given. The rule was tested using two 10 minutes motion JPEG video clips and a 30 minutes MPEG-2 VBR video clip under 55 different network load conditions and gives results within a 3% error margin. The paper is organized as follow: Section 2 introduces the QOS mapping framework and Section 3 presents the QOS mapping rules. In Section 4, a measurement system to validate the mapping rules is presented. The system enable the measurement QOS concurrently at the network and application level. In Section 5, measurement results that validate our mapping rule is given. Finally, some related work is reviewed in Section 6.
2. QOS Framework 2.1. End-to-End QOS Specification A QOS specification is an abstraction that is used for QOS mapping. It is associated with a media stream and is closely related to the concept of a connection and of its endpoints. A media stream is the information flow that is transmitted from the source endpoint to the destination endpoint of a connection1. For example, a QOS specification at the application level permits the selection of an appropriate transport protocol and when mapped to the network level, it permits the reservation of the appropriate network resources. In this section, we give our definition of a QOS specification and we specialize it to each level of the protocol stack. 1 We
assume that each stream is composed of a source and a destination endpoint. Multicast (point-to-multipoint) streams are also considered; however, for QOS mapping, it is sufficient to consider each source and destination pair separately.
2.2. Per-Level QOS Specification Quality of service is defined at various levels of the information transport protocol stack. This work mainly focuses on two levels with the associated protocol data units (PDUs): APDUs at the application level (e.g., video frames, audio packets, data buffers) and NPDUs at the network level (e.g., ATM cells and IP datagrams). Transport PDUs (TPDUs) are used internally for fragmentation by the transport protocols (e.g., TCP segments and AAL5 frames). Transport QOS issues are not directly addressed in this paper as 2 The
PDU gap loss is the number of consecutively lost PDUs [2].
Time Delay
Time Delay Distribution
Class II ε% Blocking Clipping S II
Time Delay
Gap Distribution
Time Delay Distribution
Class I
SI
Average Throughput
A QOS specification consists of three parts: a service descriptor, a traffic description and a QOS profile. The service descriptor is an attribute such as voice telephony, audio, video conference, etc. The traffic description is associated with the source as it specifies the media stream statistics and it is used by the flow control subsystem at the source to regulate the injection of traffic into the network. The QOS profile is associated with the destination as it specifies the guaranteed quality of service that a stream conforming to its traffic description should receive. The QOS profile often reflects the interactivity characteristics of a media stream through its delay and loss constraints. Finally, when compared with QOS measurements, the QOS profile gives an indication to whether the QOS guarantees are provided. Two descriptors are used for characterizing the traffic: the maximum protocol data unit size (e.g., maximum video frame size, audio packet size, AAL5 frame size, ATM cell size) and the maximum PDU rate (e.g., video frame rate, audio packet rate, peak cell rate); when combined, the descriptors give the peak rate of the media stream. The parameters used in the QOS profile are the maximum PDU delay, the maximum PDU loss rate and the average PDU gap loss2 . The QOS specification is end-to-end; i.e., for any level in the protocol stack, the QOS is specified (and measured) from the moment an level L protocol data unit (PDU) crosses the boundary from level L to L-1 at the source endpoint to the moment it crosses the boundary from level L-1 to L at its destination endpoint. In particular, the end-to-end network QOS will be given in terms of cell level statistics in between the network adapter of the source endpoint to the network adapter of the destination endpoint. It assumes that the QOS routing subsystem is responsible for finding a route with the required end-to-end cell level QOS for a given traffic flow. It is the responsibility of the routing subsystem to combine the per-hop QOS provided by each switch and to compute the overall end-to-end network QOS guarantees. In summary, a QOS specification is a triple (S; T; P ), where S is a service descriptor, T is a traffic description and P is a QOS profile. In the next section QOS specification examples are given for each level of the protocol stack.
