MASAA : A Case Study In Building a Distributed Integrated Media Database

Transcription

1 Copyright 2000 IEEE. Published in the Proceedings of the Hawai i International Conference On System Sciences, January 4-7, 2000, Maui, Hawaii. MASAA : A Case Study In Building a Distributed Integrated Media Database Roger Zimmermann, Mohammad R. Kolahdouzan, Cyrus Shahabi Integrated Media Systems Center University of Southern California Los Angeles, California [rzimmerm, kolahdoz, cshahabi]@cs.usc.edu Abstract The possibility of immersing a user into a virtual environment is quickly becoming a reality. At the Integrated Media Systems Center (IMSC) we are exploring the unique challenges presented by such systems in the context of our own implementation, the Media Immersion Environment (MIE). A challenging aspect of immersive environments is the persistent storage of vast amounts of heterogeneous data. In this report we present our efforts and experience in building a storage infrastructure for the MIE. The different characteristics, requirements, and heterogeneity of the encountered data present new challenges for their acquisition, storage, access, and analysis. In the context of a sample application, collaborative immersipresence, hosted on the MIE, we investigated these challenges in a realistic setting to build the current version of the MIE storage infrastructure. We report on our progress so far and describe remaining challenges, extensions, and future research directions. 1 Introduction With current technological advances in computer graphics and animation, high-speed networking, signal and image processing, and multimedia information systems, it is now feasible to immerse a person into a virtual environment. The goal for immersive environments is for people to interact, communicate and collaborate naturally in a virtual space while they are actually residing in different physical locations. Our work at the Integrated Media Systems Center (IMSC) is focusing on different challenges in realizing such an immersive environment under the Grand Challenge MASAA: Multimedia Acquisition, Storage, Access and Analysis. This research was supported in part by cash/equipment gifts from Informix, Intel, JPL/NASA, contract no and NSF grants EEC and MRI System Integration Experiment termed the Media Immersion Environment (MIE) [10]. The MIE has applications in many domains, including those where it is expensive to have physical presence (e.g., distance learning), impossible to have physical presence (e.g., space exploration), important to have multiple presences (e.g., remote medicine), not safe to have physical presence (e.g., nuclear studies), and enjoyable to not have physical presence (e.g., entertainment industry). A key aspect of such collaborative, immersive environments is the necessity to seamlessly store vast amounts of data that are produced during collaborative sessions (i.e., the provision for persistent storage). The inherent structure of such recordings will make it possible to query elements of the immersion and customize the results towards user preferences. Hence, a collaborative, immersive environment can be viewed as consisting of two components, (a) a collection of clients, also called immersipresence stations and (b) an underlying communications and storage infrastructure (see Figure 1). In this study we focus on the storage infrastructure. Building such a system is a challenging task for the following reasons. First, the Media Immersion Environment provides a framework for the creation of various integrated media systems, offering an overarching, unifying theme for fitting together the different information technologies. In an integrated media system, advanced media technologies are used to combine, deliver, and transform information by way of images, video, audio, avatar animation, graphics, haptic, and text in real time. Various integrated media systems will be developed within the MIE framework and hence its 1

2 storage infrastructure should be general enough to handle the encountered data types. Second, the amount of heterogeneous data generated is vast. Audio and video clips are generally large in size, but novel data types, such as haptic information also require generous storage space, especially for long sessions. Third, an easily accessible programming interface is required to allow developers to focus on the application functionality. The different requirements of the various data types and information sources led us to employ multiple, specialized data servers to optimize the storage performance. We will now present our efforts and experience in building such a storage system for the MIE testbed. Section 2 describes one example application, termed cooperative immersipresence, to elaborate on the media types and requirements that must be supported by the storage infrastructure. Section 3 details the implementation challenges and how we addressed them or are planning to address them. We introduce different storage servers and describe their specialization. Section 4 discusses remaining challenges, extensions and future research directions for our current paradigm. 