Andrew Davis
TechOnline

This article takes a look at a new class of machines called video servers. Here we examine the major application types and look at the performance requirements and how they affect the hardware and software design.

There are two major applications of the digital technology. One is for ad insertion, which is of interest mostly to the cable television industry; the second is for a general class of problem known as video on demand (VOD), which can be a cable TV application or an enterprise training application as well. The ad insertion application, or rather the economics of ad insertion, have been driven by storage costs. VOD, because of the need for seamless, real-time, two-way interaction between users (subscribers) and the server, had to wait for compression technology and reliability improvements, as well as storage efficiency.

A video server is the multimedia equivalent of a data file server. It is a full computer system, including storage, processing, and input output functions. In addition, it contains video encoders and decoders to convert analog content into digital and vice versa. The difference between ordinary data files and video data files places extraordinary demands on the architecture of a video server. Video data, particularly MPEG-2 data, differs from typical computer files in that the MPEG file is a large, continuous, and complex stream of information that must be managed as it travels through then entire system.

VOD can actually be thought of as two classes of applications. Near video on demand (NVOD) is a set of several video streams, with starting and ending times that are separated by some fixed interval (for example by 10 minutes). Hence NVOD is a schedule-based broadcast service. Because the video streams are transported separately, a subscriber or viewer can choose when to begin viewing the content by picking a start time that matches the NVOD schedule. Should the subscriber stop viewing in the middle of the program it is possible to redirect the viewer through software control to another stream to pick up the video content near the point where he or she left off. Since there is only a redirection and no subscriber interaction with the video stream, only one video stream per time interval is needed an danynumber os subscribers can access the ocntent on any one of the video streams at any give time. The critical parameter for NVOD then is the number of separate intervals needed, which determines the required number of video streams the server must process.

In contrast to NVOD, VOD allows subscribes or viewers to see video content at any instant they choose. To make this possible, the video stream must be capable of being started on subscriber command. Since any subscriber can access the video content at anytime, the hardware and software must support a separate independent stream for each subscriber. The critical parameter here, then, is the number of video streams that can be processed at one time. In many industrial or real-world applications, the specification for the number of simultaneous streams a server must process can be determined by probability analysis, not unlike telephony traffic theory, which determines how many switch ports are needed to support a given number of users and phone lines. Because each user can request any VCR-like function at any time in a true VOD environment, it is necessary to have separate, independent video streams for each active user. Here again, the specification for the number of streams that can be processed at any instant is a critical parameter, often determined by some sort of probability analysis. The ability of the server to respond in real time to input commands rom the viewer is also a critical performance requirement.

VOD obviously requires more bandwidth than NVOD since VOD systems have to process a larger number of video streams. In most community cable systems, if VOD is offered as a service, three to six video channels are usually set aside, regardless of the number of titles offered.

When looking at the hardware and software architecture for video server computer systems, the following parameters are usually the first to be considered:

• Server Output Capacity
  This is measured by the number of simultaneous video streams which the server can process. Traditionally, these streams originated in external storage devices and were switched onto the distribution network as required. Today many designs distribute the stored video data across RAID disks which allows multiple access to the files. With RAID technology, a stream is not stored as a complete data file in one place, but is created as required by playing segments stored on each of the disks in the array.
  
  
• Storage Capacity
  In video server terms this is defined as the total hours of stored content available to the system. Some systems may not use an all-disk storage design, but rather use a hierarchical design, where some or all of the data is backed up on lower cost secondary (and slower) storage.
  
  
• Independence of Video Streams
  With independence, the video data in each stream can be independently accessed and controlled.
  
  
• I/O Performance
  This term is interpreted in a way that is peculiar to the application. I/O Performance is how well the inut and output channels move entire streams of video data and this is different from how a computer file server works. Video files are relatively long and must be delivered as a continuous stream. Therefore video server I/O channels must have direct access to video content storage to be effective; buffers are not the best design.
  
  
• Operating System Performance
  This refers to the need for a video server operating sytem to meet the I/O expectations. "In-kernel" buffer management is a term often used to describe the method of managing data movement directly from content storage to output to improve the operating system performance.
  
  
• Scalability
  This is very important in most video server applications, since it addresses the way server or storage capacity is added over time. Scalability is achieved by adding CPUs or memory to a server, or by networking complete servers.
  
  
• CPU Speed
  
  
• Reliability
  Often this is used to describe the system's capability to recover from potential system and component failures to ensure subscriber video services remain available. If the video server generates $X in revenues per hour, then downtime is expensive, and easily quantifiable.