Multimedia networking abstractions with quality of service guarantees

1 Multimedia Networking Abstractions with Quality of Service Guarantees Aurel A. Lazar†, Lek Heng Ngoh* and Anupam Sahai* Abstract The concept of schedulable region previously introduced to broadband networks with quality of service guarantees is extended to multimedia devices such as audio/video processing and disk storage units. The resulting multimedia capacity region characterizes the amount of resources a physical device is able to provide under quality of service constraints. The modeling methodology supports a straightforward association of resources with logical objects and, thereby, the mapping of logical objects onto physical objects with quality of service guarantees. Examples showing the size and shape of the multimedia capacity region of various physical devices are given. 1. Introduction Multimedia networks are enabled by two basic technologies: networking and multimedia computing. Object-oriented languages provide the main high level modeling tool for both. Mapping objects onto physical resources with guaranteed quality of service (QOS) is a key requirement. For broadband networks the object-oriented methodology has been instrumental in modeling the network entities and in describing the mechanisms composing the network architecture. A general model for characterizing the capacity of key network abstractions with QOS guarantees was put forth in [14]. Within this model, the mapping of logical resources onto physical resources is a straightforward task. In order to characterize the capacity region of a multiplexer in broadband networks with QOS guarantees, Hyman et al. [8] have introduced the concept of schedulable region. The schedulable region represents the multidimensional capacity of the multiplexer; its dimensionality depends on the number of traffic classes and represents the stability region of the multiplexer under QOS constraints. The schedulable region is a resource abstraction that allows a separation of time scales: †. Department of Electrical Engineering and Center for Telecommunications Research, Columbia University, New York, NY 10027-6699, e-mail: aurel@ctr.columbia.edu. *. Institute of Systems Science, National University of Singapore, Singapore, 0511. September 28, 1994 2 the time scale of cells and the time scale of call arrivals and departures. In [9] it is shown how the separation of time scales is the appropriate tool for resolving the admission control problem. Based on a calculus of schedulable regions, the QOS in the network can be guaranteed [10]. Can the concept of schedulable region be extended to multimedia computing [11]? Is such an extension needed? Multimedia computing platforms, such as those implemented based upon the Multimedia Systems Services Architecture [7] (see also [6], [17]), provide a framework of middleware, e.g., system components lying in the region between the generic operating system and specific applications. This middleware is based upon high level operating systems abstractions. These abstractions, however, exhibit only a nominal modeling of QOS constraints. For example, most existing real-time operating systems do not support abstractions that provide real-time QOS guarantees for multimedia traffic. At most they support a fixed priority [16], or a dynamic priority based scheduling scheme [4]. To realize QOS guarantees, the mapping of task requirements onto a set of priorities becomes fairly complex and low level. Higher level abstractions are needed on which admission control decisions can be made. The situation for video on demand (VOD) systems is very similar [18], [19], [20]. While scheduling and admission control algorithms have been proposed in the literature, appropriate abstractions are not available. How to map logical objects onto physical resources with guaranteed quality of service has not been specified. In this paper we extend the methodology for characterizing resources based on the schedulable region concept to multimedia devices. Here the novel concepts of multimedia capacity region, an extension of the schedulable region to different media devices with QOS guarantees, will be introduced. Using the multimedia capacity region, the problem of scheduling multimedia becomes identical to the one in [9], i.e., a real-time packing exercise very similar to the Tetris game. User requirements are modeled through service class specifications with QOS constraints. The service class specification is in terms of JPEG, MPEG-I, MPEG-II video and CD quality audio streams with QOS constraints. We believe that these classes will be widely available for interactive multimedia applications. QOS for these classes is specified by a set of frame delay and loss constraints. Please refer to [1] for an extensive discussion on QOS specifications. We model the various resource components of the workstation using logical abstractions (objects) with