§2.6.1–2.6.3Internet Video · DASH · Content Distribution Networks

What video costs

In streaming stored video streaming stored video Simple Watching a recorded video while it is still arriving, instead of waiting for the whole file. Precise An application whose underlying medium is prerecorded video placed on servers, which users request to view on demand. From a networking point of view the salient characteristic of video is its high bit rate: from 100 kbps for low quality to over 10 Mbps for 4K. The most important performance measure is average end-to-end throughput, which must be at least the bit rate of the compressed video for continuous playout. introduced in ch. 2 — open in glossary the medium is prerecorded — a film, a television programme, a sporting event, a user’s own upload. It sits on servers, and users request it on demand.

A video is a sequence of images displayed at a constant rate, typically 24 or 30 images per second. An uncompressed image is an array of pixels, each encoded into bits for luminance and colour.

The important property is that video can be compressed, trading quality against bit rate. Today’s off-the-shelf compression algorithms can compress a video to essentially any bit rate you want. The higher the rate, the better the image and the better the viewing experience.

Compression can also produce several versions of the same video, each at a different quality. The book’s example is three: 300 kbps, 1 Mbps and 3 Mbps. A user on a fast connection might choose the 3 Mbps version; a user watching over 3G on a phone might choose 300 kbps.

HTTP streaming, and its one big flaw

In HTTP streaming http streaming Simple Storing a video as an ordinary file and fetching it with a normal web request. Precise A scheme in which the video is stored at an HTTP server as an ordinary file with a specific URL. The client issues an HTTP GET, collects the bytes in a client application buffer, and begins playback once the buffer passes a threshold. Its major shortcoming is that every client receives the same encoding, whatever rate it can actually sustain. introduced in ch. 2 — open in glossary the video is simply stored at an HTTP server as an ordinary file with a specific URL (Uniform Resource Locator) Uniform Resource Locator The address of a single Web object: hostname plus path name. introduced in ch. 2 . The client opens a TCP (Transmission Control Protocol) Transmission Control Protocol The Internet transport protocol that delivers data reliably and in order, with flow control and congestion control. introduced in ch. 1 connection and issues a GET for that URL. The server sends the file as quickly as conditions allow.

On the client side the bytes collect in an application buffer. Once the number of bytes there passes a threshold, playback begins: the application periodically takes frames from the buffer, decompresses them and displays them. So the video plays while later parts of it are still arriving.

This has been extensively deployed — YouTube used it from its inception. It also has a major shortcoming:

DASH

Dynamic Adaptive Streaming over HTTP dynamic adaptive streaming over http Simple Storing several qualities of a video, and letting the player pick, chunk by chunk, what it can afford. Precise DASH. The video is encoded into several versions, each with a different bit rate and quality level, each stored under its own URL. The client requests chunks of a few seconds’ length one at a time. While downloading it measures the throughput it is receiving and runs a rate determination algorithm to choose which version the next chunk should come from. It therefore adapts both across clients and over time for one client. introduced in ch. 2 — open in glossary ( DASH (Dynamic Adaptive Streaming over HTTP) Dynamic Adaptive Streaming over HTTP Chunk-by-chunk quality selection driven by the throughput the client measures. introduced in ch. 2 ) removes the choice by not making it once.

The video is encoded into several versions, each with a different bit rate and quality. Each version is stored at the HTTP server under its own URL. The server also holds a manifest file manifest file Simple The list telling the player which versions of a video exist, and where each one is. Precise A file held by the DASH HTTP server that provides a URL for each version of the video along with its bit rate. The client requests it first to learn what versions exist, then selects one chunk at a time by giving a URL and a byte range in an HTTP GET request. introduced in ch. 2 — open in glossary , giving a URL for each version along with its bit rate.

The client requests the manifest first and learns what exists. It then selects one chunk at a time, each a few seconds of video, by giving a URL and a byte range in an HTTP GET request. While downloading, it measures the throughput it is receiving and runs a rate determination algorithm to choose which version the next chunk should come from.

Content distribution networks

Video companies distribute on-demand multi-megabit streams to millions of users every day. YouTube alone has a library of hundreds of millions of videos and serves hundreds of millions of streams daily.

The obvious approach is one massive data centre holding everything. There are three major problems with it.