2 The Immersipresence Applications The immersive telepresence application (subsequently referred to as immersipresence for short) is designed to provide an experience for the participants in which individuals are immersed in a controlled, customized multimedia information universe; a universe that will include people, places, information collections, databases, the web and beyond. Such an experience is also termed a session and the ultimate goal is to make it appear highly similar (or indistinguishable) from a real-world setting. At IMSC we are experimenting with several specialized immersipresence applications. By specialization we mean that their efforts are focused towards specific real-world settings and domains. Two of the projects are BioSIGHT and the Haptic Museum (see [10] for other applications). The purpose of the BioSIGHT project is to develop a novel methodology to map a high school curriculum into a series of interactive visualization modules. These modules will augment the role of the teacher, and establish the value of student-centered interactivity, both in an individual setting as well as in a collaborative learning environment [5]. Central to the mission of the Haptic Museum project is the development of new technology for enhancing the exhibition of three-dimensional works of art. In particular, this project develops and explores technology that allows remote users to interact with the exhibitions. The virtual museum will allow the ability to touch priceless museum objects. Haptic interfaces will be included for the inspection and manipulation of three-dimensional art objects [9]. To support the media acquisition, storage, access, and analysis needs of immersipresence applications in general, and BioSIGHT and the Haptic Museum specifically, we designed and implemented a distributed, multi-server storage infrastructure. 2.1 Architecture Figure 1 illustrates the MIE storage infrastructure and its four organizational repositories: an object-relational database management system (OR-DBMS, currently Informix Universal Server), a real-time file server (RTFS, currently Microsoft NetShow TM Theater Server), a collaborative data server (termed Agora server), and the Ariadne information mediator [1]. This distributed repository can support the storage and retrieval of structured, semi-structured, and unstructured media types. Middleware written in Java TM glues these repositories together to provide a unified access from the immersipresence stations to the stored data. The middleware is hence loaded into the client space, along with the application graphical user interface (GUI), at run time. One of the main advantages of this design is its operation in two-tier mode which provides superior performance as compared with the three-tier architectures commonly employed in the industry. 3 Behind the Scenes We now take a detailed look at the inner workings of the MIE storage infrastructure. We will concentrate on the real-time, structured, and collaborative data support. Details of the Ariadne information mediator, that provides for the seamless integration of web information sources, can be found in [1]. 3.1 Real-time Data Support An immersistation application generates data of a variety of media types, each with different requirements. Some of them require real-time storage and delivery. Among these, the more traditional isochronous data types are audio and video streams. Such data needs to be delivered at a prespecified rate to avoid undesirable artifacts such as display disruptions. Other, newer data types of this category include avatar and haptic data [20] Avatar Data Overview One approach to immerse multiple participants into a shared virtual space is the creation of avatar models for each individual, i.e., three-dimensional representations that can be placed into computer generated environments. To achieve realistic interactions the avatars feature a detailed face model 2

3 Pentium III WWW Pentium III Pentium III Real Time File Server Socket SUN StorEdge A1000, Main Repository of Images & Data SUN Enterprise 450, Informix Universal Database Server CGI bin Ariadne Information Mediator High Speed Network Real Time RMI / JDBC Streams Middle Ware Socket Socket Middle Ware Immersidata Station Immersidata Station Collaborative Data (Agora) Server Figure 1. Two-tier, distributed, multi-server MIE storage system architecture. with color texture that can be animated to mimic facial expressions. To be effective, such avatar models must be very accurate and the animation quality very high. For the storage and retrieval of avatar data one key aspect is very important: it requires real-time processing, in particular it must be captured and played back at a rate of 30 frames per second. For each frame the coordinates of about 18 points need to be captured. A similar technique to that of haptic acquisition has been employed for avatar data acquisition. This data is then integrated with off-line recorded data representing face muscles, eye-regions, wrinkles, and volume morphing features in order to re-construct and render a realistic face animation Haptic Data Overview Haptic data is information that is associated with threedimensional objects that are part of an immersisession and may be shared among the participants. Touch and feel sensations are an integral part of the experience and are designed to make the immersion more realistic than if only visual information is provided. Haptic sensations generally require special hardware and software support. For example, at IMSC we have access to a device called PHANToM [17], which can be described as a stylus that can be moved in three-dimensional space. An integrated servo mechanism allows the device to exert constraining, inertia, and vibrating forces. A second device is the CyberGrasp system [22]. It consists of a CyberGlove that tracks the hand and its movements of a user in threedimensional space. Additionally, a lightweight, unencumbering force-reflecting exoskeleton fits over the CyberGlove and adds resistive force feedback to each finger. With the CyberGrasp force feedback system, users are able to explore the physical properties of computer-generated 3D objects 3

