Video On Demand has become increasingly popular. Giant television
providers in the United States have committed to provide VOD services
in the near future. Interactive Video On Demand (IVOD) is an extension
of VOD in which additional functionalities such as Fast Forward,
Fast Rewind, and Pause are implemented. These functionalities
pose new requirements and challenges on the system implementation.
An IVOD system has three components: Client's "set-top box",
network, and servers with archives. The clients' set-top boxes
are their interfaces to the IVOD system. It has a network interface,
a decoder, buffers, and synchronization hardware. Clients input
their commands using remote controls. Network of an IVOD system
must be a high speed network. Currently available technologies
that are suitable for transferring IVOD data include SONET, ATM,
ADSL, and HFC. Servers with archives are places where user commands
are processed and where movies are stored. Issues such as admission
control, servicing policies, and the storage subsystem structure
must be considered when designing the IVOD system. In addition
to the technical issues described above, non-technical issues
such as standards, property rights, and cost must also be considered.
Table of Contents
|List of Figures||iv|
|List of Tables||iv|
|4||Quality of Service||7|
|6.1||Network Requirements of the IVOD System||9|
|6.2||Existing Network Technologies||10|
|7.3.2||Movie Storage Data Distribution||16|
|7.3.3||Data Block Placement||17|
|7.4||User Traffic Characterization||19|
|8||Ways of Handling Interactive Functions||20|
List of Figures
|Figure 1||A Centralized Interactive Video On Demand System||2|
|Figure 2||A Centralized IOVD System with Local Buffers||3|
|Figure 3||A Distributed Interactive Video On Demand System||3|
|Figure 4||Communications Between Clients and Servers||4|
|Figure 5||A User's Set-Top Box||5|
|Figure 6||An IVOD Storage Hierarchy||6|
|Figure 7||The Overall Interactive Video On Demand System Architecture||6|
|Figure 8||A Network Model for IVOD Systems||10|
|Figure 9||An ADSL Local Distribution Network||11|
|Figure 10||A Hybrid Fiber Coax Local Distribution Network||12|
|Figure 11||A Typical Structure of an IVOD Server||13|
|Figure 12||Data Flow Graph of an IVOD Server||13|
List of Tables
|Table 1||Interactive TV Trials in the United States||22|
|Table 2||Interactive TV Trials Outside the US||23|
With the explosion of Internet, people have endless hype, opinions,
forecasts, and beliefs about it . Interactive Television,
they feel, is the vision to their beliefs: people will soon be
able to purchase products, view movies, play video games, browse
Internet, participate video-conferences without leaving their
houses . Of all the new things that people can do with television,
video-on-demands is highly supported by Hollywood since it can
lead to new markets and can bring them unpredictable profits.
People have been passive participants in receiving what TV service
providers offer since television was introduced. Interactive Video
On Demand (IVOD), unlike traditional television delivery, provides
users with flexibility in choosing the kinds of information they
like to receive . An IVOD system is capable of serving a large
number of end users to concurrently access large number of repositories
of stored data, often movies . IVOD is basically an extension
of Video On Demand (VOD). In addition to the freedom of choosing
movies, users can interact with movies and decide the viewing
schedule . In other words, IVOD system supports VCR-like functions,
such as fast forward, rewind, pause. The enormous communication
bandwidth and disk bandwidth required, and the Quality of Service
(QoS) demanded necessitate a careful design of the system in order
to maximize the number of concurrent users while minimizing the
An IVOD system comprises of 3 major components: the "set-top box" at the client's site, the distribution network, and the server. There are many design issues to consider in building each of these components. In this paper, we first discuss several different alternatives for the system structure -- the placement of video servers. We then provide a general description on each of the system components. In Section 3 and Section 4, some interactive functions and the QoS expected from an IVOD system are presented. The next three sections discuss each of the system components in details: The hardware requirements for the "set-top box" at clients' sites are outlined in Section 5. Network requirements and existing potential network technologies for implementing an IVOD system are studied in Section 6. Section 7 presents various issues to consider in building an IVOD server, including admission control to guarantee QoS, servicing policies employed to serve viewers, and user traffic characterization to improve system's performance. Design issues to consider in building the storage subsystem in an IVOD server are also discussed. They include the structure of the storage subsystem, the movies' distributions in the subsystem, the placement schemes of data blocks on storage devices in the subsystem, and the disk scheduling algorithms used to retrieve data blocks for viewers. Several proposed ways of supporting interactive functions are outlined in Section 8. Lastly, Section 9 lists some of the existing VOD trials.