associated multimedia capacity regions. By using the multimedia capacity region to quantitatively model resources, the mapping the logical resources onto physical resources becomes straightforward. The operating system details are internalized, and a uniform logical abstraction is presented to the middleware designer. How the multimedia networking abstractions with QOS guarantees (the schedulable and the multimedia capacity regions) can be employed by the middleware designer is presented in [12], [13]. September 28, 1994 3 This paper is organized as follows. In section 2 the modeling of broadband networking abstractions is briefly reviewed. In section 3 we describe in detail the abstractions for modelling of the multimedia components, like the audio/video unit and the disk storage unit. In section 4 the multimedia capacity of these components are established experimentally. Section 5 concludes the paper. 2. Modeling Broadband Networking Abstractions with QOS Guarantees In this section the characterization of physical resources in a broadband network under cell level QOS constraints is briefly discussed. The modeling of a (non-blocking) broadband switch with QOS guarantees is reviewed in section 2.1. Its main building block is a multiplexer. In section 2.2 the schedulable region as the resource abstraction of a broadband multiplexer with cell level QOS is briefly presented. 2.1 Modeling a Broadband Switch A broadband switch accepts traffic from a set of input links, and routes each incoming cell though a non-blocking switch fabric to the appropriate output port, where it is queued in a link control unit, for transmission over the output link. This link control unit is essentially a multiplexer. It consists of a set of buffers, a buffer manager, and a scheduler, and it mediates the contention among cells from different input links and among those of different traffic classes. A traffic class is characterized by its statistical properties and QOS requirements. Data granularity is given by the size of the packets or cells (53 bytes). Typically, the QOS requirements reflect cell loss and delay constraints. In order to efficiently satisfy the QOS requirements on the cell level, scheduling and buffer management algorithms dynamically allocate communication bandwidth and buffer space. The system architecture of a non-blocking broadband switch is shown in Figure 1. The performance of the switch is largely determined by its main building block, the multiplexer. In the next section we will characterize the capacity of a broadband multiplexer by making use of the concept of schedulable region. 2.2 The Schedulable Region of a Multiplexer The capacity of the link, the size of the buffer, and the scheduling and buffer management algorithms used will determine how many calls of a given class the multiplexer will be able to support [8], while guaranteeing cell level QOS to each class. The set of points in the space of possible calls for which QOS can be guaranteed on the cell level is called the schedulable region. As such, it represents a stability region. Its dimensionality is given by the number of traffic classes. September 28, 1994 4 The schedulable region of a multiplexer with a 100Mb/s link capacity is shown in Figure 2. On each of the three axes the number of calls of a particular traffic class is shown. The three traffic classes in Figure 2 correspond to video, voice and data streams. The specifics are: Class I traffic is characterized by a frame duration of 62.5 ms and a peak rate of 10Mb/s, Class II traffic is modeled as an on-off source with constant arrivals with an exponentially distributed active period and 64K/s peak rate, and Class III traffic is mod- Input Link Four-Port RAM I.1 II.1 III.1 IV.1 I.2 II.2 III.2 IV.2 I.3 II.3 III.3 IV.3 Output Link Buffer Manager and Link Scheduler Input Link Four-Port RAM I.1 II.1 III.1 IV.1 I.2 II.2 III.2 IV.2 I.3 II.3 III.3 IV.3 Output Link Buffer Manager and Link Scheduler Input Link Four-Port RAM I.1 II.1 III.1 IV.1 I.2 II.2 III.2 IV.2 I.3 II.3 III.3 IV.3 Output Link Buffer Manager and Link Scheduler Figure 1. The System Architecture of a Broadband Switch eled as a Poisson source with 1Mb/s average rate. Other details including the exact QOS parameters are given in [8]. The surface depicted in Figure 2 delimits the capacity region of the multiplexer. Any combination in the number of calls below this surface has its QOS guaranteed. September 28, 1994 5 From the point of admission control, the schedulable region is a sufficient representation of the multiplexer and summarizes the net effect of all cell-level details including the scheduling and buffer management algorithms, the traffic statistics and the QOS parameters. Note that while the schedulable region is depicted as a three dimensional volume, it is in fact an n-dimensional space, where n is the number of traffic classes recognized by the admission control entities. 