About application streaming

The Streaming Stack is based on WebRTC and integrates with GPUs from different vendors to enable low latency video encoding mandatory for any kind of real time streaming use case.

This section covers usage of the Streaming Stack, for installation instructions please refer to the installation page.
We also assume you know [how to get access to the Stream Gateway API].

Streaming Stack Overview

The Streaming Stack is highly customizable but requires some knowledge about how it works.
Each component of the stack fills a specific role:

Gateway Agent NATS Coturn
Provides the initial place where a client and a container go to start a peer to peer discussion. Those ‘chatrooms’ are called sessions and are central to the rest. A middleware between the Gateway and AMS. It secures access to AMS, launches containers, relays status information, and more. A messaging queue to allow components to communicate. A STUN/TURN server needed for WebRTC to work behind NATs and firewalls.

You can see how each component interact when creating a new streaming session

Streaming an Application

As mentionned above, sessions are an important concept for the Streaming Stack. They contain userdata, application information and more, but most importantly, they provide an entrypoint for both the client and the container to start the signaling process.
Signaling is a process by which both peers establish optimal codecs, network routes and content types.

1. Creating the session

WebRTC is a peer to peer protocol, but clients have to find each other through a central server first.
This is the primary purpose of the Stream Gateway. Sessions are equivalent to chatrooms that allow two peers to discover each other.

The session is created by calling the Gateway API at POST /1.0/sessions. The returned object contains information about the created session as well as a websocket url you have to use to start the signaling process.

2. The signaling process

When they can contact each other, both peers have to agree on a number of parameters. This process of exchanging messages
is called the signaling process.
Details about exchanged messages aren’t covered in this page, but you can find more information here.
The following provides an overview of the process:

  1. The container creates an offer containing the desired streams (video, audio, binary) as well as codec
    information.
  2. The client replies to that offer by accepting or refusing it.
  3. At that point, both peers agreed on the media type, but don’t know how to send data directly to each other (keep in mind that, so far, all communication is done through a websocket on the Gateway).
  4. They start going through the ICE protocol and interact with STUN servers to establish an optimal path.
  5. Each network path is bundled in an ICE candidate. There are usually multiple ICE candidates per peer and both sides negotiate the best candidate pair.
  6. Both peer agree on the best network path (ICE candidate) and start the actual streaming. At this point, they stop communicating through the Gateway websocket and talk directly to each other.

Hint: The server itself does not need to know about the messages content, it just has to forward messages from one peer to the other

3. Establishing the stream

When optimal ICE candidates have been selected and codecs capabilities have been agreed upon, the actual tracks are being sent. These tracks can be either video, audio or binary. The binary data channel can be useful to send arbitrary data like controls.
The streams can then be consumed to display the final content.

Supported Video Codecs

In order to support low latency encoding the list of supported video codecs is limited as not all codecs are either support by any of the supported GPU models or are not performing well enough for low latency purposes.

Currently supported video codecs are:

  • H.264
  • VP8

IMPORTANT: The use of H.264 requires a license from the MPEG LA. It’s your obligation to ensure you have the rights to streaming H.264 encoded video content to your users.

In the future we plan to add support for:

  • VP9
  • AV1

When additional codecs become available depends on when they are supported by the GPU vendors
in their hardware encoding solutions or if a viable software encoding solution exists.

Anbox Cloud combines both software and hardware video encoding in order to utilize available resources in the best possible way. Hardware video encoders usually have limited capacity of how many simulatenous video streams they can encode for low latency scenarios. The Nvidia T4 can for example encode 37 video streams at 720p and 30 frames per second (see “Turing H.264 Video Encoding Speed and Quality” for more details). Depending on the used CPU platform additional compute capacity might be available to support additional session via software encoding.


Last updated 15 days ago.