Gap Loss η
Class III Γ Average Time Delay T
Figure 1. A typical network QOS profile. they are mainly relevant to the internals of a specific protocol. A transport protocol implementation is expected to deliver to the application QOS specified at the APDU level, and it is expected to receive QOS guarantees from the network provider at the NPDU level. ATM Network QOS Specification. It is assumed that future multimedia networks will have de facto standard services with predefined traffic descriptions. It is easy to envision voice telephony at 8kbps and 64kbps, talking-head video at 128kbps, compressed CD quality audio at 256kbps and uncompressed at 1.4Mbps, MPEG-2 video at 1.5, 4.5 and 13.5Mbps, etc. The assumptions are based on a bulk pricing model that determines a large proportion of the traffic to fall into these categories, thereby, allowing schedulers at all contention points to take advantage of statistical multiplexing. Custom services with non standard rates will also be supported using peak rate resource allocation, and thus will be higher priced. It is assumed that there is a common predefined set of QOS profiles throughout the network. In this paper, we propose three distinct types of QOS profile (see Figure 1) derived from the traffic classes defined in [8]. Class I QOS profile is characterized by zero cell loss probability and a bound S I on the maximum cell delay. Class II QOS profile is defined by % cell loss probability, average cell gap loss and a bound S II on the maximum delay. Moreover, cells belonging to class I and class II QOS profiles that are delayed more than their respective delay bounds are considered lost. Finally the class III QOS profile is chararterized by a bound T on the average time delay and a lower
bound ? on the average throughput. No restrictions are placed on traffic streams in the class III QOS profile and no loss bounds are guaranteed. End-to-end retransmissions are needed to recover form cell losses of traffic streams with class III QOS profile. Transport QOS Specification. The next generation of transport protocols will need to offer services to multimedia streams with two basic types of timing requirements: real-time (non-real-time) and interactive (non-interactive). Other requirements, such as reliability (error resilience, sequencing, etc.), throughput, efficiency, etc., are well-known and understood. These are not addressed here. Below, the focus is on the timing requirements that are imposed on transport protocols with the emergence and deployment of new multimedia applications. A flow is said to be a real-time (RT) stream if it is associated with an isochronous data flow. A RT stream, therefore, requires continuous data delivery to the user with tight deadlines (i.e., data that arrives after its deadline is considered lost). Applications using RT streams are often associated with multimedia devices (such as audio and/or video). Isochronous data streams exhibits stringent time delay bounds due to their playout characteristics. These streams can be played upon reception, with potentially some playback delay (for error resilience or controlling the delay), or stored and played back at a later time. Conferencing and broadcasting are examples of applications requiring RT streams such as MPEG-2 and motion JPEG for video, and CD and telephony for audio. RT streams may or may not allow for retransmissions, depending on whether they are interactive. Interactive (I) streams are streams for immediate usage, with almost no playback delay (or relatively short compared to the end-to-end delay). They require very low response time, but not necessarily high bandwidth and isochronicity. Human response time and tolerance provide bounds on delay latency and losses for such streams. For example, the end-to-end delay for audio without echo cancellation must be less than 40 msec, for audio with echo cancellation less than 150 ms (to have an acceptable response to hand-off) and less than 250 msec for video. For a VCRtype control stream a latency of about 500 ms is considered acceptable. In short, the criterion for classifying a stream in the interactive category is its response time. The playback delay of interactive streams needs to be minimized and within human levels of tolerance at all time. Furthermore, audio and video may have to be synchronized if they are not multiplexed in the transport layer. Non-interactive streams can tolerate more network fluctuations since local resources (RAM and disk space) may be used to compensate for the less stringent QOS requirements imposed on the network. Table 1 shows some application and/or media stream examples and their classification.