5 10 Class I Ca lls 15 60 0 40 Cla ss III 0 20 0 Class II Calls 4000 2000 A pair of objects modeling a multiplexer and the associated schedulable region represents a complete characterization of the physical multiplexer from the QOS point of view. This resolves the mapping question between logical resources and the corresponding physical resources with QOS guarantees. Thus, the schedulable region conceptualizes the difference between data transmission (as in the Internet) and multimedia information transport with QOS guarantees (as in broadband networks). Figure 2. The Schedulable Region of a Multiplexer with Three Traffic Classes Based on the schedulable region broadband networking abstractions such as virtual circuit, virtual path and virtual network with QOS guarantees can be readily defined. This entails the concepts of admissible load and contract regions. The reader can find further details in [9], [10]. 3. Modeling Multimedia Abstractions with QOS Guarantees In this section we extend the resource characterization of networking devices to multime- September 28, 1994 6 dia computing. We define the appropriate abstraction for characterizing multimedia resources and call it the multimedia capacity region. Underlying this definition is a set of service classes with QOS requirements. The latter will be discussed in detail in section 4.2. We start in section 3.1 with the modeling of a multimedia workstation with multiple processors. Section 3.2 presents the multimedia capacity region of the audio/video unit. Section 3.3 discusses the multimedia capacity region of the disk storage unit. 3.1 Modeling the Multimedia Workstation The system architecture of a multiprocessor based multimedia workstation is shown in Figure 3, [3], [5]. The Audio Video Unit (shown as AVU below), is responsible for multi- ATM Network Multimedia Workstation Switch Fabric Camera mP Monitor mP M AVU mP M M MPU DSU disk Figure 3. The System Architecture for a Multimedia Workstation media processing, and supports media processing tasks in a deterministic manner, and runs on a dedicated processor(s). The input/output subsystem is similarly modelled, sep- September 28, 1994 7 arately through the Disk Storage Unit (DSU), and is also run on a separate processor(s). The Main processor unit (MPU) runs thesystem tasks, both to increase speed and to remove external interrupts, as well as the other Operating System (OS) overheads associated with all the application tasks. We see an advantage in this configuration because the three different processors are manipulating different granularity of data, and thereby, operate on different time scales. A video conferencing application might require processing of frames, the best effort data swapping of pages and finally data stored on the disk transfer of blocks. The MPU also runs the multiprocessor based multimedia operating system and controls the AVU and the DSU. As already mentioned in above, all the OS related tasks, such as demand paging, swapping and other interrupt processing related Operating System tasks are executed on this processor. The dynamics of the above mentioned tasks are usually difficult to model. This architecture enables us, however, to isolate the AVU and DSU services with better understood statistics, from MPU services with unpredictable statistics and, thereby, to substantially simplify the admission control problem with guaranteed QOS. By isolating the execution of various multimedia applications to specialized processors, a characterization of the AVU and DSU as a device abstraction that guarantees QOS becomes straightforward. As in [2], [17], real-time scheduling and memory management algorithms running on these processors will be used to support such an abstraction. Note that, the architecture of the multimedia workstation described above runs counter to the developments in broadband networks. For broadband networks the integration requirements of video, voice and data led to an integrated architecture supporting a single protocol data unit, the cell. Note also that the scheduling model for the AVU and the DSU will be very similar to the link scheduling model of the multiplexer in section 2. Each processing unit services a fixed number of service classes (instead of traffic classes) and has a fixed processing capacity provided by its processor (instead of the link capacity of the multiplexer). 