4 Figure 2. The PHANToM (left) and CyberGrasp haptic devices. they manipulate in the simulated world of an immersisession. Haptics data is represented with several complementary components: three-dimensional scene information, temporal positional information of the objects involved in the scene, the nature of the collisions between those objects, and the forces associated with those collisions. Two characteristics of haptic data are important for the design of a storage server that supports the storage and retrieval of haptic data. First, haptic data is temporal in nature and requires real-time retrieval for a smooth, realistic, and hiccup-free display. Second, this data type is large in size. Especially if fine-grained motions and actions are to be recorded then the amount of data generated that describe every detail of such an experience is voluminous Real-time File Server To support the real-time data types of immersisessions, for example audio, video, avatar, and haptic data, a real-time file server (RTFS) is required as an integral part of the system architecture (see Figure 1). At the current stage of our infrastructure implementation, we utilize Microsoft s NetShow TM Theater Server [13] in combination with local disk drives to support the immersistations basic real-time storage needs. However, as part of the evolution of our infrastructure we plan to replace this commercial solution with our own RTFS implementation for the following reasons: 1. To investigate protocols between the clients and the server that dynamically adjust the necessary bandwidth requirements. One of these protocols is Superstreaming [19], a technique that can improve server utilization. Furthermore, Super-streaming can take advantage of fast networks by adaptively increasing the streaming rate to the maximum rate that is available. 2. To support very high bandwidth media types that take advantage of Internet-2 class networks. Haptic data playback at very high resolutions may require higher data rates than the traditional MPEG-1/MPEG-2 streaming rates. 3. To support composite media streams with both interand intra-time dependencies that may need to be retrieved partially or entirely overlapped [6]. 4. To incorporate novel compression algorithms that are currently under investigation at IMSC. These include techniques to compress avatar and haptic data, which require different distortion metrics than, for example, video streams. 5. To investigate end-to-end quality-of-service (QoS) protocols that allow bandwidth reservation. The architecture of our proposed RTFS implementation is illustrated in Figure 3. To achieve the high bandwidth and massive storage required for a multi-user RTFS, disk drives are commonly combined into disk arrays to support many simultaneous I/O requests. A distributed, multi-node design based on commodity personal computer hardware components provides a cost-effective and scalable solution. To store a large media object on such a platform it is commonly striped into blocks: [14, 21, 7, 3, 12]. There are two basic techniques to assign these blocks to the magnetic disk drives that form the storage system: (a) in a round-robin sequence [15, 3], or (b) in a random manner [11, 4, 16]. Traditionally, the round-robin placement utilizes a cyclebased approach to scheduling of resources to guarantee a continuous display, while the random placement utilizes a deadline-driven approach. We plan to implement the random block placement in conjunction with a deadline-driven 4

5 Ethernet 100 Mb/s EthernetSwitch Gigabit Ethernet 1000 Mb/s Ethernet 100 Mb/s Node 0 Node 1 Node 2... Node N Personal Computer (~550 MHz; ~256 MB) Multiple 100 Mb/s Ethernet PCI Adapters PCI bus; 1064 Mb/s Ultra2 SCSI PCI controller Disk 0 Disk 1 Disk 2... Disk N ~ 160 Mb/s Disks: high-performance, Ultra2 SCSI (e.g., IBM Ultrastar 36ZX) Figure 3. Real-time file server architecture based on commodity personal computer hardware components. scheduler because this combination provides a low startup latency even under high loads. A low startup latency is very desirable in support of interactive applications such as collaborative immersipresence stations. To achieve high reliability and availability for all the session data that is stored in the RTFS we will implement a parity-based data redundancy scheme that, in addition to provide fault-tolerance, can also take advantage of a heterogeneous storage subsystem [24, 23]. A large-scale disk storage system tends to evolve and become heterogeneous for several reasons. First, disks are mechanical devices that might fail. Because the technological trend for these devices is one of an annual increase in both performance and storage capacity [8], it is usually more cost-effective to replace a failed disk with a new model. Moreover, in the fast-paced disk industry, the original model may be unavailable because it was discontinued by the manufacturer. Second, if the disk array needs