2. System Architecture
As with other networked systems, IVOD can be designed as centralized multimedia systems or distributed multimedia systems. A centralized IVOD system places processing servers and media archives in a single site as a central node. Requests from clients are processed at the central node, and videos demanded are delivered through the network to the client sites. Figure 1 illustrates a centralized system architecture. Centralized IVOD systems are simple to manage, but they usually suffer from poor scalability, long network delay, and low throughput. The performance of centralized IVOD systems can be improved if local servers are added. These local servers have video buffers, but no media archives. Popular movies can be stored in local video buffers so that they can be delivered to clients more quickly. Videos that are not buffered at local sites can be delivered to clients from the central archive when they are requested. A distributed IVOD system has local processing servers and media archives. Clients' requests are handled by local servers (Figure 3). If the movie requested is not in the local archive, the local server can request the movie from remote servers located across the network.
Figure 1. A Centralized Interactive Video On Demand System
Figure 2. A Centralized IVOD System with Local Buffers
Figure 3. A Distributed Interactive Video On Demand System
A distributed IVOD system can be viewed as many small regional
IVOD systems connected together. The distributed IVOD system spreads
users' requests to many sites, thereby moving the processing servers
and media archives closer to the clients. Local servers reduce
network delay and congestion as experienced by central servers,
but distributed systems are more difficult to manage. The choice
of the system structure depends on the available storage, communication
systems, costs, application demands, and other factors. However,
the desired QoS of IVOD systems (described in Section 4)
makes the distributed structure more preferable.
Each IVOD connection requires a bi-directional communication between
the client and the local server. Each server has a number of video
selections available for users. The server processes the client's
requests and tries to respond to the clients' demands as soon
as possible. An IVOD system should be able to handle hundreds
or even thousands of clients with different preferences simultaneously
. The quality of each service should remain in specific bounds
throughout the entire session. An IVOD service usually starts
from a client requesting information from a server; the server
responds via the network to the client.
The system architecture of an Interactive Video On Demand system
basically consists of three major parts: a client, a network,
and a server. Each part can be subdivided further into components
and interfaces. Figure 4 depicts the communications between clients
Figure 4. Communications Between Clients and Servers
A client subscribing to an Interactive Video On Demand service
has a display device (usually TV) and some audio devices (e.g.
speakers) to present the movie requested. He/She interacts with
the system via an input device such as a remote control, a mouse,
or a keyboard. A controller is needed at the client site to take
the client's commands and to send the signal to the server through
its network interface. The controller also stores video signals
it receives from the server into its buffers, decodes the compressed
signals, and sends the decoded signals to the display at the appropriate
time. The controller is assembled in a box, known as the "set-top
box." Figure 5 depicts the components at the client site.
Figure 5. A User's Set-Top Box
2.2 The Network
An IVOD service requires real-time display of the video purchased
by the client. A typical video stream consists of frames of pictures,
sounds corresponding to those frames, and captioned text. The
large quantity of information needed to be transmitted to the
client continuously with minimal delay poses high performance
requirements on the network. An IVOD network should be a high
speed network with reasonable error rate as retransmission is
unacceptable. Since video information is delay sensitive, the
delay variation (jitter) should be kept to a minimum.
2.3 The Server
A server of an IVOD system processes commands from users. It accepts
or rejects the clients' requests based on the current state of
the system and the network load (refer to Section 7.1).
It also performs scheduling on the retrieval of data for all active
clients. A multimedia archive is connected to the server. The
archive contains a collection of videos available to the users.
Depending on the system requirements and the budget available,
a range of storage devices can be used: cache (RAM) is the most
expensive but has the lowest access time. Disk-arrays provide
fault-tolerance at a reasonable price and access rate (10 msec).
Optical discs have a capacity of 650 MB with access time 100 msec.
Digital Versatile Disc (DVD) is state of the art. Each disc can
stores 4.7 gigabytes of information. The content of movies stored
on DVD discs can be easily configured to suit viewers' preferences
with the help of authoring tools. Tape drives are on the lower
price range, but with longer access time. A typical IVOD storage
system uses a combination of storage devices to optimize the tradeoff
between cost and efficiency. Figure 6 depicts a general IVOD storage
Figure 6. An IVOD Storage Hierarchy
Combining all the components above, an IVOD system is constructed. The overall system architecture of an IVOD system is shown in Figure 7.
Figure 7. The Overall Interactive Video On Demand System Architecture
As mentioned before, IVOD is an extension of VOD with additional
interactive functionalities added. Possible interactive functions
Other interactive features include the ability to avoid or select
advertisements, to investigate additional details about news events
(through hypermedia, for example), and to browse, select, and
purchase goods . Seven main interactive functions: Fast Forward,
Fast Reverse, Jump, Play, Stop, Pause, and Slow Down will be discussed
in this paper.
Quality of Service (QoS) can be used in many different aspects.
For instance, QoS in networks may include guarantees on the throughput,
network delays, delay jitter, error rate, etc. QoS in this paper
refers to the required standards of IVOD services expected by
users. They include:
Fast setup time: Initial service delay (connection setup
time) should be within a few minutes.
Independence of Quality of Service to different customers:
Quality of service provided to the current customers should not
be degraded due to the joining of new customers to the service.
