By Deepak Kataria, LSI Corporation

Executive Summary

There is a critical shift underway in the marketplace today, as media services are becoming increasingly digitized and interactive, and advances in compression technology are reducing the bandwidth requirements. The network access bandwidths widely available as consumer services are increasing. The convergence of these trends coupled with advanced telecommunications capabilities and media-rich content is fundamentally changing the nature of on-demand service delivery. Because the technology makes content effortlessly accessible, customers have come to expect anywhere, anytime delivery of very high-quality video and other rich media types to both wired and mobile devices. Video-on-demand (VoD), for example, is expected to grow to a very respectable $27 billion industry by 2011—a mere four years away. All elements of the supporting infrastructure, therefore, must be in place if this remarkable transition is to successfully take place. VoD is expected to grow exponentially, creating a host of challenges—and opportunities—for the fixed line network operator in terms of congestion management, content location services, billing, recommendation engines, storage, and timely delivery.

In order for this transition to take place, the following elements must be in-place: high-speed access and transport, capable server solutions, adequate and widely dispersed storage (both user and network-based), media processing, delivery route diversity, content processing, and QoS (traffic and congestion management). At the same time, an architecture optimized for the delivery of on-demand, media-based content must be in place and fully functional.

This paper explores the network and system architectures and the delivery strategies required for the move to on-demand video to be successful. Central to the paper's theme is an examination of VoD access and transport architectures and the technology components embedded within them, including a proposed hybrid approach that involves of a distributed proxy server model and a managed, scalable peer-to-peer distribution scheme. The model is designed to dramatically reduce the cost of widespread video distribution.

The Future of VoD

There is a critical shift underway in the marketplace today, as media services are becoming increasingly digitized and interactive, and advances in compression technology are reducing the bandwidth requirements. The network access bandwidths widely available as consumer services are increasing. The convergence of these trends coupled with advanced telecommunications capabilities and media-rich content is fundamentally changing the nature of on-demand service delivery. Because the delivery technology makes content effortlessly accessible, customers have come to expect anywhere, anytime delivery of very high quality video and other rich media types to both wired and mobile devices. Furthermore, content owners are looking for new and more effective ways to get their content in front of more people and to monetize their content in as many ways as they possibly can. The "perfect storm" of content availability, technological capability, and a high-speed, universally available telecom infrastructure has opened the door for the accelerated transition to video-on-demand on a wide-scale, lucrative basis.

Critical Factors: Media

Ongoing changes within each of the media, technology and telecom sectors guarantee that video-on-demand will become a major force in media delivery over the next few years. We'll examine each sector in turn, beginning with media. In the final analysis, this is what the industry is about—delivering all forms of available content to as many different device types as possible. Of course, this cannot be a haphazard effort in the hyper-competitive world of media selling. Media placement must be a concerted, planned, well-analyzed effort if it is to be successful. A good understanding of each customer's purchasing history, download preferences, and preferred technology and delivery modalities maximizes the potential of future sales. By analyzing past purchases over a long period of time and comparing the results to purchasing trends within the greater market, media owners, in partnership with their service delivery partners, can create individual customer profiles that can then be used by search and recommendation engines to more accurately predict and recommend future sales. 
 

Figure 1: The standard diagram used to represent Long Tail Economics

It is equally critical that these organizations understand the economics of the media industry. Contrary to popular belief, the bulk of the value of a content owner's media library does not lie in the small number of "first tier" properties that are currently on the big screen or flying off rental store shelves. Ironically, the majority of the value lies in the vast number of older properties that, if placed in front of the right customer at the right time, would yield massive sales. The ability to collect rental and preference data, analyze it en masse, determine and make potential recommendations, therefore, are critical elements of revenue growth within the media space. This model for analyzing and then acting upon the demands of a massive, aggregated market is called Long Tail Economics1. The details of Long Tail are beyond the scope of this article, but suffice it to say that Long Tail Economics is something of a restatement of the 80:20 Rule, in the sense that 80% (or so) of the value of any given content library lies in the yellow "long tail" shown in Figure 1, not in the short, dense red region of current bestsellers. The provider that devises a way to monetize the content within the Long Tail wins the game.

Closely related to Long Tail is an equally fascinating phenomenon known as Zipf's Law. Originally coined to describe the repetitive nature of natural language speech in terms of the frequency with which any particular word is spoken, Zipf can now be used as a modeling tool for the manner in which products and services in a dynamic marketplace are sought after and purchased. For the purposes of video-on-demand, Zipf observes that at any point in time, a small number of VoD titles will be requested frequently (recent releases, for example) while many others will be requested increasingly frequently, often in inverse proportion to its rank in the "long tail." So, the most popular title will be requested approximately twice as often as the second most frequently requested title, which occurs twice as often as the fourth most frequent title, etc. The implications of Zipf's Law are significant. On the one hand it can be used to create probability models for rental requests; on the other hand it can be used to determine the degree to which certain titles can be compressed for archival storage. Similar to Long Tail Economics, Zipf is a useful algorithm.