to be expanded due to increased demand for either bandwidth or storage capacity, it is usually most cost-effective to add current-generation disks. Support for various disk models also protects customers from being locked into a single source when buying magnetic disk drives. For these reasons heterogeneous storage systems are a common occurrence. Hence, a RTFS should manage these heterogeneous resources intelligently to maximize their utilization and to guarantee jitter-free real-time delivery. Our first generation networking support will be based on Fast Ethernet technology between the RTFS nodes and potentially Gigabit Ethernet for the backbone. The timely delivery of the data will be ensured via the Real-Time Protocol suite (RTP, Real-Time Control Protocol, RTCP, and Real- Time Streaming Protocol, RTSP). Employing such Internet standard protocols facilitates the interoperability of the realtime file server with a wide range of applications. Unlike Asynchronous Transfer Mode (ATM) networks, Ethernet does not natively support bandwidth reservation to guarantee a certain Quality-of-Service (QoS) for isochronous data transmissions. We plan to incorporate the Resource Reservation Protocol (RSVP) when it is more widely adopted across routers and networks. To rapidly start our protocol suite efforts we are currently investigating the Darwin Streaming Server [2], an open source project that is available from Apple Computer, Inc. The Darwin server supports many of the protocols that we are implementing and hence it provides a tool for the testing of compatibility and interoperability issues. 5

6 3.2 Structured Data Support Immersistation applications also generate structured data that does not have real-time storage and delivery constraints, and can be easily stored in a database management system (DBMS). This category of data includes application control data, which represents the interactions of a user with an application (very low data rate), and meta-data, which represents information about the immersisession as well as information about the real-time data (i.e., its structure). An immediate advantage of storing this data is the ability to play back a session. Assume a scenario where some students in a shared session are to learn about biology materials within an educational application. We store all of the students interactions with the application (i.e., who did what at what time) into the database, so a session can be played back later. This can be useful when the students want to review a course, or for an instructor to grade the students. A second advantage is the ability to query the database. In the context of our example, one can query the database for all the sessions in which a specific person was present (querying session meta-data), all the sessions in which a specific participant looked confused (querying extracted meta-data from avatars), or all the sessions during which a particular object was being investigated (querying application control data). Another interesting advantage is the ability to employ data-mining techniques to analyze the stored data. For example, one can analyze the behavior of individual students in a collaborative session to find out how each one would have been using the system if they were working alone, based on their interactions in both the current session and the previously recorded similar sessions, as well as other students (from training sets) interactions within the same session. We use Informix Universal Server v9.14, an object relational database as our database server. This DBMS is hosted on a Sun Enterprise 450 platform with an attached StorEdge A1000 RAID storage system. We now describe three different types of structured data generated within our immersipresence environment in more details Application Control Data Overview Application control data represents interactions of the users with an application. These interactions are usually in the form of mouse movements and click actions, but what we really consider as application control data is the impact of (a series of) these actions. For example, every mouse movement while viewing a mosaic picture may cause the rotation of a picture by some degrees, and the collection of these actions constitute the control data. On the other hand, to mark an object in an image, a user may need to move the mouse around the object and specify its corners, but only marking the object is considered as an application control data and the detailed information about this action (e.g., the path on the image that the mouse was traversing prior to specifying the corners) is disregarded. The application that we use to investigate this type of data is a rapid prototype of a digital microscope (as part of the BioSIGHT application), which allows users to navigate through biological images of the human body by performing actions such as zooming, focusing and scrolling 1. The program can be executed over the Internet (via a Java applet) and works similar to a shared white-board: the effect of each action performed by a user will be simultaneously viewed by other users. Figure 4 shows a simplified conceptual schema based on an object-oriented data model that we use to store this data type. The prototype application is designed to generate control data that is compatible with the schema. We strive to develop a general schema that can support as much actions as possible for other similar