I NI
RT Conferencing (motion JPEG, MPEG-2, telephony) Shared whiteboard Broadcasting (MPEG-2, CD) RealAudio, RealVideo
NRT Remote control Web browsing Telnet Email, file transfer Distributed computing
Table 1. Application categories. Application QOS Specification. The types of applications considered here are multimedia-based. For example, video conferencing, audio broadcasting, web browsing, etc. For these services, it is quite easy to give the traffic description using the peak APDU rate (e.g., 30 frames per second for video) and the maximum APDU size (e.g., maximum MPEG-2 video frame size; selected based on the Q factor). See [3] as an example of the later. The parameters of the transport QOS profile depend on the isochronicity, latency, error resilience and reliability characteristic of the media stream. Typically, the gap loss is desired to be set to 1; i.e., no consecutive APDU loss. The delay and loss depend on the application timing requirements. For example, non-interactive applications may be able to support loss, at the expense of occasional retransmissions and buffering. User QOS Specification. The QOS at the lower levels is specified in terms of PDU loss, gap loss and delay. For a typical user, this kind of description is too sophisticated. For that reason, the user level QOS is usually described via a subjective specification. For example, terms such as excellent, good, fair, bad, etc., are used to specify QOS while adjective small, medium, large are used to specify window size preferences. The methodology for the subjective assessment of QOS is not address in this paper. An example of such methodology can be found in [6].
3. QOS Mapping The process of translating QOS specifications between two different levels of the protocol stack is called QOS mapping. For example, user-to-application mapping is needed to ease the process of selecting QOS at the human-machine interface. It is a mapping from a set of user preferences to a quantitative description of the service desired. Given that a quantitative description is available, application-to-network mapping is needed to reserve the appropriate network resources. QOS mappings are needed at both, connection establishment time and at the renegotiation time. The issue of finding a route providing the required end-to-end cell level QOS is not addressed here. It is assumed that it is the responsibility of the QOS routing system to find one. (The
Application
Transport
Network
Conferencing RT_I MPEG-2 qStack motion JPEG telephony NRT_I kStack Broadcasting CD audio MPEG-2
ATM (cells)
Mobile
RTP RT_NI
Sharedboard
UDP
IP
TCP
(datagrams)
NRT_NI
File transfer
Figure 2. Service mapping. aggregation of the per-hop, network level QOS on a route is also considered part of the QOS routing system.) Finally, the application-to-transport mapping is needed for monitoring and adapting to the rapid network fluctuations of QOS. The overall QOS mapping process is proposed to be carried out in two steps: a mapping between the services and traffic descriptors and a rescaling of the parameters of the QOS profile. In the next sections, only the application-totransport service mapping and the application-to-network parameter mapping will be addressed. Furthermore, the focus will be on loss.
3.1. QOS Service Mapping Figure 2 illustrates a possible service mapping structure based on the transport services proposed in Section 2.2. The first three classes (RT I, NRT I and RT NI) were exemplified earlier. The last class, NRT NI, provides a full spectrum of QOS for applications without timing requirements. It is expected to be used for backward compatibility (i.e., with current IP application that use UDP or TCP); that is, from a potentially lossy service (UDP/IP) to a lossless service (TCP/IP) with no delay constraints.
3.2. QOS Parameter Mapping Let a and n denote respectively the application and network levels. To write parameter mapping rules, the following symbols are needed:
R(l): S (l): A(l): D(l): L(l): G(l):
PDU max rate PDU max size PDU ave. size PDU (end-to-end) delay PDU loss ratio PDU ave. gap loss
(# pdu/sec), (bytes/pdu), (bytes/pdu), (sec), (number), (# pdu),
encoding frame rate (sec?1 ) number of frames window size (pixels) ave. frame size (bytes) min. frame size (bytes) max. frame size (bytes) std dev. of frame size
medium m JPEG 25 15 000 240 320 6394 2052 13 364 2132
large m JPEG 20 15 000 480 640 14970 7131 45 234 4936
MPEG-2 MPEG-2 30 54 020 296 720 2573 322 16 219 2238
Table 2. Video clips statistics. where l is either a or n. Finally, the peak rate at level l by PR(l) = S (l) R(l) (bytes/sec) and the average rate AR(l) = A(l) R(l) (bytes/sec). For losses, the following relation is proposed:
L(n) 1 L(a) A(a) G(a) = A(n) G(n) AGP (n; a) (1) where AGP (n; a) is the average number of NPDU gaps per
APDU. For example, if a frame loss of 10?3 is desired, considering an average gap loss of 1, average frame size of 2400 bytes, an average cell gap loss of 1.75 and average number of cell gaps per frame of 4, then the cell loss should be 1:4 10?4. To evaluate the accuracy of the loss mapping rule prescribed by Equation 1, experimental loss measurements have been concurrently performed at the application and network level. The results are reported in Section 5.