3.2 The Multimedia Capacity Region of the Audio/Video Unit The number of combinations of calls for which QOS guarantees can be provided is called the multimedia capacity region for the AVU, and is given in Figure 4 and Figure 5. Each point on the plot represents the number of simultaneous CD Quality, MPEG-I, MPEG-II and JPEG audio-video data retrievals that can be supported by the AVU. The size and shape of the multimedia capacity region of the AVU is determined by the onboard memory and processor capacity, the scheduling and memory allocation algorithms, the service class statistics and the QOS parameters. One significant difference between the schedulable region and the multimedia capacity region is in the number of classes supported. The number of service classes at the user level is expected to far exceed the number of traffic classes at the multiplexer. However, September 28, 1994 8 a number of service classes can be mapped onto a single traffic class of the multiplexer, and therefore, supporting a large number of service classes will not require an increase in the number of traffic classes. The Multimedia Capacity Region is shown in Figure 4 and 5. Figure 4 depicts the 3 dimensions of the MCR with no MPEG-II sources active, whereas Figure 5 shows the MCR with 1 MPEG-II service class active. The multimedia capacity region shown above abstracts away the lower level details of the AVU, such as the various operating system overheads and the scheduling policy used. It provides a resource model for the AVU in terms of call support for the different service classes. The strong similarities with the schedulable region of a multiplexer are apparent. On the abstract level the only difference is dimensionality of the Multimedia Capacity Region vis-a-vis the Schedulable region, and the size and dimensionality of the data blocks that the AVU manipulates. This is in terms of frames as opposed to cells as in the case of the broadband network multiplexer. Details about the experimental set up for obtaining the data in Figures 4 and 5 are given in section 4.3 3.3 The Multimedia Capacity Region of the Disk Storage Unit (Video on Demand) Figures 8 and 9 show the multimedia capacity region of the DSU of a video on demand prototype. Each point on the plot represents the number of simultaneous MPEG-I, MPEG-II and JPEG video data retrievals that can be supported by the DSU. Figure 8 represents the capacity of the DSU if video data is retrieved without loss or delay. Figure 9, on the other hand, represents the multimedia capacity of the same DSU when data is retrieved with the allowable QOS loss and delay parameters. Both of these plots have been obtained using experiments which are described in section 4.4. September 28, 1994 9 MPEG-I Video 16 14 12 10 8 6 4 2 0 0 1 16 14 12 10 8 6 4 2 0 2 0 3 4 1 5 2 6 7 3 8 4 9 10 5 11 CD Quality Audio 6 12 13 JPEG Video 7 14 8 15 9 16 Figure 4. Multimedia Capacity Region for the AVU [D=55ms, loss = 0%, num (MPEG-II) =0] MPEG-I Video 16 14 12 10 8 6 4 2 0 0 1 16 14 12 10 8 6 4 2 0 2 3 0 4 1 5 2 6 7 3 8 4 9 10 5 11 CD Quality Audio 6 12 13 7 14 JPEG Video 8 15 16 9 Figure 5. Multimedia Capacity Region for the AVU [D=55ms, loss=0%, num (MPEG-II) = 1] September 28, 1994 10 MPEG-I Video 16 14 12 10 8 6 4 2 0 0 1 16 14 12 10 8 6 4 2 0 2 3 0 4 1 5 2 6 7 3 8 4 9 10 5 11 CD Quality Audio 6 12 13 JPEG Video 7 14 8 15 16 9 Figure 6. Multimedia Capacity Region for the AVU [D=55ms,loss=10%,num(MPEG-II)=0] MPEG-I Video 16 14 12 10 8 6 4 2 0 0 1 16 14 12 10 8 6 4 2 0 2 3 0 4 1 5 2 6 7 3 8 4 9 10 5 11 CD Quality Audio 6 12 13 7 14 JPEG Video 8 15 16 9 Figure 7. Multimedia Capacity Region for the AVU [ D=55ms, loss=10%, num(MPEG-II)=1] September 28, 1994 11 MPEG-I Video 12 10 8 6 4 2 0 0 MPEG-II Video 2 1 2 8 9 6 7 1 0 12 10 8 6 4 2 0 3 4 5 JPEG Video Figure 8. Multimedia Capacity Region for the DSU [MPEG I: data loss=0%, frame delay=0ms; MPEG II: data loss=0%, frame delay=0ms; JPEG: data loss=0%, frame delay=0ms] MPEG-I Video 12 10 8 6 4 2 0 0 MPEG-II Video 2 1 2 9 8 7 6 1 0 12 10 8 6 4 2 0 3 4 5 JPEG Video Figure 9. Multimedia Capacity Region for the DSU [MPEG I: data loss=2%, frame delay=0ms; MPEG II: data loss=0%, frame delay=0ms; JPEG: data loss=5%, frame delay=0ms] September 28, 1994 12 4. Experimental Evaluation of the Multimedia Capacity Region In this section the methodology and experimental set up for determining the multimedia capacity of the audio/video and storage unit is presented. The dependency of the MCR on the various parameters is investigated. The experimental methodology used is discussed in section 4.1. In section 4.2 we give a general description of the service classes employed. In sections 4.3 and 4.4 the experimental evaluation of the multimedia capacity region of an AVU and a DSU is described. 