applications, and be easily extendible for applications with different natures Session Meta-Data Overview Session meta-data represents information about immersive sessions and it can be divided into two different classes: General Session Meta-Data: the set of information which is in common among different applications (haptic, avatar, shared white-board). In particular, we collect the following information units (only the important ones are mentioned here) for an immersisession: session name, start and end time of the session, participants and the time that each one joined and left the session, keywords and major and minor topics. Specific Session Meta-Data: the set of data which is specific to each application. For example, for a haptic application (with the PHANToM device), we need to store information about the objects that were being explored, its distinct constituting polygons as well as the sampling rate. This data is required for both querying purposes and playback (e.g., a haptic session should be played back with the exact same sampling rate that it was recorded). A simplified conceptual schema that we use to capture this data is depicted in Figure Extracted Meta-Data Overview This type of data represents the structured data that is extracted from both real-time and application control data. Such data needs to be generated in order to support contentbased indexing, querying and retrieval of the actual data. To illustrate, for a haptic application, we extract data to specify 1 A demonstration is available at 6

7 Time Actions Time Person (Person ID#) (Object ID#) (Action ID#) Action Id Object Height,Width Color Image Image Actions Focus Zoom Load Scroll Figure 4. Conceptual schema of application control data. the time distribution during which an object was touched, the identity of the distinct polygons that were touched, and relative speed/acceleration of the PHANToM device. Consequently, we can support queries to identify the portions of an object which are of interest to different people based on the surfaces that were touched the most. For the shared white-board application, we extract data specifying what images were chosen, what parts of each image were displayed and for how long, and what objects in an image were selected. Given this data, we can identify the most interesting parts of images for students, or compute the success rate of a group of students trying to find an object in a series of images. For performance reasons, we generate the extracted data offline (after a session is completed) for real-time data, but the extraction is done in an online manner for application control data. Therefore, queries can be issued to the system while a session is still in progress. 3.3 Collaborative Data Support The MIE infrastructure (see Figure 1) can also be used for acquiring data generated within a collaborative session (e.g., the digital microscope of BioSIGHT). These data generated by two or more clients (as they collaborate within a session) are shared through a centralized controller (Agora 2 server in Figure 1). This centralized controller not only coordinates and dispatches the collaborative data but also stores them in the alternative storage servers for future querying, access and analysis. For the remainder of this section we discuss the details of the Agora server within the context of the digital microscope application which operates similar to an immersive shared white-board. The Agora server, a Java application that is part of the RIMS toolkit (see Section 3.4), makes connections to the user programs through sockets (Java s Socket class). Agora uses different ports to support simultaneous connections to different applications (e.g., digital microscope, haptic museum) and multi-session environments. It also establishes a connection to the Informix database server using RMI (Remote Method Invocation) and JDBC, and automatically glues each application to its associated database/tables based on a pre-defined configuration. The server handles the storage of commands into the database server, and distribution of interactions (i.e., application control data) between the users. Recall that the digital microscope allows two or more users to select biological images (e.g., image of a blood cell) and navigate through the images by issuing typical commands such as zoom-in, move-left, scroll, etc. The users 2 Agora is the name of the place where people discussed and exchanged ideas about life in ancient Greece. 7

8 Immersive Session Major & Minor Topics Session # and Name Attended Role People Keywords Start and End Time Join and Leave Time Rating, Participation Name ID# d... Avatar Session Haptic Session Object # Initial Magnification Sampling Rate Figure 5. Conceptual schema for session meta-data. virtually share the same interface and hence whatever action one user performs, the other users see. The user program, upon receiving any command from a user (usually a push button), transforms the command to a standard form (based on the conceptual schema) and sends the standard command off to the server program. After receiving a command, the server module distributes the command to all participants. If a command requires other types of information, for example an image, the server retrieves the image from the Informix database server and then distributes the additional information to the participants. Each user, including the originator of the command, will act based on the received commands from the server. The server also adds some meta-data information (e.g., time, the person who issued the command, etc.) to the commands and stores them into their designated tables. 