Equally important, and central to the theme of this article, is the increasingly important role that video-on-demand is beginning to play in the market. Traditionally considered the nemesis of media delivery because of the perception that it cannibalizes the sale of DVDs and tape, VoD is now being recognized by the media sector as a viable and potentially long-term ally in the battle for media relevance. It also facilitates the delivery of the "triple play" of voice, video and data. Long considered something of a Rosetta Stone for service providers struggling to ensure their own long-term relevance, triple play assumes VoD as one of its three elements. And since streaming content is expected to grow to a $27 billion industry in the United States alone by 20112, it's clearly a force to be reckoned with.
Critical Factors: Technology

The technology sector plays an equally critical role in the delivery of video-on-demand. Consider the elements of the evolving VoD universe that lie between the customer and the desired content. Today, the content is resident on a storage device (usually a DVD) that is in the hands of the customer. But as technology and market acceptance make VoD a lucrative reality, an interesting transition takes place. Content that traditionally resides on a device at the customer's premise is migrating into the network and taking up residence on one or more servers from which it is streamed to the customer, on-demand. This implies a need for massive storage arrays, complex interaction among peer servers, powerful packet processing capabilities, data collection and billing algorithms, and a network capable of transporting increasing volumes of media-rich content.

Of course, none of this is relevant if it fails to deliver content to the customer in a timely and accurate fashion. Measurable Quality-of-Service (QoS) must be guaranteed, and in a network and processor-dependent delivery environment, traffic management is key. A delivery architecture must be designed and put into place that assumes worst-case scenarios and offers full-proof workarounds for them.
Critical Factors: Telecom

The final telecom sector represents something of a linchpin in the overall delivery scheme. Each of the three sectors is fundamentally dependent on the other two, and this is particularly true of the telecom component. The richest content available and the most complex and capable data collection, storage and billing model do nothing for the customer without a dependable, broadband infrastructure underlying them. A number of relatively recent advances in the telecom domain have in many ways reinvented the role of the network, particularly as it relates to on-demand content.
The biggest change, of course, is the inexorable evolution from a relatively rigid TDM-based network to a far more flexible IP-based network. And while the shift is still in the early stages as far as the telcos are concerned, that won't be the case for long. Implementations are already underway to convert entire service provider networks to IP and MPLS, thus guaranteeing the availability of QoS that is indistinguishable from that provider by the TDM-based legacy Public Switched Telephone Network (PSTN). And along with the network conversion come IP-dependent service innovations: witness the continuing successes of IPTV, VoIP and IMS.

It is worthwhile to mention IMS here because it will play a central role in the delivery of on-demand video in the not-too-distant future. Originally crafted by the Third-Generation Partnership Project (3GPP) as a way to deliver content to mobile devices, IMS is now considered by most to be the mainstay technology platform for multimedia content delivery over all networks. Its promise is far from subtle: Once IMS is universally deployed, a service provider will be able to deliver any content to any user anywhere in the world, over any network, using any access modality, to any device—and billed according to customer preferences that are stored in a network database. This remarkable capability captures the essence of streaming video and VoD as business imperatives. A media, technology and telecom infrastructure with the ability to analyze customer preferences and act on them, that guarantees consistently high QoS, and that delivers content on the customer's terms, not those of the service provider, is as good as it gets. The road to service delivery, however, is complex. In the following section, we examine the evolution of VoD architectures and what it will take to get there.

VoD Architectures

As VoD increases in popularity and as service providers, technology designers and media owners converge, the delivery architecture evolves as well. The best known VoD architecture in use today in large networks is often referred to as the SHE-VHO-VSO model. The architecture is somewhat hierarchical: A "Super Headend" (SHE) aggregates content on a wide-area or national basis and delivers live national content across the wide area, typically using IP as the network layer protocol. A "Video Hub Office" (VHO) works with the SHE to integrate local and regional content into the mix, thus ensuring availability of both national and local content. Finally, the "Video Switching Office" (VSO) maps the diverse content from the network into the access domain for distribution to the customer.