4.1 Experimental Methodology As already explained in Section 3.2, the MCR is a n-dimensional space (for n service classes), where each interior point represents the number of calls supported for all the service classes, while maintaining their QOS constraints. Typically, as in Figure 4, if the number of service classes supported stand-alone is, equal to 3 for MPEG-II, 15 for MPEG-I, 7 for JPEG, and 15 for CD Quality audio, the number of candidate points on the MCR will be on the order of 15x15x7x3. In order to determine the exact shape of the MCR, the viability of the above points has to be determined. Such an approach is extremely time consuming and methods to cut down the number of searches must be adopted. In our experiments, we started with a hyperplane as an approximation of the MCR. The hyperplane was obtained by first experimentally determining the three operating points representing the maximum number of standalone service classes that can be multiplexed with QOS guarantees. (The hyperplane is uniquely determined by these points.) Each of the points residing on the hyperplane (only the positive values) are evaluated whether they reside within or outside the MCR. If an operating point is found to be inside the multimedia capacity region, its designated neighbor is tried out next. The designated neighbour is the next point chosen by moving along the direction of the normal to the hyperplane. If the designated neighbour is found to be within the MCR, this process is repeteadly tried out till a neighbouring point is found to be lying outside the MCR. This process is exhaustively repeated for all the points on the hyperplane untill the MCR is obtained. 4.2 Service Classes Four service classes have been considered for the evaluation of the multimedia capacity of an AVU and a DSU, namely JPEG, MPEG-I, MPEG-II video streams and CD Quality Audio. These four service classes have different statistical characteristics and QOS requirements. For simplicity, each of the service classes is defined below together with its QOS requirements. The experimental data was employed in a live environment. • Class I: JPEG video streams with a peak bit rate of 2.5 Mbits/s. These streams correspond to a video window size of 320x240 pixels and a frame rate of 30 frames/s. The maximum allowable delay is fixed at 55 ms and a 2-5% frame loss rate is tolerated. Note that, for a fixed compression coefficient (called the Q factor), the exact size of each video frame (in bytes) is variable reflecting the actual information content. September 28, 1994 13 • Class II: MPEG-I variable bit rate (VBR) video streams with a peak rate of 1.5 Mbit/s. These streams were generated for a video window size of 352x240 pixels. The general frame pattern comprises the I, P and B frame types. In the experiments, the video stream was generated using the frame pattern IBBPBBPBBI. The frames tolerate a 55 ms delay with 2-5% frame loss. The frame rate was fixed at 30 frames/sec. The exact size (in bytes) of each frame varies within each frame type • Class III: MPEG-II video streams with a constant bit rate (CBR) of 10 Mbits/s. This service class is characterized by a video window size of 640x480 pixels. No more than 55 ms frame delay and 2-5% frame losses are allowed. The frame pattern is fixed to IBBPBBPBBI with a frame rate of 30 frames/s. The MPEG-II data used in the experiment was generated using a software MPEG-II encoder. • Class IV: CD quality stereo audio with constant bit rate of 1.411 Mb/s. This service class is characterised by a 55ms delay and no losses. 4.3 The Multimedia Capacity Region of the AVU For our experimental purposes, we have used a 2-processor SPARC 10/512 machine, running Solaris 2.3. One of the Sparc processors was dedicated to the systems tasks (the MPU) whereas the other processor was assigned for media processing (the AVU). We used digitized data already stored in Unix flat files, since we did not have access to a digitizer or compression card which can handle multiple data streams. The data read from flat files was sent using ATM-AAL3/4, over an ATM LAN. Since the data is read from a unix flat file, the Unix filesystem, buffer management, takes care of retrieving data blocks ahead of time and in parallel, and hence the experimental results are independent of the data layout on the disk. Figure 10 shows the process level model for the AVU participating in a video conferencing session. The compressed digitized media frames are grabbed by the workstation software, and thereafter processed for transmission and sent to the receiver. Note that Figure 10 shows only the modeling of sub-processes for transport from the sender side; the modeling of sub-processes needed to support the video stream arriving from the receiver side are not depicted. We have used a static priority, capacity based scheduling (CBS) policy [16], to schedule the various service classes. The CBS policy was implemented at the user level using the Solaris 2.3 real-time scheduling facilities. The implementation enables the scheduling of periodic threads running at static priority levels. The timing resolution of the periodic thread on our platform, was less than 1ms for external events, and 10ms for internal processing events. For using the CBS policy, the capacity requirements of the four service classes were first benchmarked. The capacity requirements specify the time required for the CPU to grab a media frame (point A in Figure 10 above) and send it out through the switch fabric of the multimedia workstation (point B in Figure 10). 