3.4 A Middleware for Unified Data Access The underlying infrastructure of our Media Immersion Environment, MIE, is based on a distributed system. There are different file and database servers located in different places to support the environment. There is a need for a middleware to, first, make the distributed architecture look like a unified structure for its users. Second, to provide a unified standard method (i.e., through an Application Programming Interface, API) for the users to access the servers. RIMS, Rapid Integration of Multimedia Servers, is a toolkit that we have developed for this purpose 3. This toolkit includes three parts: servers, APIs and utilities. The current version of the toolkit contains a) the Agora (collaborative data) server and Ariadne mediator, b) APIs for connections to Informix database server, Agora server, Ariadne, RTFS and Microsoft NetShow server, and c) utilities for signing applets, and defining new sessions, (i.e., conceptual schema of their commands and their associated databases and tables). The middleware (API) is implemented in Java and is loaded into the client memory space at run time along with an application GUI. This middleware also provides the following functionality: 1. Glues the servers (DB, RTFS, Ariadne) in order to provide a unified and standard view to the clients. The client applications along with the middleware are downloaded and executed using a standard web browser from a web-server. Note that Java security does not allow an applet downloaded from one server to connect to other servers (i.e., DB, RTFS and Ariadne). Therefore, currently we have the applet signed by a third party which in turn authorizes it to connect to the other servers who thrust the third party. Once 3 An early release of the RIMS toolkit and its documentation are available from 8

9 this middleware is loaded into the client space, it supports all required interactions between the clients and the servers (i.e., storing and retrieving data, submitting queries and receiving the results of the queries). 2. Makes socket connection and interacts with the Agora server. For those clients participating in a shared immersisession (e.g., digital microscope), as they interact with the application, their actions are transmitted to the Agora server. These action are then time-stamped and stored into the database server. 3. Stores (session and extracted) meta-data into the database servers. Part of this information is populated directly by the middleware (e.g., the time a user joined/left a session), and part of it by the collaborative server (e.g., the time that a session was started/finished). In addition, certain parts of the session meta-data are manually entered by the user and subsequently stored into the database server through the middleware (e.g., keywords, major and minor topics). This information is requested from the users through the GUI before allowing them to make connections to the other servers. 4. Queries the database. Users can query the database by submitting the query to and retrieving the results through the middleware. 4 Challenges, Extensions, and Future Research Directions The MIE is envisioned as an evolutionary platform that will change dynamically over time. It will serve as a testbed for the integration, evaluation, and demonstration of advanced multimedia technologies. Therefore, we expect the MIE storage infrastructure to change and evolve over time to meet the needs of new and improved applications. There remain many challenges and extensions that have not yet been addressed in the current version of our infrastructure. We will mention a few of them here briefly. The synchronization between different media streams will increasingly becoming an important requirement for advanced media applications that incorporate multi-modal user interfaces. The human perception is very sensitive to variable delays in the presentation of media streams that have time inter-dependencies. To address this challenge we need to replace the current NetShow TM Theater Server with our own RTFS. This is because the current commercial video servers (including NetShow Theater) can only guarantee intra-stream time dependencies. However, these products cannot support a sophisticated multimedia presentation (e.g., generated dynamically as a result of a query) that requires retrieval of multiple streams (where each stream might be in different modalities: e.g., haptic, video and audio) with inter-stream time dependencies. We intend to implement the techniques we studied in [18] within our RTFS to tackle this challenge. The Application Programming Interface (API) that enables developers to access the functions of the storage infrastructure from their applications needs to be streamlined and opened for extensibility without sacrificing performance. Currently, the middleware functions that provide access to multiple storage servers are tightly coupled with our applications and their graphical user interfaces. We are in the process of separating the general functions required to access different servers from the application codes, in order to define them formally as an application programming interface, API. We also intend