Mechanical Sympathy: The Software/Hardware Divide

Improving Video Encoding System Efficiency @Netflix

Every show on Netflix is encoded for the device type and bit rate that is optimal for the customers. For any one show that can be tens of different encodes, generated by a large-scale system that scales up to hundreds of thousands of CPU cores concurrently. As the performance team at Netflix, we are motivated to tune the system to generate encodes quickly and efficiently. 

In this talk, we explore the challenges of improving the efficiency of video encoding systems at Netflix scale, and cover actions we take to achieve it by leveraging high-resolution telemetry tools, hardware features, and resource-aware task scheduling.

It’s Not Your Machine; It’s Your Code.

Can we use information in tweets to decide how and when to trade stocks? Yes. But can we do this efficiently at scale, and on a tight budget? Well, this talk will discuss what such a task entails. We will look at some of the performance issues encountered in a simulated high frequency trading scenario, and how we can solve them with a hardware-conscious approach.    

By peeling back layers of software abstractions, we will learn skills and tools to decide how to make the most of our hardware, and how to circumvent limitations imposed by commonly used software.

Pragmatic Performance - Tales from the Trenches

In industry, one if often required to retrofit existing systems to extract more performance out of them. These performance optimization efforts are also performed under deadlines that make the "best'' solution, however attractive, often infeasible. In this talk, we would like to share techniques that have served us well at Target.  

While the scale of data that systems have to contend can be overwhelming, one can use that same scale to one's advantage. By observing statistical patterns in the data, we are able to alter the course of the computation to a more performant path. These course corrections can be done in real time or in an off-line manner. In this talk, we will present two examples that have made their way into production.

Mechanical Sympathy Panel

In this panel, we’ll continue the conversation about mechanical sympathy by bringing our speakers back to discuss where they see future trends and new ideas they have yet to try. We will touch on new tools that can provide even better insights into what is happening in our systems. We will also talk about different design choices that we as developers are able to make as hardware continues to evolve and the problems just get bigger and more complex. This will be an opportunity to learn more about how you can apply what you’ve heard in the talks to projects that serve your business.