Continuity of media streams: There should be no or little jitter in presentation
Prompt response to interactive functions: The loading of
extra data from the server due to interactive functions, e.g.
fast forward, should be invisible to the customer in the ideal
case. It should appear as if the client is operating his/her own
Transparency of the involvement of multiple media streams:
Multiple media streams must be synchronized. For instance, if
a video object is to be combined with an audio object at the client's
site, this should be done in such a way that lip-synchronization
All of the above QoS require cooperation of all three components
in the IVOD system: server, network, as well as the "set-top
To support interactive services, considerable functionalities
must be built into the set-top box. Four important components
for set-top box are Network Interface, Buffer, Decoder and Synchronization
5.1 Network Interface
The network interface allows the client to receive data from server.
Moreover, it provides a mechanism to translate user commands received
by the sensor to appropriate signals for transmissions on network.
In order to save storage space, disk bandwidth, and network bandwidth,
movies are usually encoded before they are stored. Thus, a decoder
at the client's site is needed to decode the arrived media streams
before presenting them to the viewer.
Due to delay jitters in the network, the arrival time of a video
stream cannot be determined exactly. In order to guarantee starvation-free (continuous) playback, the server must ensure a media unit is
available at the client prior to its earliest predicted playback
time. By taking into account the maximum network delay (max),
the server can transmit a media unit at least max prior
to the unit's earliest predicted playback time. However, if the
actual network delay experienced by the media unit is less than
max, the media unit will arrive at the client earlier
than its scheduled playback time, and will have to be buffered
. For details on the computations of the maximum buffering
that is needed at the client's site, refer to .
If the buffer size is large enough and the Jump size is relatively
small, Jump Forward or Jump Backward may not require data delivery
from the server. In other words, the data required are already
in the client's buffer. The response to the jumps will be faster
as compared to those requiring data from the server. Buffering
also makes the response to Fast Forward and Fast Rewind faster
if the initial data required is in the buffer.
As mentioned above, decoding is required at the client's site.
Buffering of data that may be required for future display allows
data to be decoded while the current data is being displayed.
Less powerful decoding hardware or decoding algorithm can be used
since real-time decoding is avoided. Thus, the cost of the "set-top
box" can be reduced.
5.4 Synchronization Hardware
A movie consists of both video and audio streams. They must be
synchronized before being presented. Synchronization is required
at the client site to support scalable video . In scalable
video, a video stream is decomposed to a base stream and one or
more additional streams. The additional streams are to be combined
with the base stream to produce higher quality video. Each of
those streams is stored as individual media files on the server.
Depending on the quality demanded and the bandwidth available
to the client, the presentation may require zero or more of the
additional streams . Therefore, different media streams must
be synchronized before presenting to the customers.
Lastly, the user interface should be simple and easy to use. Preferably,
the same remote control can be used for both the IVOD system and
the video cassette recorder. In addition, the cost of set-top
box must fall within reasonable limits (under a few hundred dollars)
for the technology to succeed. Open and interoperable systems
that let users subscribe to several different services are preferred
Unlike traditional computer communications that are bursty and
short-lived, the deliveries of videos involve sending enormous
amount of data to customer homes continuously for a long-period
of time (the length of the movie presentation). Videos are usually
encoded using MPEG-1 or MPEG-2 standards. Since MPEG encoding
standards exploit the intra-dependancy between frames, the resulting
encoded video streams are usually of variable bit rates. These
characteristics of videos lead to new criteria and challenges
on the network.
6.1 Network Requirements of IVOD Systems
High Speed Network. High speed networks are definitely
needed for IVOD systems. For example, videos compressed using
MPEG standards require a bandwidth between 1.5 and 6 Mbps. A system
supporting 100 users requires a bandwidth close to 600 Mbps. Ordinary
10 Mbps Ethernet or 28.8 kpbs telephone lines cannot support such
Connection-Oriented Transfer. IVOD services are real-time
multimedia application services. In other words, the time at which
packets arrive at the destination is strictly bounded. Any packets
arrived later than their expected time are useless and discarded.
Moreover, retransmissions are not possible . Consequently, connection-oriented
services, which reduce the rate of dropped and late packets, are
Latency and Jitter. The delay between a video stream being
sent from the server and it being received by the client needs
to be minimized. Variations in delay (jitter) must be kept within
rigorous bounds to preserve the quality of the presentation .
In order to provide IVOD services, the network needs to deliver
guaranteed services to delay-sensitive variable-bit-rate video
traffic. The delay sensitive characteristic of video services
requires the network to support some resource reservation schemes
for each video stream.
6.2 Existing Network Technologies
As mentioned above, a typical video stream requires an average
bandwidth of 1.5Mbps.
A typical IVOD network can be divided into two levels: backbone and local distribution networks. The backbone connects servers and routers/access nodes together, while the local distribution network links a client site to its nearby router/access node. Each local distribution network link needs a bandwidth of at least 1.5 Mbps - the bandwidth of a single IVOD video stream. The bandwidth of the backbone is usually on the order of hundreds megabits per second, depending on the number of simultaneous connections the IVOD network has to support. Figure 8 illustrates a network model for IVOD systems.