to extend the functionality of the API to support access to other databases (e.g., DB2, NCR TOR). Subsequently, we intend to release this API publicly for other research groups who want to rapidly integrated multiple multimedia servers for their applications. The main advantage of this API is that it provides access to multiple servers from a Java client that can be activated from a typical web-browser. We are also planning to include data-mining utilities into the RIMS toolkit to support the analysis of behavior of the users in immersisessions. Finally, as our own servers (e.g., RTFS) become available, we plan to integrate them into the RIMS toolkit and release their code as well so that other research groups can replace commercial products with our open source servers. This will provide them with a flexible and extendible multimedia storage infrastructure. The security aspects of our infrastructure have mostly been ignored so far. However, once the MIE testbed is deployed outside our campus space this will become an important issue. At the same time the system will need to prove its scalability to grow regionally and eventually nationally and internationally. The Next Generation Internet (NGI) will provide the necessary bandwidth to deploy distributed advanced collaborative media applications. The storage infrastructure needs to keep up with such growth. 5 Acknowledgments We would like to thank the following people with whom we have engaged in intriguing discussions: Craig Knoblock, Chris Kyriakakis, Margareth McLaughlin, Dennis McLeod, Gerard Medioni, Ulrich Neumann, Chrysostomos L. (Max) Nikias, Albert (Skip) Rizzo, Alexander Sawchuk, Gaurav Sukhatme, Wee Ling Wong, and Sherali Zeadally. 9

10 References [1] J. Ambite, N. Ashish, C. Knoblock, S. Minton, P. Modi, I. Muslea, A. Philpot, and S. Tejada. Modeling Web Sources for Information Integration. In AAAI/IAAI, pages , [2] Apple Computer, Inc., Cupertino, CA. Darwin Streaming Server, URL: projects/streaming/. [3] S. Berson, S. Ghandeharizadeh, R. Muntz, and X. Ju. Staggered Striping in Multimedia Information Systems. In Proceedings of the ACM SIGMOD International Conference on Management of Data, [4] S. Berson, R. R. Muntz, and W. R. Wong. Randomized Data Allocation for Real-Time Disk I/O. In COMPCON, pages , [5] The BioSIGHT Project, URL: usc.edu/. [6] S. Chaudhuri, S. Ghandeharizadeh, and C. Shahabi. Avoiding Retrieval Contention for Composite Multimedia Objects. In Proceedings of the International Conference on Very Large Databases, Zürich, Switzerland, September 11-15, [7] H. Chen and T. Little. Physical Storage Organizations for Time-Dependent Multimedia Data. In Proceedings of the Foundations of Data Organization and Algorithms (FODO) Conference, October [8] E. Grochowski. Average Price of Storage and Internal (media) data rate trend, IBM Almaden Research Center, San Jose, California. URL: technolo/grochows/g06.htm and g16.htm. [9] M. McLaughlin, L. Ellison, J. Lucas, and S. Goldberg. The Interactive Electronic Exhibition: Reconstructing the Boundaries between Museums and their Constituencies. In International Communication Association Conference, Montreal, Canada, [10] D. McLeod, U. Neumann, C. Nikias, and A. Sawchuck. The Move Towards Media Immersion. In IEEE Signal Processing, volume 16, pages 33 43, January [11] R. Muntz, J. Santos, and S. Berson. RIO: A Real-time Multimedia Object Server. In ACM Sigmetrics Performance Evaluation Review, volume 25, September [12] R. Ng and J. Yang. Maximizing Buffer and Disk Utilization for News On-Demand. In Proceedings of the International Conference on Very Large Databases, [13] Microsoft NetShow TM Theater Server, URL: [14] V. Polimenis. The Design of a File System that Supports Multimedia. Technical Report TR , ICSI, [15] K. Salem and H. Garcia-Molina. Disk striping. In Proceedings of International Conference on Database Engineering, February [16] J. R. Santos and R. R. Muntz. Performance Analysis of the RIO Multimedia Storage System with Heterogeneous Disk Configurations. In ACM Multimedia Conference, Bristol, UK, [17] SenseAble Technologies. PHANToM Users Manual, [18] C. Shahabi. Scheduling the Retrievals of Continuous Media Objects. Ph.D. Dissertation, University of Southern California, Los Angeles, California, [19] C. Shahabi and M. Alshayeji. Super-streaming: A New Object Delivery Paradigm for Continuous Media Servers. To appear in the Journal of Multimedia Tools and Applications, [20] C. Shahabi, G. Barish, B. Ellenberger, M. Kolahdouzan, D. McLeod, S. Nam, and R. Zimmermann. Immersidata Management: Challenges in Management of Data Generated within an Immersive Environment. In Proceedings of the Fifth International Workshop on Multimedia Information Systems (MIS 99), Indian Wells, California, October 21-23, [21] F. Tobagi, J. Pang, R. Baird, and M. Gang. Streaming RAID- A Disk Array Management System for Video Files. In Proceedings of the First ACM Conference on Multimedia, pages , Anaheim, CA, August [22] Virtual Technologies, Inc., Palo Alto, CA. CyberGrasp TM, URL: [23] R. Zimmermann. Continuous Media Placement and Scheduling in Heterogeneous Disk Storage Systems. Ph.D. Dissertation, University of Southern California, Los Angeles, California, December [24] R. Zimmermann and S. Ghandeharizadeh. Continuous Display Using Heterogeneous Disk-Subsystems. In Proceedings of the Fifth ACM Multimedia Conference, pages , Seattle, Washington, November 9-13,