October 1, 2005
The 10GigE Migration
By Paul W. Sherer, Arroyo Video Solutions
Only a few years ago, 10 Gigabit Ethernet (10GigE) seemed at best a vision for the future. But today it is widely available in network interface card (NIC) interfaces for servers, within switches, and supported by wideband fiber transport infrastructures for both local and remote intersystem communications applications.
Within the same timeframe, video services have migrated rapidly away from the original point-to-point architecture of application software interface (ASI) transport and time division multiplexing (TDM)-over-fiber, to GigE transport. Internet protocol (IP) video and video-on-demand (VOD) over GigE infrastructures are now widely deployed.
The availability of such high-speed infrastructure presents novel opportunities for the design of systems carrying IP video and VOD. With such flexibility, subscriber "churn" will be reduced, and cable operators will be able to architect solutions vastly richer and more functional than ever seen in the past.
This article retraces the evolution toward 10GigE switches, interfaces, and transport. 10GigE enables a new array of architectural methods for designing and deploying IPTV and VOD service offerings—and is readily migrated to other forms of real-time and near real-time applications.
Evolution from ASI to 10GigE
When Ethernet got its start in the early 1980s, the market was very confused about transport standards for local area networks (LANs). IBM Token Ring, Datapoint ARCNET, and a host of others competed for dominance as network interface standards. Ethernet itself started off with slow data rates—2 Mbps, 5 Mbps and 10 Mbps. All media were shared access media. And the debate was dominated by arguments over whose media access control (MAC) method gave the best control over throughput and priority. In the late '80s, 3Com and other vendors redefined the landscape with the invention of the "collapsed backbone" concept.
A collapsed backbone is defined as the use of the backplane of a switch to network point-to-point forms of Ethernet. The capacity of the backplane is orders of magnitude higher than that of the physical interfaces. And solutions began to appear that were nonblocking in the backplane.
Meanwhile, work appeared at the transport layer—metro Ethernet over fiber, ring networks, and point-to-point networks that essentially "uncollapsed" the backbone. Standards evolved adding services such as virtual LANs (VLANs), priority, multicast features and link aggregation (n*ETH). Costs were reduced, and feature-rich Ethernet became the mass market media interface, with other competing standards rapidly dropping out of the game.
Today, Ethernet is the dominant form of enterprise, metro and (increasingly) backbone technology. While the traffic is IP, the interfaces are increasingly standardized GigE and n*GigE. Ethernet is used in more than 85 percent of all network connection interfaces, and more than 300 million hub and switch ports have been installed.
At the same time, the video industry migrated from analog to digital and defined its own industry-specific set of interfaces and standards—in particular ASI—for the transport of real time video. Figure 1 outlines the structure of systems using ASI.
The industry evolved ASI. Transport products were developed that carried ASI in a branching tree topology from the headend to the edge. Variants were evolved that carried ASI over metropolitan rings as an out-of-band service, and encoders and decoders were developed with ASI interface cards.
On the surface, things appeared to work out well, at least for broadcast-only applications. But as video evolved to a more switched model (multicast groups) and an on-demand model (unicast), the ASI networking toolkit did not scale accordingly.
A comparison of what can be done with Ethernet (and IP) systems vs. what can be done with ASI systems is particularly interesting:
• Ethernet scales to n*GigE data rates; ASI scales to sub-gigabit data rates.
• Ethernet has standardized switching solutions (which are also compatible with IP routers); ASI has limited point-to-point solutions and no switching capabilities.
• Ethernet supports unicast, multicast and broadcast addressing; ASI is unicast only.
• Ethernet supports very flexible switching topologies, such as tree-and-branch, ring and arbitrary topologies; ASI supports simple point-to-point topologies.
• Ethernet supports VLANs—isolated broadcast domains; ASI does not.
• Ethernet supports priority tagging with up to eight priority levels; ASI does not.
• Ethernet supports jumbo frames up to 9 kB; ASI does not.
• Ethernet economics and performance are driven by global mass market drivers; ASI is a tiny and expensive market of very limited performance.
• Ethernet innovation rate is high and maps to server, switch and transport innovation; ASI's innovation is limited and is primarily focused on how to "bridge" from legacy ASI to next generation GigE.
10GigE products, both NICs and switches, are being driven rapidly by the enterprise server market. Volumes are tripling year to year. Pricing is currently dropping to less than $5,000 per switch port and less than $1,000 per NIC. As such, 10GigE is no longer a technology of the future, but rather a viable alternative that allows IT managers and cable operations personnel to take advantage of the fat pipes for the consolidation of input/output (I/O) ports and cabling on servers and switches.
Applications design using 10GigE
A design trend that started with GigE is now accelerating with 10GigE. High-speed, low-latency transport and switching structures enable the network to be utilized as the backplane of a distributed solution. This design trend started with GigE and is now accelerating with 10GigE and is in stark contrast to the old view of systems design, in which the architecture approach was box-centric, with boxes connected together with slow point-to-point interfaces.
Some of the features of Ethernet switching (n*GigE and 10GigE) that affect network-centric solution designs include the following:
• High data rates can be used to ensure rapid completion of high priority functions. Startup bursts for cache-fill operations enable short latency startup of real-time video stream to "eyeballs." Also, high data rates can be used to rapidly enable resiliency functions such as propagation of duplicate copies of streams to other servers within the streaming server group.
• VLAN broadcast domain isolation can be used to minimize traffic flooding and also to implement clustering functions (e.g., VLAN-specific video streaming group).
• Multicast distribution can be used for keep-alive heartbeats and other resiliency functions.