Figure 8. A Network Model for IVOD Systems
6.2.1 Backbone Network
IVOD services require high-speed and low-jitter networks to support
hundreds or even thousands of simultaneous connections. The required
bandwidth of the backbone is on the order of hundreds megabits
per seconds. Two attractive solutions are SONET and ATM.
SONET. SONET is a synchronous fiber optic network. The
entire bandwidth of a fiber optic link is devoted to a single
channel. Nodes connecting the channel transmit data at different
time slots. A basic SONET channel (STS-1) has a bandwidth of 51.84
Mbps. SONET can also multiplex multiple digital channels together
 to support more viewers. For example, three STS-1 channels
are multiplexed to form a STS-3 channel with 155.52 Mbps bandwidth.
SONET is suitable for delivering IVOD streams because bandwidth
is guaranteed and jitter is zero.
ATM. ATM is the acronym for Asynchronous Transfer Mode.
It is asynchronous because it allows data arriving at irregular
intervals . ATM is suitable for transferring IVOD data because
it is a connection-oriented packet switching network. ATM transmits
data at a rate from 1.544 Mbps to 622 Mbps, over copper and fiber
Note that a backbone with ATM running over SONET can also be adopted.
6.2.2 Signaling Schemes
A local distribution link requires a bandwidth of 1.5 Mbps. Many
signaling schemes can deliver video data at such a data rate over
existing communication links . Two such schemes are described
Asymmetrical Digital Subscriber Loop (ADSL). ADSL is a signaling scheme used on copper-wire networks (e.g. telephone networks). It can deliver data at high speed with few signal distortions over existing copper wires. ADSL consists of a pair of ADSL units. One is installed in the client site; the other is attached at the central phone office. ADSL uses advanced integrated circuit designs, complex digital signal processing techniques, and software-based algorithms to compensate distortions in copper wires . ADSL can provide a subscriber with a down-link of 1.536 Mbps wide-band signal, an up-link of 16 Kbps, and a basic-rate ISDN channel/the analog 4 kHz telephone channel on existing twisted copper wire . These characteristics satisfy the bi-directional communication and bandwidth requirements posted by IVOD services. If the client site is within 5.5 kilometers of the access node, no additional equipments are necessary to ensure strength of the received signal. ADSL is achievable because it divides the signal on a range of carrier frequencies, dynamically adjusting to achieve the most efficient channel allocation . Extensions of ADSL include HDSL, SDSL, S-HDSL, and VDSL. HDSL has a data rate of 6 Mbps, and it can support MPEG-2 video streams up to about 2 km . An ADSL local distribution network is shown in figure 9.
Figure 9. An ADSL Local Distribution Network 
Hybrid Fiber Coax (HFC). HFC is currently being installed by cable TV companies. It migrates the all-coaxial cable design to a network with fiber and coaxial cable. Variations on HFC exist and have been implemented. They include fiber-to-feeder, fiber-to-the-hub, fiber-to-the-zone, fiber-to-the-curb, and fiber-to-the-tap. Fiber-to-the-feeder and fiber-to-the-hub are networks with fiber trunks, coaxial distribution links and subscriber drops. The others are networks consist of fiber trunks and distribution links, but coaxial subscriber drops . The migration to HFC involves a replacement of the current 300-450 MHz coaxial cables by new 750 MHz coax cables . Each channel is subdivided in 125 6-MHz subchannels. Bi-directional communications are implemented using splitband systems. In North America, the frequency band between 5-54 MHz is used as up-link. The remaining bandwidth is guard band and down-link. Since clients share the same physical medium, collisions occur when they try to access the medium at the same time. The problem of sharing the transmission medium limits the number of clients attached to a coax tree structure . The network architecture of HFC is depicted in Figure 10.
Figure 10. A Hybrid Fiber Coax Local Distribution Network 
IVOD services are feasible on HFC because the frequency band of
the coaxial cables is splitted so that an up-link is available
for sending clients' requests to servers. Currently, Rogers Cables
Canada is upgrading its existing cable system to HFC, hoping they
will soon be able to deliver Internet access services, if not
Video On Demand services.
Server is the heart of the IVOD system. It provides high quality
services to users by using strategies that minimize cost. Server
has a storage subsystem where movies are stored. Many researchers
have explored ways to optimize the server capacity and have discussed
issues concerning building a multimedia server in general. This
section will discuss admission control, servicing policies, the
storage subsystem, and traffic characterization, which are issues
related to optimizing server performances. Figure 11 and Figure
12 depict the general structure and details of an IVOD server:
Figure 11. A Typical Structure of an IVOD server 
Figure 12. Data Flow Graph of an IVOD Server 
7.1 Admission Control
Any incoming clients who want to use the IVOD service will have
to request for and set up a connection with one of the servers.
If the required movie or parts of the required movie are not stored
in the current server, transfers of data between servers are needed.