Figure 2: The architecture of a traditional single server

This model works well for traditional content delivery, but as interest in on-demand services grows and as it becomes a larger component of the revenue stream, several service considerations come into play. Given that the overarching goal of any VoD distribution model is to minimize delivery latency, maximize network service dependability, create an easily managed (and therefore billable) service environment, and ensure a dependable, predictable environment for media streaming, differing architectural approaches emerge. Because traffic bottlenecks are one of the most common and vexing considerations in server-based networks, a move away from a single-server solution is a natural progression. Traditional media streaming, illustrated in Figure 2, delivers content from a single server that is typically owned by a Content Delivery Network (CDN) provider such as Akamai. In this environment, requests for content are transmitted to a central site that oversees the entire content delivery process. Once the request has been validated, the central site transmits the request to a proxy server located within the network, which in turn delivers the requested content to the customer.

The Distributed Server Model

In a perfect world, this model works well. However, because of the vast complexity of most networks, content streamed from a single site is fraught with potential problems, as we'll see shortly. As a result, a multiserver architecture is often deployed, and content is streamed from multiple servers to a single client. This spreads the potential for failure across multiple devices and also eliminates the chances of an ingress or egress bottleneck at the server. In this model, content downloads are driven by the receiver, not by the transmitter, which eliminates the need for coordination among the various servers involved in the download. Furthermore, requests for content at the servers occur at the block level, which means that there is no need for complex and time-consuming packet processing. Finally, bandwidth utilization is adaptively managed among the pool of participating servers, which means that (1) server loads are kept relatively uniform and (2) the potential for a server failure to cause a severe service interruption is nil.

Peer-to-Peer Streaming

Following on the heels of the distributed server model is a variation on that theme. The server that is ultimately responsible for streaming the content to the customer is often referred to as the streaming engine. In a peer-to-peer environment, such as that used by PPLive, PP-Stream (no jokes, please) and CoolStreaming, the primary streaming engine partners with peer servers and downloads relevant blocks from its partner devices for delivery to the customer. This results in a sharing of the transport responsibility among diverse servers and paths, increasing the overall efficiency of the system.

The Challenges

The evolution from a single server model to a peer-to-peer model proceeded in lock-step with the demands of the market, but even the peer-to-peer model has its challenges—the biggest one being absence of a service provider managed delivery. All VoD architectures are fundamentally dependent on three variables: the access technology, network transport, and the server or servers that control content distribution.

Figure 3: Problems can occur in the access, transport and server domains of VoD environment.

Figure 3 illustrates the potential problems that can occur in these three "regions." The broadband access component rarely experiences problems; as long as the selected access technology offers adequate bandwidth, problems don't occur unless there is a physical breach in the loop. Network transport, on the other hand, can experience a physical failure within the network core, but far more likely is a service-affecting scenario that involves traffic management.

When the network becomes congested as the result of inadequate engineering, a lack of route diversity, or a temporary situation that causes suboptimal transport behavior, packet delivery becomes sporadic and therefore unpredictable—which in turn makes QoS unpredictable. Transient conditions in the network can lead to reduced bandwidth availability, which in turn can lead to packet loss, delay, and jitter (variable delay between packets), all of which lead to inferior QoS and a poor viewing experience for the customer.

One solution to this problem as we noted earlier is to stream content simultaneously from multiple servers as shown in Figure 4. In this scenario, requests from clients terminate at the central management site and are then transmitted simultaneously to multiple proxy servers, which proceed to deliver the desired content to the customer in parallel streams.

Figure 4: The multiple server scenario. Content is streamed to the user from multiple sources.

The third and final region that can affect service quality is the server itself. As a highly specialized and robust PC, the server suffers from the same potential problems that plague standard PCs: hard drive failures, memory crowding and software failure leading to a crash. A failure in any one of these three "regions" can be a showstopper. The architectural evolution that has taken place in the short time during which VoD has been commercially available is a testament to the need to protect against a simple failure that leads to a catastrophic service outage.

So we have now described three distinct approaches to VoD streaming: a single-server approach (SHE-VHO-VSO); a multiple edge server approach; and a peer-to-peer approach. Each of these has its advantages, but none of them are perfect. To satisfy the increasingly rigorous demands of a media-hungry market, a new approach is needed, an approach that is in reality a hybrid of the previous three.

Operational Considerations for the Future

Before we describe this proposed hybrid technique for video distribution, let's first examine the realities of the VoD market. At the heart of any VoD system is the server array that makes it possible to store, locate and distribute content on an on-demand basis. These servers comprise streaming engines (sometimes called video pumps), a switching function at layer two, and one or more hard drives for content storage. Storage requirements in this environment are enormous: it isn't uncommon for a VoD server to require space for more than 10,000 hours of on-demand programming. They are typically paired with some form of asset management system that automates the acquisition, management, analysis and storage of warehoused content.