25000 frames for each of the service classes were benchamarked to obtain the capacity values. For the CD quality audio ser- September 28, 1994 14 vice class, the size of each media frame was chosen to be 6.5KBytes. Benchmarking experiments for determining the capacities of the four service classes are tabulated in Table 1. Hardware WorkStation Encoder Application Transport A A/D Converter Kernel Video Video Device Driver B ATM Network Video Figure 10. Process Level Model of an AVU Through benchmarking we found that the worst case scheduling time assigned to a service class represents statistical outliers. Therefore, by following the recommendations of the scheduling time assignment of the CBS policy an over-reservation of the CPU capacity for most of the frames would take place. We have chosen instead the value of the capacity for each service class to be the average value plus half the standard deviation of the benchmarked values. Table 1: Capacity values used for the scheduling of the service classes Service Classes Chosen Capacity Value (mS) JPEG 4.0 MPEG-I 2.0 MPEG-II 8.0 CD Quality Audio 1.5 The MCRs shown in Figure 4 and 5, use the CBS scheduler to schedule the service September 28, 1994 15 class streams with QOS constraints as shown. For the convinience of the reader, the four dimensions of the MCR is translated into 2 separate plots of 3 dimensions each, with 0 MPEG-II service class calls active as in Figure 4, and with 1 MPEG-II service class call active as in Figure 5. There are no frame losses introduced during the scheduling of the service classes, although the service class definitions have them. This is because the Capacity Based Scheduling algorithm does not allow frame losses. We have developed another version of the Capacity Based Scheduler, which introduces frame losses at the scheduling level. This scheduler has been used to get the MCRs in Figure 6 and 7. Together this figures show the effect of the lossy CBS scheduler on the Multimedia Capacity Region. As the definition of the service classes for JPEG, MPEG-I, and MPEG-II permits 2-5% frame losses, we can see that the MCR depicts an increase in capacity (vis-a-vis without frame losses, as in Figures 4 and 5), because it can now support a larger number of calls for the different service classes. Intuitively, this is because, with frame losses allowed, the scheduler can schedule and support a larger number of service classes at the same time. The multimedia capacity determined experimentally and depicted in Figures 4 and 5 above is closely related to the notion of schedulable bound of the CBS policy. It is more general, however, as it applies to any scheduling or memory management algorithm (or both). 4.4 The Multimedia Capacity Region of the DSU Figure 11 shows the software and hardware components of a DSU configured as a disk sub-system of a Video on Demand (VOD) prototype. Main Memory SCSI-2 hard disk video data retrieval threads MPEG-I, MPEG-II and JPEG Video Segments disk Scheduler workstation data bus Figure 11. Process Level Model of the DSU. September 28, 1994 Dedicated CPU 16 The host system is a 2-processor SPARC 10/512 machine running Solaris 2.3 operating system. It consists of a SCSI-2 hard disk with an average seek time of 10 ms, maximum seek time of 40 milliseconds and a maximum external data transfer rate of 80 Mbits/s. The disk holds three service classes consisting of video segments encoded with the MPEG-I, MPEG-II and JPEG standard algorithms, respectively. The characteristics of these video segments are described in section 4.2. During VOD operation, the video is retrieved on a frame by frame basis from the disk and onto the workstation’s main memory over the internal data bus. The latter can sustain a throughput of 160 Mbits/s. Two important parameters that contribute to the size of the multimedia capacity region of the DSU are the disk layout strategy and the disk access scheduler. There has been much research work reported on the way video data can be arranged on the surface of the disk in order to efficiently support multiple video stream retrievals [18], [19]. Other research has proposed algorithms that perform video data caching schemes to maximize the number of VOD users that can be supported [20]. Whilst these results are important for increasing the total number of video streams that can be simultaneously admitted, no data striping and caching have been implemented on our initial prototype. Instead, each video segment is stored as a file. Video streams are accessed through a standard file system interface with no optimizations. Current work is focusing on storing video on contiguous tracks on disks and retrieving the data using low-level disk read primitives. The function of the disk scheduler is to control the execution of the video data retrieval threads. Its objective is to satisfy the QOS requirements of all three service classes (as specified in section 4.2), and at the same time to maximize the number of video streams that can be simultaneously retrieved. In our