Hence, admission control for the remote server is also required.
In order to guarantee new clients have continuous playback of
the video and to ensure the QoS contracted to existing connections
are not jeopardized, the server must have enough resources such
as storage subsystem read bandwidth, memory buffers, processing
bandwidth, and network bandwidth before admitting a new connection
request . Before establishing a connection, a set of QoS parameters
sent by the client will be checked by the server. If requested
QoS can not be achieved, negotiations can be made between the
client and the server, or the connection will be denied. An alternative
is that the server does not commit the resources requested by
the client until the connection is up for some threshold of time,
usually a few minutes . This reduces the probability of committing
resources to connections that just last for a few minutes or even
a few seconds.
An admission control algorithm must have some knowledge about
the capability of the storage subsystem, e.g. the minimum number
of blocks that the server can read in a time slot. A time slot
is the interval between serving a video stream divided by the
number of active streams in the server. Neufield,
Makaroff, and Hutchinson  suggested that the minimum number
of blocks that a server can read in a time slot, called minRead,
is the only storage subsystem information required for the admission
control algorithm. According to them, minRead can be determined
by running a calibration program that uniformly spaces data blocks
across the disk to maximize seek times (assuming a SCAN algorithm
is used). The estimation of minRead should be as accurate
as possible because conservative estimations of minRead
degrade server performances. A simple admission control algorithm
can be described as follows: sum the block schedules for all active
streams together with that of the requested stream. If the sum
is greater than minRead in any time slot, the connection
will be denied. An improvement can be made to this simple scheme
by allowing the server to read ahead when the server is idle.
The server must also ensure enough network resources are available
before accepting a new connection. Unfortunately, variable bit
rate video streams create difficulties in determining the amount
of resources needed. If the network reserves resources according
to the average video stream rate, delay or packet loss may occur
when the servers are sending at their peak rates. If the network
reserves resources according to the peak video stream rate, the
network may be under-utilized most of the time . Traffic policing
is performed on admitted connections to ensure the connections
have not over-used any network resources.
In addition to checking the storage system and network bandwidth,
the server also ensures the availability of memory buffers and
processing bandwidth. The server, based on the required QoS, determines
the amount of buffer space and the processing bandwidth needed
by the requested stream.
7.2 Servicing Policies
Servicing policies determine the design and the implementation
of various IVOD components. The server capacity, which is the
number of simultaneous viewers that can be supported by the server,
depends on the strategies used in allocating video streams to
viewers. Obviously, from the service providers' point of view,
server capacity should be as large as possible. The simplest way
of allocating streams to viewers is to dedicate a single stream
to each viewer. However, this scheme is expensive and inefficient
since the server capacity is bounded by the maximum number of
streams that can be handled by the server. If the network in the
IVOD system supports multicast, e.g. ATM, sharing of video streams
among several viewers is possible. Sharing of video streams can
significantly increase the number of simultaneous viewers and
can save network bandwidth . Three service policies have been
proposed in :
On-Demand Single Cast (ODSC). ODSC is the simplest of the
three schemes. In this service policy, each client has a dedicated
video stream that is assigned to the viewer when its connection
is granted. Since the client has complete control over his/her
own video stream, the implementation of the interactive functions
mentioned in Section 3 becomes easy. The drawback of this
scheme is that the server capacity is limited. The waiting for
services to become available can be unexpectedly long if client
requests come at times when the server is fully utilized .
Phase Multicast (PMC) or Batching. To use this policy,
the data network must support multicasting. Each video stream
is shared by viewers of a multicast group. Video streams
are started at fixed intervals or phases. Connections that
are admitted in between startings of streams are grouped and served
by the next video stream . The video streams' starting intervals
can be fixed by the service provider, or they can be determined
by the server dynamically. Without any doubts, the longer the
time interval between consecutive streams, the lesser the service
costs and the greater the number of viewers can be supported by
each stream. However, long setup time delay may cause clients'
dissatisfaction. Thus, a tradeoff exists between cost and quality
of service provided.
A serious drawback of this scheme is the difficulty in implementing
interactive functions. Since the video streams are now shared
by others, the delivery of each video stream cannot deviate from
its scheduled time. A discussion on how to support interactive
functions under PMC is out of the scope of this paper.
On Demand Multicast (ODMC). ODMC is a hybrid of PMC and
ODSC. During light load, the server uses ODSC to serve viewers.
It switches to PMC during heavy load. Thus, ODMC addresses problems
of PMC and ODSC, which include the inability of ODSC to cope with
overload situations and the underutilization of the server during
light load when PMC is used .
Simulations had been run on the above three service policies under
the condition that no interactions are allowed. Thus, the system
simulated was really a VOD rather than an IVOD. It was found that
the ODMC scheme is suitable for server that is utilized lightly
by providing fast setup time. At very high levels of utilization,
the PMC service policy should be used. The ODSC service model
should only be used when support of extensive random access, which
is what IVOD needed, is required. For details on the simulation
and the performance results, refer to .
