Wednesday, July 04, 2012

Google I/O 2012 Day 2 Keynote by Urs Hölzle, Senior VP of Engineering: Google Compute Engine Announcement

YouTube’s video archive of Urs Hölzle’s Google Compute Engine announcement during the Google I/O 2012 Day 2 Keynote includes an interactive transcript. Following is a excerpt from the transcript starting at 00:36:07 and ending at 00:47:01:

image>> Sundar: To do that, I want to invite Urs Hölzle, our first V.P. of engineering, one of the first ten employees at Google, and the person, more than anyone, responsible for building all of Google's infrastructure. Urs.

[ Cheers and applause ]

imageThank you, Sundar. We've heard a great many things today about applications, clients, and -- but I'm here today to talk about servers and the infrastructure that powers all these applications.

A long time ago, at Google, we started working very hard to build the world's fastest, most scalable, and most reliable infrastructure. We needed our infrastructure to deliver incredible scale and incredible performance at a very low cost. And over the years, we built one of the world's largest set of data centers connect these data centers. And then we started looking at ways to deliver that infrastructure to you so that you can build great applications for your users.

So in 2008, we launched App Engine. App Engine lets you write simple, intuitive code to build your apps. And then we take them and we manage them and we scale them for you. So if you have one request per second or 1,000 requests per second, it just works.

imageWe're thrilled about the popularity of App Engine. Today, it's supporting over 1 million active applications, 7.5 billion hits per day every day, and 2 trillion data store operations per month. That easily makes it the largest public, NoSQL data store infrastructure in the world.

[ Applause ]

Thank you.

So more and more developers are using App Engine. And every day, we hear success stories from them. Here's one from Japan, where earlier this month, a fully one-third of Japan was watching a national song contest. And App Engine was powering the live streaming and the voting for this event. At its peak, it ran at 24,000 requests per second. Yet, no other App Engine user noticed that this was going on. So the ability to sail through this kind of traffic spike allows you to focus on your application while we focus on handling the traffic.

So we're very proud of what App Engine has enabled all of you developers to do.

But many of you have said you want even more options. And, specifically, you've told us that you want virtual machines on demand with industry-leading performance, with industry-leading scalability, and so easy and cheap to run that you'd prefer them to your in-house servers.

Well, today, I'm going to show you what it is like to run in Google's cloud.


imageI'm here to announce Google Compute Engine, infrastructure as a service that delivers the kind of performance --

[ Cheers and applause ]

-- and scale and value --

[ Applause ]

You haven't seen anything yet. Thanks.

But the infrastructure as a service that delivers the performance and scale and value that only Google can deliver.

So Compute Engine gives you Linux virtual machines at Google scale. Spin up two VMs or 10,000 VMs. It just works. You have multiple storage options, high-performance networking between VMs so that you can form them into a cluster and great connectivity to your end users using Google's global backbone.

So let me show what Compute Engine and App Engine together can do. One of our beta testers was the Institute for Systems Biology. The ISB uses genomic analysis to decipher how genes function and to find relationships between the genes that signal the potential for new cancer drugs. They wrote an App Engine app that connected to Compute Engine virtual machines. And you'll see it here on the screen here.


They show the human genome in a circular form. And the application allows researchers to visually explore the associations between genes, mutations, and other factors. And each of these associations is shown with a blue line like that.


Now, this may look simple, but this kind of analysis requires a tremendous amount of computation. So ISB built an in-house cluster with 1,000 nodes to handle this kind of computation.

Yet it still took them ten minutes to compute just a single association using the entire cluster. So, fortunately, it took only a few days to port this application to Compute Engine. And let's see how Compute Engine performs on this application. Now, remember, on their in-house 1,000-core cluster, what you would see here is one line appear, you know, click, and then you'd wait for ten minutes to see the next line. So let's see what Compute Engine can do with this. (Ticking).

All right. So this is what happens when you add 10,000 cores to your application. Instead of it taking ten minutes, you get a new association every few seconds.

That's the kind of scalability and performance that Google can deliver.

And so today, anyone with large-scale computing needs can now access this same infrastructure with Compute Engine virtual machines. And this infrastructure comes with a scale and a performance and a value that is unparalleled in the industry, because you benefit from the efficiency of Google data centers and our decade of experience in running them.

What you get from Google is not just the scale, but also amazing stable performance.

So our virtual machines and storage are predictably fast so you can rely on a consistent level of performance, no matter who else is running on Compute Engine right now, just as App Engine was able to handle a huge spike in the Japanese application.

For example, when Invite Media ported their ad server from another cloud provider to Compute Engine, they were able to double the number of connections that each single VM can accept at -- using VMs of comparable size, and, more importantly, they were able to reduce the number of connection errors by over a factor of ten because of the predictable performance.

You can hear more directly from Compute Engine beta users in the technical sessions about Compute Engine which you will see show up on your I/O calendar after this talk. So be sure to check them out. Now, finally, we know that you want top scalability and top performance. But you also need value. And Compute Engine is a great value, delivering up to 50% more compute per dollar than other cloud providers, so you don't have to choose between getting the best performance and getting the best price. We worked very hard for a decade to lower the cost of computing, and we're passing these savings on to you.

[ Applause ]

Compute Engine is now open in limited preview. And it's amazing to finally be able to talk about this. It's amazing to see 10,000 cores working for a single application and what that can do. So 10,000 cores, I think that's really cool. But you know what really cool?

[ Laughter ]

Well, we know that some of you need even more scale. And we have the technology to help you.

For computations that are very, very computationally intensive but don't need that much I/O, we can scale much, much higher. You may have noticed this ticker counting up since the start of my presentation. Well, this ticker is not a conceptual counter. It shows the actual count of the number of cores available to the genome app right now. And they've all been adding since the beginning of this presentation. So what would you say to 770,000 cores available to your app?

[ Cheers and applause ]

So if you're ISB, you would say, actually, my app kind of taps out at 600,000 cores. It doesn't scale any further. So let's leave 170,000 idle and allocate 600,000 to the genome app and see how fast it runs. Switch to the app. Right. There's a little initial delay. And then we really get going.

[ Cheers and applause ]

imageSo this is the same computation now running on 600,000 cores. That is cool!

[ Laughter ]

And that is how Infrastructure as a Service is supposed to work.

So this, ladies and gentlemen, truly is the best time ever to deliver and to build for the Web. You now have access to the scale and performance of Google's infrastructure at a great price. And it's up to you to figure out how to make the best use of that. So with that, back to Sundar. Thank you.

[ Cheers and applause ]


Very impressive. Adding 761,886 cores in an 00:04:07 demonstration period corresponds to 3,085 cores per second. For more details about ISB’s genome project, see Google’s Behind the Compute Engine demo at Google I/O 2012 Keynote article.