4. A QOS Monitoring Platform A measurement system has been developed to measure QOS concurrently at the application level and the network level. The system was implemented in the firmware of an HP broadband network analyzer [4] and permits real-time measurements of loss and delay. Loss measurements in the order of 10?9 and delay measurements with a precision in the microsecond range can be performed.
Experimental setup Two modules of the broadband analyzer were used (see Figure 3); one as a source and the other as a sink. The source transmits dummy video frames of the size specified by the video traces described below. Each cell transmitted is timestamped, numbered and contains the sequence number of the last cell of the frame it belongs to. The sink receives the cells and evaluates the cell delay, frame gap loss, cell gap loss and gap per frame histograms. Two motion JPEG and one MPEG-2 VBR video clips were used as
MPEG-2 VBR
HP Broadband Network Analyzer Sender
Poisson Constant Controlled 20 Mbps load (CBR) cross traffic
1.05
Receiver
"fg.txt" 1
0.95
0.9
Fore ASX100
ATML Virata
Scorpio Stinger
Taxi (100 Mbps) OC-3 (155 Mbps) Contention Point
NEC Fore model 5 ASX100 Constant Poisson load 20 Mbps garbage sink
0.85
0.8 0
5
10
15
20
25
30
35
40
45
50
Figure 3. Network topology. Figure 4. Results for the MPEG-2 video clip. multimedia streams. The motion JPEG clips, composed of 15 000 frames, were recorded with different window sizes: medium (240320 pixels) and large (480640 pixels). The MPEG-2 video clip has 54 020 frames (approximately 30 minutes at 30 frames per second) with a window size of 296720 pixels and GOP of 12/3. A summary of the statistics of the video clips is given in Table 2. The network topology and the interference traffic streams are illustrated in Figure 3. The video stream is generated by the module on the left of the broadband analyzer. At the first hop (the Fore ASX-100 switch), the video stream is multiplexed with a Poisson cell stream with an average of 20 Mbps. The Poisson stream is used to add some cell delay variation and to interleave the video cells so that they are not all consecutive. The combined flow (video cells and Poisson cross traffic) goes through a set of four ATM switches and OC-3 links. The contention point is located in the middle, at the Scorpio switch. Two cross traffic streams are injected: a constant bit rate (CBR) stream and a controlled bulk arrival cross traffic stream. The three streams (combined flow, CBR and controlled bulk arrival) have the same priority and compete for the available buffer. The CBR stream is used to help filling up the queue at the contention point so that buffer overflow can occur more easily (77.5 Mbps CBR is injected when the medium window motion JPEG clip or the MPEG2 VBR clip are played, and 74 Mbps CBR is injected when the large window motion JPEG clip is played). The controlled cross traffic consists of batches of cells injected into the network at line speed (155.52 Mbps). The batch arrival process is Poisson with an average arrival rate that can be set from 0.016 to 100 (i.e., mean batch interarrival time from 10 msec to 60 sec). The batch sizes are geometrically distributed with an average that can be set anywhere from 10 to 50 000 cells. The batch arrival rate and the average batch size can be changed dynamically to obtained different cross traffic load scenarios.
motion JPEG, large window 1.05 "cl15k-l.txt" 1
0.95
0.9
0.85
0.8 0
0.5
1
1.5
2
2.5
3
3.5
4
Figure 5. Results for the large mJPEG clip. Finally, at the last hop, the stream is demultiplexed. The dummy video cells are sent back to the broadband analyzer for real-time measurements of QOS and the Poisson cross traffic is sent into the “garbage sink.”
5. Experimental Results Figure 4 experimentaly validates Equation 1. In order to obtain the graph of the figure, all terms of the equation were brought on the same side, thus, the expected resulting value should be 1. As one can see, the average is 1.0067 while all the points are withing 3% of error. In fact, 43 out of 50 points have the value 1.01. Similar measurements are in progress for the motion JPEG clips. Figure 5 shows the preliminary results obtained; the values are in the range from 0.93 to 0.97.