7.3 Storage Subsystem
Server subsystem is where compressed movies are stored. The compressed
video and corresponding compressed audio streams for a movie can
be stored in the same server or in different ones. The storage
subsystem is also the place where most of the optimizations are
done to improve the performance of the IVOD system. It is an important
factor in determining the server's cost. So far, the standard
used in an IVOD system that has been agreed on is the use of MPEG-2
for video encoding .
7.3.1 Hierarchical Structure
Even with MPEG-2 compression, a movie will occupy approximately
4GB . If the fixed magnetic disks are used as the sole storage
medium in IVOD systems, the cost of archiving thousands of movies
is on the order of millions. Therefore, large tertiary storage
devices such as tape, optical jukeboxes, or the new technology
Digital Versatile Disc (DVD) jukeboxes should also be used .
Tertiary devices are highly cost effective because they provide
large storage capacities at low cost. However, their random access
time is slow, and their throughput is low. Consequently, servers
should be organized as a hierarchy that combines the cost-effectiveness
of tertiary storage and high performance of fixed magnetic disks
. Refer to Figure 6 for a server storage hierarchy.
A hierarchical storage model can be described as follows: Popular
movies are stored in RAM, less popular ones on the hard disks
and the least popular ones on tertiary storage. This hierarchical
approach reduces operating costs while offering a wide selection
of movies to its users .
7.3.2 Movie Storage Data Distribution
Several approaches are possible for managing the storage hierarchy.
One approach is to store the beginning segments of the multimedia
files in magnetic disks. This approach reduces in start-up latency
and provides smooth transitions in the playback . It improves
the system performance if viewers make another selection within
the first minutes of the movie [15, 5].
An alternative is to calculate movies' popularity on a daily basis.
Based on their popularity, movies are rearranged or replicated
as necessary during off-peak hours. Consequently, movies are available
for viewing during peak hours based on anticipated demand .
7.3.3 Data Block Placement
The way in which data blocks of media files are placed on disk
can significantly affect the data retrieval scheme and the system
performance since the time required for retrieving data depends
totally on the location of data blocks. The data retrieval scheme
can in turn affect how interactive functions, such as Fast Forward
(FF) and Fast Backward (FB) searches, are supported. One possible
approach of implementing FF and FB is to read more data blocks
for that particular client in a round. However, extra disk bandwidth
requires more storage devices to serve the same number of clients,
thereby increasing the cost of the storage subsystem. Cheng, Wen,
Lee, Wang, and Oyang  suggested that if the data blocks are
placed by a placement scheme that effectively utilize disk bandwidth
during normal playback, one can design a data retrieval scheme
which requires no extra bandwidth to support interactive functions.
However, an extra buffer of size several file blocks is needed
when a stream is in the FF or FB search modes.
Two general strategies used in data block placement schemes are
load-matching and load-balancing .
Load-matching. Placement of more frequently accessed data
blocks in "best" (outer) tracks while those that are
accessed occasionally are placed in inner tracks. This strategy
exploits the track dependent transfer rate that can have a ratio
of 1:1.8 from inner to outer tracks .
Load-balancing. Placement of data in a way such that constant
streaming capacity is provided independent of viewers' choices
and location of data.
Both of the above strategies try to maximize the server's capacity,
which is the number of simultaneous viewers supported. Details
on implementation of load-matching and load-balancing and description
on interdisk and intradisk load-balancing can be found in 
Two main schemes of data block placement are disk farm and disk
array (RAID). Each disk holds several entire movies in disk farm;
whereas in RAID, each movie is scattered over multiple drives,
e.g. block 0 on drive 0, block 1 on drive 1, and so on so forth.
This organization is called striping. RAID is preferred to disk
farm because it offers better performance .
Load matching is a possible approach under this scheme: Popular
movies are stored in outer tracks while less popular ones are
stored in inner tracks. However, popularity of movies may change
several times in a day. Thus, the capacity of the server will
decrease when such changes happen. One can dynamically rearrange
the movies on disks but this can be expensive [7, 15]. Under this
scheme, the number of concurrent access to a single multimedia
file is limited by the throughput of the disk. However, this scheme
is easy to implement.
RAID (Redundant Array of Inexpensive Disks)
Unlike storing data in a single disk, both load-balancing and
load-matching can be achieved in RAID by "striping"
all movies onto all disks [7, 15]. Under this scheme, data is
"striped" across an array of disks. The coarser the
striping, the larger the buffer required. Figure 13 describes
the structure of a RAID.
