Akamai and beyond TCP: Scaling video delivery into the future by reinventing the core distribution protocols
Kit Knox, Vice President, Media Engineering
In its role delivering content for a vast number of customers, Akamai sits in a unique position to observe the biggest challenges in continuing to scale video delivery systems into the future. Having lived with TCP since 1974 as our primary mechanism for reliable message delivery, the time is now to design and deploy systems that meet the specific needs of our video platforms. This talk will cover ongoing R&D and production deployment results as Akamai works to replace TCP to provide secure protocol acceleration, multicast, and peer-to-peer enabled protocols.
Distributed encoding and AI encoding at scale
Mike Coward, engineering manager
Qi Huang, research scientist
As video becomes more popular on Facebook, the volume of user uploads rises along with the processing complexity from our products. Both trends greatly challenge the infrastructure team to provide a performing solution with flexible pipeline customization support. Given this load, we have to optimize for both speed and bandwidth so videos can be uploaded quickly and watched reliably independent of network quality. We'll discuss how distributed encoding improved our processing time and we'll describe how we apply machine learning to encode scenes individually at the highest possible quality while minimizing data consumption.
Facebook Live Video for News Feed
Peter Knowles, production engineer
Bhavana Radhakrishnan, software engineer
The recently announced Facebook Live Video product provides the ability for people to share live video on Facebook. This talk will cover the general architecture of the Live Video infrastructure, and focus on the network and caching topology required to implement such a massively scalable live video streaming system. There will be a description of the broadcast and playback media streaming protocols selected for the project, as well as an overview of the custom stream fanout proxy required to sustain such a large number of simultaneous viewers. Some of the scaling challenges that we faced during development will also be discussed, along with examples of how these were overcome.
360 video: Enabling high-quality streaming at scale and optimizing for VR
Evgeny Kuzyakov, software engineer
David Pio, software engineer
360 video is an immersive form of video that captures the world all around you. This presents a unique challenge of very high resolutions and bit rate, coupled with the fact that in VR viewers do not view the entire 360 world simultaneously. In this talk, we detail the transformations and optimizations applied to make high-quality 360 video encoding and streaming possible at scale.
Using AI to understand video
Manohar Paluri, research lead, Facebook AI Research and Applied Machine Learning teams
An enormous amount of video content is made available online every minute, and artificial intelligence plays an important role in helping people navigate this content. Facebook AI Research is pursuing various efforts to understand the semantics of these videos. This talk will describe our 3D Convolution-based fully supervised networks for video classification; our voxel to voxel prediction work, where we design a generic deep architecture for various video understanding problems that go beyond predicting labels; efforts in using video as a supervisory signal to understand the visual world; and some encouraging results in the unsupervised learning efforts.
Building the Complexity-Based Encoding optimization model
David Ronca, Director of Encoding Technologies
The Complexity-Based Encoding (CBE) optimization model developed by Netflix allows us to algorithmically determine the best video encoding recipe for a given title, at scale. Determining the optimal bit rates to use for a specific title is a very difficult problem to solve, but we faced many other significant challenges that were prerequisites to CBE. These challenges included building a robust and scalable encoding system, video quality metric at scale, and the ability to leverage idle web server instances for the massive re-encoding effort. Further, before we could roll out the new model, we had to ensure that the recipe change would play on the many millions of customers’ connected devices. This talk will discuss the each of the major challenges that we faced while working on the CBE project and how we solved them.
Live video for millions: Yahoo NFL game
Omer Luzzatti, Senior Director, Yahoo video platform
Yahoo's free, online live stream of a regular-season NFL game was the first time a game was provided only OTT out of the teams' local markets. This talk will discuss the challenge of scaling to a live TV audience of millions, describing approaches such as caching within the ISPs, multicast, and UDP streaming. It will describe Yahoo's platform, adaptive bit rate algorithm, CDNs dynamic selection, and collaboration with the ISPs.
HDR: Where are we this week?
Steven Robertson, software engineer, YouTube
High dynamic range video is developing toward broad access at a blistering pace, but we're not there yet. At YouTube, our challenge with new video technology isn't just to make sure it works, but also to make sure it's relevant — that it makes enough of a difference to enough viewers so that creators want to take advantage of it. This deeply practical talk will go over everything you need to start playing with HDR from creation to consumption, the current state of HDR support in VP9 (with numbers), and a detailed look at how we're planning to bring HDR to mobile.