• Jumbo frames can be used to minimize packet overhead and maximize throughput (e.g., performance is often limited to "packets-per-second" in servers).
• Priority tagging and switch support can be used to handle different priority network functions (such as high priority for transporting active video to a consumer and lower priority for transporting redundant copies of the video to backup servers).
GigE and 10GigE enable tiered caching
Caching is a well-known method for optimizing the mix of high and low cost components in an end-to-end system. The most expensive but highest performance storage is random access memory (RAM), while the least expensive storage is disk drives. High-speed connectivity between resilient elements, with the option of placing those elements in different locations, enables resiliency.
In order for a caching architecture to be effective, it must contain a suite of reactive algorithms that dynamically tracks changes in content popularity. For VOD items to be tracked, they have to include the titles that are being viewed and the most popular points within the titles that are being viewed. Popularity often cannot be predicted; a good example is the sudden death of a major entertainer, leading to a sudden uptake of libraries of content featuring that entertainer.
In addition, new forms of navigation can be enabled within caching streaming servers, such as chaptering (allowing entry into a piece of content at a set of locations) and linked virtual assets (program segments, advertisements, etc). Some chapters or virtual assets may be much more popular than others and thus remain in cache.
Four tiers of caching are logically part of a VOD solution (Figure 2).
Tier 1 is closest to the edge and is where the content is played out in real time. It also accounts for a small percentage of the hardware because of the caching efficiencies delivered by dynamic popularity algorithms.
Tier 2 is local storage and may be within a single system or in a peer system connected by high-speed local data capacity.
Tier 3 is a local content library that enables popular content to be held locally and accessed via inexpensive networking resources.
Tier 4 is the permanent source of content and can be located anywhere, even in a national headend.
Each caching tier has resiliency within groups of systems at that same tier and with systems at higher tiers within the architecture. The operator does not need to statically assign popularity tags to content or manually propagate content to libraries or specific caching nodes.
10GigE (and n*GigE) reduces the distributed node-to-node transport latency sufficiently so that a variety of cache management operations can be utilized. Figure 3 shows the dramatic latency reduction enabled by 10GigE—a 1-hour show can be filled in 1.35 seconds, or 2,667 independent and unique shows can be filled in 1 hour. These performance numbers assume full utilization of the available capacity. Actual numbers will vary because of the multifunction use of the transport and packet overhead.
High-speed transport and switching can be used as the vehicle of special operations, such as creating resiliency for the most popular titles. Since content is being filled into caches much faster than real time, sufficient data capacity exists to perform other management operations as lower priority operations. Functions such as a "provision multiple", whereas a stream is composed of several independent streams (e.g., program plus subscriber targeted ads), can also be assembled from multiple content locations.
Using n*GigE and 10GigE, operators can build nationwide networks today. 10GigE switches and transport are used both as a clustering mechanism within hubs and as a transport mechanism between hubs. Figure 4 outlines a typical design of such a system.
The end-to-end solution consists of groups of vault servers (with thousands of hours of content stored in each vault), intermediate groups of caching servers (whenever latency reduction or data capacity mitigation is needed), and edge groups of streaming servers. The addressing and switching tools provided by the GigE switching system are utilized, such as jumbo frames, VLANs, multicast pruning of server clusters, and prioritized transport. High-speed connectivity also strengthens the fault tolerance of the end-to-end system. For example, vaults and streaming caches can be replicated and rebuilt in near-real time. New systems can be added transparently, and content can be dynamically load-balanced as a background (but low latency) operation.
Finally, even the edge network interface can benefit from 10GigE transport. New breeds of edge modulators are being considered that can aggregate up to 128 6-MHz channels of downstream content. Using today's modulation (64/256-QAM), the aggregate capacity needed to serve such a cluster is on the order of 5 Gbps. Using future modulations, the spectral density can double, and thus the deployment of 10 Gbps in the backbone and within streaming servers enables the flexibility to grow into future 10 Gbps service capacity tiers in the edge network.
Commodity hardware solutions
Figure 5 outlines the tradeoff between proprietary hardware and commodity hardware. At a specific point in time, proprietary hardware can provide a higher performance solution to the market; in the early '90s, such solutions were developed for VOD. But streaming prices were orders of magnitude of what they are today, so these solutions were not ready to scale.
Today, thanks to the rapidly accelerating pace of processor performance, address space expansion, increase of on board memory, increase of I/O bus speed, and deployment of high speed NIC adapters such as 10GigE, the capabilities of standard servers have surpassed those of proprietary solutions. It is now feasible to use off-the-shelf servers to process and stream 10GigE's worth of streaming content. In addition, 10GigE can be used to handle key system functions such as content propagation and resiliency within a distributed system.
Because of the rapid product development cycles within the commodity server market, it is now also possible to develop a deployment strategy based on commodity hardware, and use software to enhance and differentiate service. Proprietary hardware solutions never leverage the volume economics of the commodity market, and lag behind commodity solutions in the adoption of new technology architectural innovations such as 10GigE.
10GigE and rapid advances in server hardware are enabling operators to rapidly evolve their solution architectures for deploying scaleable IP TV and VOD services. Development cycles are much faster than proprietary hardware solutions, and hardware selection is also much more flexible.
Switch, NIC and motherboard hardware prices will drop rapidly over the coming months and years. High speed, standards-based Ethernet switching architectures are already enabling network-centric, scaleable and fault tolerant solutions, giving operators the flexibility to deploy a range of architectures, from highly centralized to highly distributed, all feeding a high capacity edge network at the access edge.
Paul Sherer is president and CEO of Arroyo Video Solutions. Reach him at firstname.lastname@example.org.