Figure 13. RAID 
Striping allows parallel accesses of data from the same physical
sectors of all disks in the array. Hence, logical sectors and
physical sectors have identical access time. The effective throughput
can be increased by a factor that is equaled to the number of
disks in the array. The increased transfer rate makes RAID a good
structure for the storage subsystem since high bandwidth is required
in supporting IVOD . RAID can have several parity disks to
provide fault tolerance. Those extra disks contain block-by-block
EXCLUSIVE OR of the other disks to allow data recovery for the
faulty disks. The number of drives' failures that can be survived
is determined by the number of parity disks added. This scheme,
however, is ill-suited to interactive functions such as FF and
FB since many of the blocks read in parallel will be discarded
7.3.4 Disk Scheduling
Real-time constraints makes traditional disk scheduling algorithm,
such as first come first serve, short seek time first, and scan,
inappropriate for IVOD. Here are two suggested scheduling algorithms
The best known algorithm for real-time scheduling of tasks with
deadlines is the earliest deadline first algorithm (EDF).
The media block with the earliest deadline is fetched first. The
disadvantage of this algorithm is excessive seeks and poor utilization
of the server's resource .
A variant of EDF is a combination of SCAN and EDF that is called
the Scan-EDF scheduling algorithm . Scan-EDF, like EDF, services
blocks with the earliest deadline first. However, when several
blocks have the same deadline, those blocks are served using the
SCAN algorithm (the disk head moves back and forth across the
disk and fetch requested blocks as it passes them). Clearly, the
effectiveness of SCAN-EDF depends entirely on the number of requests
that have the same deadline .
Under round-based algorithms, a server serves all streams in units
of round. During each round, the server retrieves a certain number
of blocks for each stream. Since MPEG-2 results in variable-bit-rate
compressed streams, the number of blocks that must be retrieved
for each client in each round will vary according to the compression
ratio achieved for each block .
To ensure continuous playback of media streams, the server must
retrieve a sufficient number of blocks for each client in each
round to prevent starvation for the round's duration. Thus, the
server has to know the maximum duration of a round as round length
depends on the number of blocks retrieved for each stream. A simple
scheme that retrieves the same number of blocks for each stream
(generally referred to as a round robin algorithm ) is inefficient
since the maximum playback rate among all streams will dictate
the number of blocks to read. This results in streams with smaller
playback rates retrieving more data blocks than needed in each
round. This may overflow some clients' buffer as well as decrease
the capacity of the server. Consequently, more clients can be
accommodated by reducing the number of data blocks retrieved per
service round for streams with lower playback rate .
One proposed approach to minimize the round length
is to make the number of blocks retrieved for each stream in each
round proportional to its playback rate . This scheme is called
Quality Proportional Multi-subscriber Servicing (QMPS)
algorithm. This algorithm is provably optimal in . For more
information on QMPS, please refer to .
7.4 User Traffic Characterization
Even though customers access the IVOD system randomly, having
a priori knowledge about the users' access pattern can lead to
a more efficient design of the storage scheme and a more efficient
utilization of the storage and network bandwidth . For instance,
if traffic characteristic indicates that a movie is popular in
a particular site, the system can replicate the movie to increase
availability. Similarly, knowing the typical user access pattern
can be beneficial in designing schemes for resource managements,
such as popularity tables updates, data redistributions, and system
reconfigurations. Preferably, these resource managements should
be done during off-peak hours .
8. Ways of Handling Interactive Functions
Several ways are proposed on the handling of interactive functions.
Most of the researches are done on implementing FF and FB. Several
proposed methods are outlined below:
Chen, Kandlur, and Yu  proposed that FF and FB can be supported
using MPEG frame skipping. Their approach composes of a storage
method, a segment placement scheme, and a playout method with
a segment sampling scheme or a segment placement scheme as alternatives
for selecting segments.
Usage of D-frames provided by MPEG-1 [7, 19]
MPEG-1 D-frames contain only the block averages that can be used
for browsing. Thus, only D-frames are presented to the client
after decoding when the client is in FF or FB modes. This method
is attractive since no additional processing is required as the
D-frames are always generated by MPEG-1. However, the resulting
outputs have poor resolutions. Unfortunately, D-frames are not
available in MPEG-2.
Block categorizing 
Unlike the previous two schemes, in which the client is responsible
for supporting the interactive functions, FF and FB are supported
by the server in block categorizing. This scheme works as follows:
Each block is marked as being relevant or irrelevant to FF/FB.
Both types of blocks are retrieved during normal playback while
only those marked relevant are retrieved and transmitted in FF/FB
modes. This scheme can be achieved by scalable compression, which
combines one or more additional streams with the base stream to
produce higher quality [12, 19]. However, scalable compression
poses additional overhead on splitting the blocks and recombining
them before presenting. This scheme requires extra bandwidth if
differential compression is used since most frames are relevant
9. Existing Trials
Table 1 and 2 list some of the VOD trials existed in the United
States and in other parts of the world . In 1995, VOD services
had started to appear in the market. However, the responses were
poor. Things started to changes in the middle of 1996. Major operators
such as Americast, Tele-Communciations Inc, Rogers Cablesystems
Ltd. have committed to buying million set-top boxes . A typical
set-top box costs about three hundred US dollars. Table 1 and
2 show that many companies have just started to offering VOD services.
