Multimedia Servers: Applications, Environments and Design / Edition 1

Chapter 1:Introduction

In the real world, the load placed on a multimedia service may be unpredictable. Therefore, from the point of view of an operator of a multimedia service, the server should operate well under varying conditions. The service may serve a large or a relatively small number of users, and the demand for videos may be rapidly changing. For a large-scale service, interruptions may be expensive.

Scalability. The number of users supported by a multimedia server can be small (as in the MTV broadcast video application) or potentially very large (as in the VOD example). Hence, the multimedia server should be capable of operating efficiently over a large range of numbers of users. As will be shown later, various sophisticated optimization policies are feasible with a large number of users [1,19]. Additionally, the number of users may grow with time. Therefore, the multimedia server should allow for easy incremental growth [15].

Reliability. In the VOD application, failure of the multimedia server can lead to large loss of revenue and goodwill. By using redundant components as backup, multimedia servers can offer very high reliability. Hence, video blocks or parity blocks may be replicated to provide protection against failure [30,7,56].

Dynamic adaptation to workload. In a multimedia system, the workload may change unpredictably. For example, the Bell Atlantic trial found that most of the demand was for new movies, which change periodically. As in the case of movies released in theaters, the actual demand may exceed or undershoot the predicted demand. It may also vary with time and have pronounced peaks andvalleys. Policies that allow the multimedia server to adapt to varying loads may be required [17].

1.2.3 Architectural Requirements

Large-scale multimedia servers are built using preexisting system components (e.g., storage devices, processors, and so on). The characteristics of these already existing components (e.g., processing speed, space, connectivity) strongly influence the structure of multimedia servers and the optimization policies used to address cost performance.

Standard logical subcomponents. It is desirable to use standard utilities (e.g., backup programs) for managing multimedia servers. This can be achieved if the multimedia server is architected in terms of standard logical subcomponents. For example, if the storage component of the multimedia server is architected as a file system (i.e., it provides a file system interface), it can be backed up by standard backup programs [30].

Topology. The underlying design of the components used to build the multimedia server may dictate a particular topology (e.g, centralized, partitioned). As later chapters will show, the topology of the multimedia server has an important influence on its performance.

Cost performance. Particularly in large-scale multimedia servers, small increases in efficiency may lead to large reductions in multimedia server costs. For example, caching popular videos in memory may reduce disk storage costs [18,14]. As mentioned earlier, with network bandwidth being a significant bottleneck, caching and prefetching in the network maybe important [51, 59, 62]

1.3 Overview of Encoding and Data Compression Technologies

In this section, we provide an overview of video encoding and data compression technologies, a working knowledge of which is essential to all three communities-multimedia application developers, application deployers, and server designers.

in the physical world, video and audio data is continuous (analog) in nature. To be able to store this data in a computer, the audio and video data needs to be digitized. The first step in this process is to sample it at a sufficiently fine interval. For example, a still image may be divided into a large number of pixels. However, each sampled value is still analog in nature; in the previous example, the image parameters at each pixel (e.g., brightness) are analog. The final step in digitization is to quantize the sampled values; for example, by rounding off the sampled value to the nearest integer. The sampling and quantization ultimately determine the final quality of the resulting video. The greater the number of samples and the finer the quantization levels, the greater will be the resulting quality.

The digitized data obtained by this process is typically too large to be stored or delivered over a network. For example, consider a 1024 X 1024 monochrome display where the brightness at each pixel is represented by a single byte. Each pixel in such a display can have 256 brightness levels. Storage of a single picture will require 1MB of storage, and delivery of a video will require 30 MB per second (MB/s). The resources required for storage and delivery can be reduced by taking into consideration the large degree of redundancy in the digitized data as well as the presence of fine and unnoticeable details. For example, successive lines of pixels in an image may be very similar. In the case of video data, successive frames in the video may also be very similar. By encoding only the differences between successive pixels or lines, the resource requirements of the video and audio data can be substantially reduced.

Because compression techniques exploit spatial and temporal redundencies in video, it follows that compression ratios could vary. For example, in a video segment with very little motion, there would be a very small difference between successive frames. Because the compression techniques encode the differences between successive frames, the compression ratio for this segment would be very high. Correspondingly, a segment with large amounts of motion could have a low compression ratio. Therefore, the bandwidth requirements of the compressed video could fluctuate. This is referred to as variable bit rate (VBR) encoding. VBR streams pose challenges in video storage, retrieval, and delivery. An alternative technique is to attempt to keep the bandwidth of the compressed stream constant and allow the resulting picture quality to fluctuate (imperceptibly one hopes). This technique is known as constant bit rate (CBR) encoding.

1.3.1 MPEG Standard

The Motion Pictures Expert Group (MPEG) has defined a set of standard audio and video compression algorithms based on the ideas above. The MPEG-1 standard defines a compression standard for audio and video data that has a bit-rate of around 1.5 Mbps [29]. MPEG-1 also defines the MPEG transport protocol, the format by which a set of concurrent audio and video streams (e.g., a movie and multiple soundtracks) can be interlaced and stored as a single stream. The interlaced stream has a header that describes the streams being combined, and this is followed by the interlaced blocks of the actual streams.

The MPEG-2 standard extends MPEG-1 to define a higher-quality compression standard that has a bit-rate of 4 Mbps to 9 Mbps [29]. MPEG-3 originally targeted high definition TV (HDTV) but is now defunct because it was discovered that MPEG-2 satisfies the HDTV requirements equally well. MPEG-4 is a standard currently under development for low bit-rate applications such as mobile communications and videophones [29]. Because MPEG-1 is widely used and is the basis for MPEG-2, we will describe it in more detail. Details of the other standards can be found in the references.

MPEG-1 video starts with an uncompressed 352 x 240 pixel frame. This frame is decomposed into one 352 X 240 luminance (brightness) channel and two color channels. Because lower resolution is needed for color, the resolution of the color channels is 176 X 120. The channels are further decomposed into macroblocks, which are the units on which encoding is actually done. Macroblocks are 16 x 16 in the luminance channel and 8 X 8 in the color channels. The data in each macroblock (or its difference from another macroblock) is encoded by applying discrete cosine transformation (DCT), which yields a sparse matrix. This matrix may be quantized into discrete levels for the purpose of reducing the storage space...