Video: The Case for Specialized Hardware Acceleration in Cloud Services

在下一篇文章中了解有关云编码和交付的更多信息内容交付峰会.

看看Brent Yates的演讲，改变流媒体交付的经济性: Full-Stack Cloud Services Using Specialized Hardware Acceleration，在… 流媒体会议视频门户.

请阅读本视频的完整文本:

布伦特耶茨: We can't really talk about hardware without talking about software first, and Neil did a great job of building on this: The way software is being developed is changing. Everyone's going to the cloud, but the cloud is not a physical place. It's just a way of doing software and building these microservices, 为问题而设计, 对于运行问题的硬件来说不是这样.

Software designers want to be able to build a piece of software, 定义需求是什么, 然后让它在哪个服务器上运行, 而不需要知道它是什么操作系统, 它有多少内存, 上面有多少个驱动器, 这是什么处理器. 这些都是次要的问题. 他们关心的是解决业务问题. 这看起来很简单，它应该是这样的, but it took us a long time to get to the perfect storm of having all the features to be able to build that kind of system.

现在有了这些自动编排层, 和语言, 和工具, 你可以这么做, 你可以专注于解决问题, and let the automatic rules throw your software out at the edge. And in this case, out on the edge that is a service provided is part of a CDN. 它给了你一些很好的特性. It's easier to manage because you're focusing on the problem, not the hardware. You can have heterogeneous server deployments where they have different hardware requirements.

但是程序员不需要知道这些. The software automatically figures out where to run it on in the most efficient way. And it allows you as a business to focus on the problem and quickly adapt to changing concerns rather than having to focus on hardware deployments.

重申一下, 云可以是公共云，就像尼尔说的, 或者是混血儿, 或者在这种情况下, 一种将云扩展到边缘节点的独特方式. And it all takes advantage of these high-CPU-core counts by breaking up the software from the big monolithic applications, and I'm pointing to Neil's slide from the second one he showed, where he had the big monolithic application and then broken up into microservices. Those lightweight micro-applications make it easier to adapt to your business logic, 这也意味着它们的数量要多得多.

There's an overhead, there's a friction with having thousands of microservices on your server. But that friction is outweighed by the efficiencies you gain in breaking your application up that way.

So why do you want to add specialized hardware to this type of deployment? Why don't just throw it out there on regular Intel hardware with 72 cores and just let it run? The reason is that CPUs are not increasing based on Moore's Law. 摩尔定律已死. 我们只得到3.CPU性能同比提升5%. 所以你把这些东西扔到你的容器里, 你提供了更多的服务, 更多的问题, 更多聊天应用, but the only way you get faster is waiting 10 years for CPUs to change, 或者用更多的服务器来解决问题.

Throwing more servers at the problem is not only capital-intensive; it's also wasteful. 它不是绿色的. 这是电力问题. Microsoft can afford to spend $9 billion dollars in data centers, 谷歌可以花100亿美元. 但这些都是巨大的资本部署. 如果我们能提高效率，那就更好了, and change the way we do business instead of just throwing servers at it.

专门的硬件可以做到这一点.

我们认为这是一个能力缺口问题. On the bottom, the blue line is CPU performance, the red line is demand. 几乎和我们谈过的每个人都是这样, 我们今天在房间里四处走动, is talking about a 30-35% year-over-year gain in how many streams they have to revive, 他们需要运行多少应用程序, 他们发布了多少数据. And the only way to solve that gap is to fill it with servers.