Price ranges from US $7.50 to US $20.00 per month plus movie charges.
Network technologies used in the trials are mostly HFC and ADSL
because they require fewer modifications of existing cable and
telephone networks. However, when HDTV at 20 Mbps is highly demanded,
networks with ATM and Satellite may even be used in local distributions.
Analysts estimated that more than US$ 1 billion were spent in
1994 on VOD infrastructure worldwide and nearly US$ 3 billions
for 1995 .
Table 1. Interactive TV Trials in the United States
|Interactive TV Trials in the United States|
|Bell Atlantic (FutureVision)||Dover Toms River NJ||Philips, FutureVision, TeleTV||nCUBE, FTTC||NVOD, PPV, transaction||Switch DigVid||2/96-Rollout||0/38,000 (passed)||Varies|
|Bell Atlantic (Stargazer)||Fairfax Vir||Stellar One||nCUBE,||VOD, Inter.||ADSL||3/95 12/96||1,000/1,000||$7.50/mo|
|Bell South (Inter. Serv)||Atlanta GA||Sci/Atlanta||H/P||VOD, NVOD transaction||Fiber/Coax||2/96-mid/97||0/12,000 (passed)||NA|
|Cox Cable (Canceled)||Omaha NB||Zenith||N/A||VOD, NVOD transaction||Hybrid F/Coax||6/94- 12/95||0/2,000||$20+ movie|
|Pacific BellHFC in SD, SJose, Org Cty||California||Sci/Atlanta||N/A||NVOD, VOD, cable||Hybrid F/Coax||96-1997||200/1 million passed(1997-?)||N/A|
|SNET (Canceled)||Hartford CN||Sci/Atlanta||H/P||NVOD, PPV||Hybrid F/Coax||4/94- Mid 96||350/150,000 passed||$20/mo +movies|
|Sprint (VDT Trial)||Wake Forest, NC||N/A||N/A||VOD, Sega, Info||N/A||10/95- Fall 97||650/1,000||N/A|
Table 2. Interactive TV Trials Outside the US
|Interactive TV Trials Outside the US |
Primary Source: Electronic Engineering Times, Nov 27,1995
|Cambridge Cable||UK||Fiber/Coax||Online Media||1994||250|
|Deutsche Telekom||Germany||ADSL/ATM/HFC Satellite||Alcatel Nokia IBM HP||1996||1,500,000|
|French Telecom||France||ADSL, Fiber to home||Phillips, SEMA||1996||1-2,000|
|Canel+||France||Satellite||Sony Pioneer Sagem Phillips Thompson||1996||N/A|
|Svenska Kabel TV||Sweden||Fiber/Coax||Digital, Vela Research||1995||500|
|Telecom Italia||Italy||ADSL||B.Atlantic, OS-9/David||1995||1,000|
|Telecom Australia||Australia||ADSL||CLI, OS-9/David||1996||2,500|
|Hong Kong Telecom||Hong Kong||Fiber to building, ATM||NEC, OS-9/David||1996||65,000|
|Israel Telecom||Two city trial||ADSL||Celerity Server, David Settop||1996||N/A|
|Korean Telecom||Korea||ADSL||Celerity Server Samsung/David Settop||1995||100|
|Nakano City TV||Japan||HFC/ATM||Fujitsu, OS-9/David||1995||300|
|Singapore Telecom||Singapore||ADSL, ATM||N/A||1995||300|
Developing this new information delivery infrastructure requires
considerable planning and effort . In addition to meeting technology
requirements, such as network and disk bandwidth, making IVOD
service a reality also involves considerations of other factors.
They include the cost of building the system, international standard
agreement, and service providers' security. The economics of providing
IVOD service cannot be ignored. A large video server can easily
cost more than a mainframe, certainly 10 million dollars. Consider
a regional IVOD service with 100,000 subscribers, each of which
rent a 300-dollar "set-top box". If the networking equipments
cost 10 million dollars and have a 4-year depreciation period,
the service has to generate 10 dollars per home per month. Charging
2 dollars per movie and 3 dollars rental for "set-top box"
requires every subscriber buying 2 movies per month to break even
. Certainly, all the cost figures mentioned above vary with
time, but it is clear that a mass market is needed for the technology
to be viable . Unfortunately, few of the existing trials have
more than 2000 subscribers. For the development of a mass market,
agreement on international standards must be reached. Establishing
standards for "set-top box", video-file server, and
standards that allow competition among service providers are important
for such systems to grow . Currently, MPEG-2 is the only standard
being agreed on for video encoding . Mechanisms to protect
intellectual property rights must also be established so that
service providers are able to maintain control of their data and
thus are able to stay in business. One mechanism to avoid movie
duplication is to limit storage buffer space in the "set-top
In fact, the technologies required to make IVOD a reality already
exist. If an IVOD system can be built with a reasonable cost,
international standards can be developed, and mechanisms for protecting
the interests of service providers can be established, we can
easily see the appearances of large-scale IVOD services in the