6 Related Work
Acknowledgements
In [11], a general framework for QOS management from user-to-user is presented. Simple mapping rules considering the PDU size, PDU rate and PDU delay are given. A QOS management system that performs QOS monitoring is also described. The one-to-one translation approach described in [9] and [10] is comparable to our approach. It considers mapping between application and network level QOS and formulates arithmetic rules. The main difference resides in the choice of QOS parameters and the rules proposed. In [9], a model for an endpoint entity called QoS Broker is presented. In [10] the model is further developed to coordinate the endsystem resource management tasks: QOS mapping, admission control and task scheduling. Similar functionality is provided in the xbind broadband kernel [1] and [7], but not addressed in this paper. The QOS broker of [9] is comparable to the QOS manager referred to in [11]. In [5], end-to-end mapping rules have been investigated. However, no exact loss mapping rule was given. Also, the monitoring setup was different. The source and sink were two workstations using a native ATM protocol stack and real-time video clips were played. However, it was not possible to measure the average number of cell gaps per frame as the driver of network card adapter did not provide any interface to recover an AAL5 frame containing errors. Furthermore, the cell gap and cell loss were estimated using probing.
The work reported here was supported in part by Hewlett-Packard Company (IDACOM Telecommunications Operation). The authors would like to thank Brian Smith for his encouragement throught the course of this work.
7. Conclusion In this paper, a framework for studying QOS mapping rules was presented. As part of this research, a platform for evaluating end-to-end QOS by performing concurrent network and application level measurements was developped. The early set of concurrent QOS measurements have shown that the loss mapping rule given by Equation 1 holds. In order to empirically validate our QOS mapping rule, more data is being collected for future analysis. Furthermore, various network topologies and cross traffic patterns will also be tested. Finally, in order to test the sensitivity to the end-system behavior, measurements using various implementations of transport protocol stacks will be carried out.
References [1] M. C. Chan, J.-F. Huard, A. A. Lazar, and K.-S. Lim. On realizing a broadband kernel for multimedia networks. In Proceedings of the Third COST 237 International Workshop on Multimedia Telecommunications and Applications, Barcelona, Spain, Nov. 1996. [2] J. M. Ferrandiz and A. A. Lazar. Monitoring the packet gap of real-time packet traffic. Queueing Systems, 12:231–242, 1992. [3] K. Fukada, N. Wakamiya, M. Murata, and H. Miyahara. QOS mapping between user’s preference and bandwidth control for video transport. In Poceedings of the Fifth IFIP Workshop on Quality of Service (IWQOS’97), pages 291– 302, New York, NY, May 1997. [4] Hewlett-Packard. The Broadband Series Test System (BSTS). URL: http://www.tmo.hp.com/tmo/pia/bsts/. [5] J.-F. Huard and A. A. Lazar. On end-to-end QOS mapping. In Poceedings of the Fifth IFIP Workshop on Quality of Service (IWQOS’97), pages 303–314, New York, NY, May 1997. [6] ITU-R BT.500-7. Methodology for the subjective assessment of the quality of television pictures. ITU-R Recommendations, BT Series Fascicle, Geneva, 1995. [7] A. A. Lazar, K.-S. Lim, and F. Marconcini. Realizing a foundation for ATM interoperability with the binding architecture. IEEE J. Select. Areas Commun., 14(7):1214–1227, Sept. 1996. [8] A. A. Lazar and G. Pacifici. Control of resources in broadband networks with quality of service guarantees. IEEE Communications Mag., 29(10):66–73, Oct. 1991. [9] K. Nahrstedt and J. M. Smith. The qos broker. IEEE Multimedia, 2(1):53–67, Spring 1995. [10] K. Nahrstedt and J. M. Smith. Design, implementation and experiences with the omega end-point architecture. IEEE J. Select. Areas Commun., 14(7):1263–1279, Sept. 1996. [11] M. Zitterbart. User-to-user qos—management and monitoring. In Fifth IFIP Int’l Workshop on Protocols for High-Speed Networks (PfHSN ’96), pages 219–234, SofiaAntipolis, France, Oct. 1996.