case, the scheduler was implemented with a non-preemptive periodic scheduling policy (with a period of 1/15 s). For the specified frame rates of each of the service classes, the scheduler allows 2 frames (or 30 frames/ s) for each MPEG-I and MPEG-II stream, and 2 frames (30 frames/s) for each JPEG stream to be served in each cycle. Non-preemptive scheduling means that the data for each service class is retrieved completely before another service class is retrieved. However, the order of retrieval for each service class is determined by the scheduler. Two sets of experiments were carried out to demonstrate the influence of the QOS parameters on the multimedia capacity region of the VOD prototype disk sub-system. In the first set of experiments different QOS constraints were considered than those described in section 4.2. In this case, all video data streams were retrieved sequentially without any delay or loss. This means that each frame was retrieved at the precise interval of 1/30 second. Unlike the AVU experiments, the delay caused by each frame retrieval is assumed to be zero, as this number is expected to be relatively small (but dependent on the exact location of the data on the disk). Given a similar setup, its influence on the overall shape of the multimedia capacity region is expected to be small. In the second set of experiments the QOS constraints on the frame loss for each of the service classes as described in section 4.2, were exactly satisfied. The frame delay was again assumed to be zero. In the first set of experiments the disk sub-system was found to support up to 6 simulta- September 28, 1994 17 neous JPEG, or 12 MPEG-I or 2 MPEG-II video stream retrievals. Other combinations in the number of these three service classes that satisfy the less stringent QOS requirements are shown in Figure 8. In the second set of experiments, each frame was grabbed without any delay but a frame loss was introduced for each service class according to the definitions in section 4.2. The loss was achieved by skipping the reading of video frames from the disk at appropriate time instances. During this time, the scheduler attempted to perform extra video retrievals by allowing the execution of other threads. The net effect of these allowances is that the disk sub-system can now support a larger number of combinations of service classes. The multimedia capacity region shown in Figure 9 illustrates that independently, a maximum of 7 JPEG, 12 MPEG-I and 2 MPEG-II video streams can be retrieved. The work presented here shows that the multimedia capacity region represents a footprint of the VOD prototype disk sub-system. The multimedia capacity region depends on the disk scheduling policy, the disk layout strategy, service class statistics, the disk hardware and the actual QOS parameters. Possible improvements or changes on either of these components will result in a disk system multimedia capacity region that can support a higher or a lower number of service class calls. The multimedia capacity region quantitatively reflects such changes. It represents a high level abstraction that separates low level information such as disk access from high level information such as call access. These naturally evolve on different time scales. 5. Conclusions Abstractions with QOS guarantees are well advanced in the broadband networking arena. We extended them to the multimedia arena, and thereby, came up with a uniform virtual resource characterization for QOS. Why did the extension of the concept of schedulable region work for multimedia devices? Both networking as well as multimedia devices are either producers, consumers or processors of media. Stability is a general concept that applies to these irrespective of the size of data they manipulate. It is, therefore, not surprising that the same quantitative characterization of networking and multimedia devices is possible (and highly desirable). Hence, we expect that the concept of multimedia capacity can also be applied to realtime protocol stacks running multiple connections on the same processor under QOS requirements. Note that the obtained multimedia networking abstractions represent objects that the operating systems should provide to multimedia computing platforms. We are doing just that on our multiprocessor platform and hope that such abstractions will be universally supported by micro-kernel implementations. This view towards modeling of resources substantially reduces the complexity of the overall multimedia network. September 28, 1994 18 On an abstract level the concepts of schedulable and multimedia capacity region are almost identical. In practice, however, the main difference appears to be in dimensionality and size. This is because the number of service classes that might be supported by a multimedia workstation can be expected to be larger then the number of traffic classes supported by a network multiplexer. The volume of traffic in number of calls, however, will be much larger for the multiplexer then the workstation. Concluding, the schedulable region and the multimedia capacity regions provide a suitable abstraction for modeling multimedia networking entities with QOS guarantees. A connection management protocol will use these abstractions (objects) to bind the physical resources together and set up communication channels among them, all with QOS guarantees. A detailed description of these mechanisms will be presented elsewhere. 