These devices also have industrial-strength processing requirements. They must handle video transcoding tasks as well as compression, both of which require powerful and diversely capable media processing DSPs.

The network has its own set of challenges. Unlike broadcast environments that stream the same content to multiple users, the VoD environment relies on a unicasting model that must, by design, support diverse control protocols such as the Real-Time Protocol (RTP) and the Real-Time Streaming Protocol (RTSP). To ensure compliance with QoS guarantees, the network must also support "trick-play" modes of operation, which include standard VCR or DVD-like capabilities: pause, fast-forward, rewind, etc. And access, typically DSL or similar, may evolve to include Gigabit Ethernet as media demands grow.

VoD: A Hybrid Approach

The solutions described above, while good, tend to have a common set of significant shortcomings. Because of their distributed server architecture and the need to ensure flawless QoS, they tend to lease far more bandwidth than they actually need at any point in time, this driving up the cost of service delivery. They also rely on an increased number of proxy servers to guard against the impact of a single server failure and to distribute the delivery of content to each customer. The proxy servers they rely on must have inordinately large storage capacity which again drives up the price of the delivered service. The result? Very high quality, but at a cost that may not be acceptable. Let's examine an alternative that permits the server elements in the VoD network to share both storage and processing resources. The technique is called Peer-Assisted Video-on-Demand.

Peer-Assisted Video-on-Demand

Peer-Assisted Video-on-Demand, sometimes also called Multiple Source Streaming VoD, is a functionally complex but highly effective architecture for the delivery of unicast video content. In a multiple source streaming system, shown in Figure 5, the raw video stream is broken into multiple blocks of video and transmitted over the network in much the same way that a message is broken into packets prior to transmission over an IP network.

Figure 5: In multiple source streaming environments, the raw video feed is broken into sequential blocks that are ultimately routed to the user.

If implemented properly, the result is better fault tolerance, bottleneck avoidance, dramatically improved use of available bandwidth, and support for significantly higher playback speeds.

Successful implementation of this methodology requires the deployment of a Multisource Streaming Scheduler (MSS). The MSS receives the streamed blocks from the various server sources (proxy servers or peer-servers), orders the blocks appropriately, and feeds the seamless, constant block rate stream to the client's device—typically a set-top box. The MSS, of course, is not without its own challenges. Variability in the delivery of data blocks can lead to buffer underflows at the MSS, which can in turn lead to playback starvation and choppy playback. Silicon-based algorithm execution can help to alleviate the problem somewhat, but more control is needed to guarantee a predictable (and billable) degree of QoS.

Figure 6: Diagram of a typical modern VoD network that relies on proxy streaming.

By implementing the model shown in Figure 6, QoS-aware VoD can be implemented over virtually any network type. Edge-based proxy servers communicate with one or more Video Block Servers, which in turn communicate with the Multisource Streaming Scheduler. The MSS interfaces with a local download buffer, which in turn relies on the services of a Local Buffer to Client Streamer function (LBCS), which ultimately interfaces with the client video player. The MSS, download buffer and LBCS together make up what is known as the Indirect Local Proxy Streamer (ILPS, sometimes called a Local Proxy Stream Server, or LPSS). This architecture, dependent as it is on simultaneous streaming from multiple servers to a single client, has numerous advantages. It is highly fault-tolerant, capable of redirecting traffic as required to avoid a hard failure in the network or service-affecting congestion. It supports streaming of high-resolution video as well as multiple simultaneous streaming sessions. This model also allows for redundant data to be transmitted as a hedge against failure, thus improving overall service quality. Finally, when transport resources are idle, servers in the system have the intelligence to "pre-fetch" content to avoid delays when traffic increases again later.

Summary

Video-on-demand has come a long way since its inception in Hong Kong in 1990. Initially something of a novelty, today it has become a mainstay element of the revenue plans for most major service providers and a strategic consideration in the product plans of both component and systems manufacturers. The convergence of media, technology and telecom may be something of a perfect storm for the industry at large, but there is a bright horizon just beyond. Peer-to-peer networking is becoming the primary video distribution mechanism in IP-based networks, and the technology elements that are coalescing into the modern VoD network lend themselves to reduced network and service delivery cost, more efficient use of bandwidth, and very acceptable QoS levels. The technology is also focused on the unique requirements of ISPs: content processing allows traffic volume to be monitored with a high degree of granularity so that traffic shaping can be performed to avoid "boundary disputes." Ultimately, the hybrid approach outlined above works well for all concerned.