So almost all major video-streaming companies use a content distribution network content distribution network Simple A fleet of servers spread around the world, each holding copies of the content. Precise A CDN manages servers in multiple geographically distributed locations and stores copies of videos and other Web content in those servers. It then tries to direct each user request to the CDN location that will give the best user experience. It may be a private CDN owned by the content provider itself, such as Google’s or Netflix’s, or a third-party CDN serving many providers, such as Akamai or Limelight. introduced in ch. 2 — open in glossary ( CDN (Content Distribution Network) Content Distribution Network Geographically distributed servers holding copies of content close to users. introduced in ch. 2 ). A CDN manages servers in multiple geographically distributed locations and stores copies of videos and other content there. It then tries to direct each user request to the location that will give the best experience.

A CDN may be private, owned by the content provider itself, as Google’s is for YouTube. Or it may be third-party, distributing on behalf of many providers, as Akamai, Limelight and Level-3 do.

Where to put the servers

CDNs adopt one of two placement philosophies. Enter deep enter deep Simple Putting many small server clusters inside access ISPs, right next to the users. Precise A CDN server placement philosophy, pioneered by Akamai, of entering deep into the access networks of ISPs by deploying server clusters in thousands of locations. The goal is to get close to end users, improving delay and throughput by reducing the number of links and routers between the user and the serving cluster. The cost is that maintaining and managing the clusters becomes challenging. introduced in ch. 2 — open in glossary puts server clusters inside access ISPs all over the world, as close to end users as possible. Bring home bring home Simple Putting a few large server clusters at exchange points instead of many small ones. Precise A CDN server placement philosophy, taken by Limelight and many others. It builds large clusters at a smaller number of sites — for example tens — typically placed in IXPs rather than inside access ISPs. It gives lower maintenance and management overhead, possibly at the expense of higher delay and lower throughput to end users. introduced in ch. 2 — open in glossary builds a smaller number of large clusters instead, usually at Internet Exchange Points.

Once the clusters exist, the CDN replicates content across them. It may not want a copy of every video everywhere, since some are rarely watched or popular only in certain countries. So many CDNs use a simple pull pull caching Simple A cache fetches and stores content the first time somebody asks for it. Precise A strategy in which a cluster that does not hold requested content fetches it on demand. It retrieves the content from a central repository or another cluster, stores a copy locally, and streams it to the client at the same time. When storage fills, infrequently requested content is removed. YouTube uses pull caching. introduced in ch. 2 — open in glossary strategy. Suppose a client requests a video from a cluster that does not hold it. The cluster fetches it, from a central repository or another cluster, and stores a copy locally while streaming it to the client at the same time. When storage fills, infrequently requested videos are removed — exactly as a web cache behaves in §2.2.5.

How you get steered to one

Now the mechanics. When a browser is told to retrieve a video by URL, the CDN must intercept the request. It has to choose a suitable cluster and redirect the client to a server in it. Most CDNs use DNS to do this.

Choosing the cluster

At the core of any CDN deployment is a cluster selection strategy cluster selection strategy Simple The rule a CDN uses to decide which of its clusters should serve you. Precise The mechanism for dynamically directing clients to a server cluster or data centre within a CDN. The CDN learns the IP address of the client’s local DNS server from the DNS lookup and chooses on that basis. One approach picks the geographically closest cluster, using a geo-location database; another picks the cluster with the best periodically measured delay. Both have weaknesses: some users are configured with a remote local DNS server, and many such servers do not answer probes. introduced in ch. 2 — open in glossary . The CDN learns the IP (Internet Protocol) Internet Protocol The network-layer protocol that defines the datagram format and addressing every Internet device must use. introduced in ch. 1 address of the client’s local DNS server from the lookup, and must choose a cluster on that basis. Strategies are generally proprietary, but two approaches are worth knowing.

Geographically closest. Using a commercial geo-location database, map the LDNS (Local DNS Server) Local DNS Server The nearby server a host queries first; a CDN identifies a client by its address. introduced in ch. 2 address to a place and pick the cluster fewest kilometres away. It works reasonably well for a large fraction of clients. It has three weaknesses:

• the geographically closest cluster may not be closest in path length or number of hops;
• some users are configured with a remote local DNS server, so its location says little about theirs — a problem inherent in every DNS-based approach;
• it ignores variation over time in delay and available throughput, always giving a particular client the same cluster.

Real-time measurements. Have each cluster periodically send probes — ping messages or DNS queries — to local DNS servers around the world, and choose on current conditions. One drawback: many local DNS servers are configured not to respond to such probes.

Check yourself