6. Acknowledgment The authors would like to thank Tan Joo Geok for generating the data streams used in the experiments. References [1] [2] [3] [4] [5] [6] [7] [8] [9] Campbell, A., Coulson, G., Garcia, F., Hutchinson, D. and Leopold, H., “Integrated Quality of Service for Multimedia Communications”, Proceedings of the IEEE INFOCOM’93, San Francisco, CA, March 1993. Coulson, G. and Blair, G.S., “Micro-kernel Support for Continuous Media in Distributed Systems”, Internal Report No. MPG-93-04, Department of Computing, Lancaster University, 1993. Gopalakrishnan, R. and Bovopoulos, A.D., “A Protocol Processing Architecture for Networked Multimedia Computers”, ACM Operating Systems Review, Vol. 27, No. 3, July 1993, pp. 19-33. Govindan, R. and Anderson, David P., “Scheduling and IPC Mechanisms for Continuous Media”, 13th ACM symposium on Operating System Principles, 1991 Hayter, M. and McAuley, D., “The Desk Area Network”, ACM Operating Systems Review, Vol. 25, No. 4, October 1991, pp. 14-21. Hehmann. D.B., Herrtwich, R.G., Schulz, W., Schuett, T. and Steinmetz, R., “Implementing HeiTS: Architecture and Implementation Strategy of the Heidelberg High Speed Transport System”, Proceedings of the Second International Workshop on Network Operating Support for Digital Audio and Video, IBM ENC, Heidelberg, Germany, 1991. HP, IBM and SunSoft, “Multimedia Systems Services Architecture”, Response to the Multimedia System Services Request for Technology, June 1, 1993. Hyman, J., Lazar, A.A. and Pacifici, G., “Real-Time Scheduling with Quality of Service Constraints”, IEEE Journal on Selected Areas in Communications, Vol. SAC9, No. 7, September 1991, pp. 1052-1063. Hyman, J., Lazar, A.A. and Pacifici, G., “A Separation Principle between Scheduling and Admission Control for Broadband Switching”, IEEE Journal on Selected Ar- September 28, 1994 19 [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] eas in Communications, Vol. 11, No. 4, May 1993, pp. 605-616. Hyman, J.M., Lazar, A.A. and Pacifici, G., “VC, VP and VN Resource Assignment Strategies for Broadband Networks”, Proceedings of the Workshop on Network and Operating Systems Support for Digital Audio and Video, Lancaster, United Kingdom, November 3-5, 1993, pp. 99-110. Lazar, A.A., “Challenges in Multimedia Networking”, Proceedings of the International Hi-Tech Forum, Osaka, Japan, February 24-25, 1994, pp. 24-33. Lazar, A.A., “A Research Agenda for Multimedia Networking”, position paper at the Workshop on Fundamentals and Perspectives on Multimedia Systems, International Conference Center for Computer Science, Dagstuhl Castle, Germany, July 4-8, 1994. Lazar, A.A., Bhonsle, S. and Lim, K.S., “A Binding Architecture for Multimedia Networks”, Proceedings of the Multimedia Transport and Teleservices, November 14-15, 1994, Vienna, Austria. Lazar, A.A. and Pacifici. G., “Control of Resources in Broadband Networks with Quality of Service Guarantees”, IEEE Communications Magazine, Vol. 29, No. 10, October 1991, pp. 66-73. Leslie, I.M., McAuley, D. and Mullender, S.J., “Pegasus - Operating Systems Support for Distributed Multimedia Systems”, University of Cambridge Computer Laboratory, Technical Report No. 282 and University of Twente Faculty of Computer Science, Memoranda Informatica 92-76, December 1992. Mercer, Clifford, W., Savage, Stefan, Tokuda, Hideyuki, “Processor Capacity Reserves for Multimedia Operating Systems”, Internal Technical Report, CMUCS-93-157, School of Computer Science, Carnegie Mellon University, May 1993. Oikawa, Shuichi, and Tokuda Hideyuki, “User Level Real-Time Threads: An Approach towards High Performance Multimedia Threads”, Network and Operating System Support for Digital Audio and Video, 4th International Workshop, Nov 1993. Rangan, P.V. and Vin, H.M., “Admission Control Algorithms for Multimedia Storage Management”, Proceedings of the 3th International Workshop on Network and Operating System Support for Digital Audio and Video, La Jolla, CA, November 12-13, 1992, pp. 56-68. K. K. Ramakrishnan et. al., “Operating System Support for a Video-on-Demand File Service”, Proceedings of the 4th International Workshop on Network and Operating System Support for Digital Audio and Video, pp. 225-236, November 1993. Yu, S.P., Chen, M. S., Kandlur, D.D., “Grouped Sweeping Scheduling for DASDbased Multimedia Storage Management”, ACM/Springer Verlag Multimedia Systems, Vol. 1, No. 3, 1993, pp. 99-109. September 28, 1994